A few years ago I started working on a book called SpaceX Foundation — a historical account of SpaceX’s first decade, told through first-hand sources: Elon’s company updates, launch dispatches, internal memos, the real-time record of a company that almost died three times and then became the most dominant launch provider on Earth.

The gap between SpaceX and everyone else is enormous and widening. Yet most of what’s been written focuses on Elon himself, not on the specific methods, culture, and decisions that actually built the company. That’s what the book is about.

While the book is still in progress, I’ve been writing an introduction essay as a way to work through the central question: why did SpaceX succeed in ways no one else has been able to replicate? And more importantly: is any of it learnable? The practices that made SpaceX dominant aren’t unique to rockets. They’re a blueprint for building anything hard.

If you want to read the book when it’s ready, subscribe here — I’ll be sharing updates and sections as they’re finalized. A private review edition is planned for later this year while I obtain legal permissions.

For now: the essay.

SpaceX has been remarkably open about how they operate. They’ve been succeeding in public for more than fifteen years now, and yet no one has replicated the results.

Competitors know their strategy. The engineering philosophy gets explained in interviews, tweets, and factory tours. Many of the ideas aren’t even new. Lockheed’s Skunkworks ran similar approaches sixty years ago — founder Kelly Johnson’s “14 Rules” read like a SpaceX operations manual.

The performance gap just keeps getting bigger. In 2025, SpaceX launched more mass to orbit than every other provider on Earth combined. Much more: every payload from China, Russia, Europe, and all American launchers wasn’t even a fifth of what SpaceX put into orbit. They’re the only company producing rockets at an industrial scale. Dare I use the word monopoly?

A Falcon 9 goes up every two to three days. Competitors manage single-digit launches per year. The same boosters have been reused twenty times each. The company has sent astronauts to the International Space Station — the first private company to do so. Starlink, their satellite internet constellation, now has over nine thousand satellites in orbit, the largest in history. Both built and launched by the same company. SpaceX is now possibly the most valuable private company on the planet.

The skeptics were confident it couldn’t happen. Apollo astronauts Neil Armstrong and Gene Cernan testified before Congress against commercial spaceflight. “Personally I think reusability is a dream,” said an Arianespace executive about SpaceX’s ambitions, “they’re not supermen.” And even if it did work, the market was too small to support the hundreds of launches needed to make reusability worth it. From a DC think tank: “The last time that California gurus predicted the era of commercial spaceflight had arrived, it turned into a disaster for the U.S. space program.” Musk was a software guy playing with expensive toys.

The early failures seemed to confirm them: three Falcon 1 explosions between 2006 and 2008. By September 2008, SpaceX had funds for exactly one more attempt, and Tesla was weeks from bankruptcy. Musk was borrowing money for rent.

Then it worked. Flight four succeeded, and NASA’s $1.6 billion cargo contract followed six weeks later. Then came Falcon 9, Dragon, ISS docking, boosters exploding on the pad, booster landings, crewed flights, and eventually Starship.

So is any of this outlier performance repeatable?

This is the puzzle. If the strategy is known and the principles are public, what’s actually hard to copy? Obvious factors explain some of this, but not enough.

Amongst them: The Space Shuttle retired, creating a gap. This was really good timing for NASA to become SpaceX’s biggest customer. But Blue Origin was founded two years earlier, and Boeing and Lockheed saw the same opportunity. The grand vision of “boots on Mars” attracted missionaries. But ambitious visions are cheap, and plenty of founders have them. Elon putting in $100 million bought early runway. But Bezos poured much more into Blue Origin, and legacy primes had multiples of this amount. Technology was also getting better: 3D printing, simulation, advanced materials. All commercially available to competitors.

These factors are all real, and none are sufficient. If they explained it, others could have caught up easily. But they’re not even close.

SpaceX is a hotbed of case study material — engineering, product, finance, strategy, manufacturing, project management, etc. If you’re interested in the company these are all important. But I’m more curious about which are repeatable. Was everything a flash in the pan? Or are some elements more broadly applicable?

The question isn’t “why did SpaceX succeed?” That’s too vague to be useful. The sharper question: What can someone building hard things actually take away?

The Strategy

What SpaceX has done, more than anything, is minimize the cost of getting things to space. The vision is humanity expanding across our solar system. But the lever is the cost of moving mass from Earth’s surface to orbit and beyond. Everything else — the launches, the landings, the reuse — serves that goal.

When you study how companies hold advantages over time, consistently being the low-cost provider might be the hardest to maintain. The reason is that it has to be baked into everything you do. It can’t be an initiative or an afterthought. It has to shape how you design products, structure the company, and choose what to build.

As you’ll see in this book, it all started from the earliest days.

Before starting SpaceX, Elon Musk wanted to get to Mars, but he didn’t set out to build a rocket manufacturer. In 2001, he tried buying Russian ICBMs to get there. The Russians quoted him ridiculous prices, so he famously reframed the question from first principles:

What is a rocket made of? Aerospace-grade aluminum alloys, plus some titanium, copper, and carbon fiber. And then I asked, what is the value of those materials on the commodity market? It turned out that the materials cost of a rocket was around 2 percent of the typical price—which is a crazy ratio for a large mechanical product.

Two percent. Your car’s raw materials are maybe 20-30% of sticker price. Consumer electronics are similar. But rockets? Ninety-eight cents of every dollar was going somewhere other than what it was made of.

Where?

Three places, it seems. Supplier markups stacking through contract layers, each tier adding 15-30% margin. Custom designs that couldn’t achieve manufacturing scale. Expendable hardware thrown away after every flight.

None of these are laws of physics. Traditional aerospace treated high costs as fixed constraints. But what if you treated them as variables? How do you actually capture that 98%?

Rethink from first principles

Start with the actual product. If you accept existing solutions, you accept their cost structure. So rebuild from physics instead.

Don’t ask “what do rockets cost?” Ask “what should rockets cost?”

Musk eventually named this the idiot index: the ratio of the actual cost of a part to the cost of its raw materials. “If the ratio is high,” he says, “you’re an idiot.”

Consider the Falcon 1 actuator. A vendor quoted $120,000 and eighteen months of development. SpaceX’s engineers built it for $3,900 by summer. When founding engineer Tom Mueller’s team asked about a critical engine valve, the supplier “kind of smirked and left” after hearing SpaceX’s timeline and budget. Mueller’s team made the valve themselves.

This pattern repeated across the vehicle. The Dragon capsule’s docking mechanism was re-invented from off-the-shelf bike shocks and catalog parts instead of adopting NASA’s existing design. There are probably a hundred examples like this, most not discussed in public.

The philosophy extended to fundamental architecture. SpaceX uses one propellant pair — liquid oxygen and RP-1 kerosene — across all stages. The vacuum Merlin engine uses a fixed nozzle extension instead of a deployable one. Fewer moving parts means fewer failure modes means lower cost.

Compare this to the Atlas V, which uses up to three different rocket types in a single vehicle, each optimized for its flight phase. Musk’s response: “You’ve just tripled your factory costs and all your operational costs.”

The Merlin engine family embodies this trade-off. Russian RD-180 engines cost $20-25 million each, while Merlin 1D production runs around $1 million. How? SpaceX eliminated hydrogen’s complexity by using kerosene, used regenerative cooling with the fuel itself, and optimized for manufacturability over maximum performance. The result was 95% of theoretical efficiency for 80% cost reduction. Merlin’s performance is slightly lower, but “good enough” at 1/20th the cost.

But identifying where to cut costs doesn’t mean you can actually cut them. You still have to build the parts. Once SpaceX concluded that atoms were cheap and process was expensive, vertical integration followed almost inevitably.

Become your own supplier

If materials are cheap and the tax is all process and overhead, you need to control the process to capture the savings. You can’t negotiate your way to 10x cost reduction with suppliers who have profits baked in at every tier.

So SpaceX became its own supplier. By building 80% of its hardware internally — engines, structures, avionics, software, and key ground systems — SpaceX collapsed the traditional aerospace stack. They outsource raw materials and commodity parts, and make everything else themselves.

“That’s something SpaceX didn’t originally set out to do,” one engineer noted, “but was driven to by suppliers’ high prices.” This wasn’t an ideological commitment to doing everything in-house; it was the result of suppliers repeatedly quoting prices and timelines incompatible with SpaceX’s cost targets.

The benefits compound. When several tiers each add 15% margin, total cost multiplies through the layers. A NASA study found SpaceX developed Falcon 9 for roughly $440 million — including most of Falcon 1 development. They estimated the same work with traditional contractors would have cost 3-10x that.

Vertical integration also accelerates iteration. When an engineer needs to change a bracket, weld, or circuit board, the manufacturing engineer is in the same building, using the same CAD systems and tooling. Materials, jigs, and processes can evolve together on the scale of weeks, enabling a rapid progression from Falcon 1 to successive Falcon 9 variants and finally Block 5, each iteration improving performance and reducing cost without waiting for suppliers to retool on multi-year cycles.

And it provides strategic control. When Russia threatened to cut off RD-180 engine sales, ULA faced existential risk. SpaceX was insulated.

The avionics example is instructive. Rather than buy radiation-hardened processors at $200,000 each, SpaceX used triple-redundant commercial processors totaling $2,000. The system voted on results. Equal reliability through software, not hardware premium.

But vertical integration creates a new problem: it concentrates fixed costs. If you own the factory, the machines, and the staff, you’re losing money every second they aren’t building something. At the traditional launch cadence of 2-4 vehicles per year, in-house manufacturing is a liability, not an advantage. To make the math work, you need volume.

Build a platform

The only way to get volume is to standardize. Build a common platform that customers have to adapt to.

The existing approach was bespoke vehicles per mission. Custom adapters, mission-specific modifications, multiple vehicle families. This optimizes each mission at the expense of manufacturing scale.

SpaceX bet the opposite: that cost savings from standardization would exceed the value of customization. Yes, customers wanted custom solutions. But they wanted low prices even more. Force them to choose, and they’d adapt. (Gwynne Shotwell, who led sales in early years, probably had to say no a lot.)

The Falcon 9 became the industry’s “Model T.” One rocket built in volume. Same nine Merlin engines on the first stage. Same vacuum Merlin on the second. Same structures, same diameter, same aluminum-lithium alloy, same welding methods, same avionics, same ground systems.

Even Falcon Heavy is just three Falcon 9 first stages strapped together with a shared upper stage. A scaled variant from the same core, not a new vehicle.

SpaceX published a “Falcon User’s Guide” with defined bolt circles, electrical connectors, and fairing environments. Customers design to SpaceX’s spec instead of demanding customizations. The standard 5-meter fairing became the industry norm. Satellite orbits adjusted to Falcon performance curves.

This flipped the negotiating power: Instead of aerospace companies serving satellite specifications, satellites adapted to SpaceX capabilities. The 5-meter fairing became an industry standard not because it was optimal for every payload, but because SpaceX made it the default and told customers to adapt. Most did.

The economics of manufacturing is what makes this work. Building 40 identical Falcon 9s annually creates automotive-style learning curves that are impossible in custom aerospace. As production scales, learning improves and cost declines. How this worked in practice is that every anomaly, wear pattern, or manufacturing defect fed back directly to the teams that designed the parts.

Finally, there’s the logical conclusion of standardization: reusability.

Reusable boosters are still the same Falcon 9 cores. You aren’t just building the same model of rocket; you are literally flying the exact same hardware. Because every booster was identical, every landing attempt provided perfectly comparable data. If making 40 rockets creates a manufacturing learning curve, flying the same rocket 20 times creates an operational learning curve that’s even steeper. The economics are devastating for competitors:

• Units 1-10: ~$60M per launch (learning the physics)

• Units 100-200: ~$30M per launch (manufacturing scale)

• Reused Boosters: <$15M marginal cost (the “Standardization” dividend)

Traditional providers, launching a handful of custom vehicles per year, never accumulate enough data to even start this cycle.

You can probably see why all three tactics were necessary.

First principles identified the waste. Vertical integration provided the control to eliminate it. Standardization allowed the volume to make that control profitable.

Without all three, the system breaks. First principles alone gives you a target you can’t reach — suppliers won’t cooperate. Vertical integration alone means high fixed costs with no volume to amortize. Standardization alone means a commodity product still built expensively through legacy supply chains. They work together so that each flight makes the next one cheaper.

What does this sound like? A flywheel! Lower costs enable lower prices, which capture market share, which increases volume, which drives costs lower still.

The incumbents understood this too late. They optimized components locally. Better engines, lighter materials, incremental gains. SpaceX optimized the system for cost, accepting component-level compromises for system-level dominance.

This flywheel relies on high volume. Competitors couldn’t imagine there was even a market for so many rockets. And they weren’t willing to take the risk of spending all that money just to find out.

The role of the “get to Mars” vision as a driving factor here can’t be emphasized enough. If your goal is just to beat traditional rockets, you wouldn’t spend so much effort on reuse or building a vertically integrated manufacturer from the ground up. I like Ben Thompson’s thought on this:

. . . if you start with the dream, then understand the cost structure necessary to achieve that dream, you force yourself down the only path possible, forgoing easier solutions that don’t scale for fantastical ones that do.

In a world where atoms are cheap and process is expensive, the real innovation was not a single engine or material, but the decision to redesign the entire stack around the economics of cost.

But a cost target doesn’t build itself. First-principles strategy says what to build. It doesn’t say how to build it without catastrophic mistakes along the way.

The Engineering

If the strategy is to rethink everything from first principles, how do you actually execute that without major consequential failures?

The standard answer is to analyze exhaustively before building. Traditional aerospace follows this path religiously. A 2020 NASA report on the Commercial Crew program noted that Boeing “utilizes a well-established systems engineering methodology targeted at an initial investment in engineering studies and analysis to mature the system design prior to building and testing.” Plan extensively. Freeze requirements early. Minimize test failures. This is the “measure twice, cut once” approach.

SpaceX inverted this.

Here’s the problem with the traditional approach: you can’t think your way to perfect solutions for problems you don’t fully understand. Your model is always wrong in ways you don’t know yet. Complex systems have emergent behaviors that only appear when the pieces are actually bolted together.

This is the paradox of first-principles design. If you’re questioning every inherited assumption (which you should), you’re venturing into territory where analysis alone can’t tell you what works. The physics might be known, but how the physics will interact with your specific materials, your specific manufacturing tolerances, your specific assembly process? That’s not something you can derive from first principles. That’s something you have to discover.

Use reality as your validation tool

The alternative is to use reality as your primary validation tool. “SpaceX focuses on rapidly iterating through a build-test-learn approach that drives modifications toward design maturity,” the same NASA report observed. Where Boeing invests up front in analysis, SpaceX invests up front in prototypes.

The core philosophy is deceptively simple: failures are data, not disasters. Tight feedback loops lead to a high rate of innovation and adaptation, quickly finding better solutions and what not to do. Speed is a tactical advantage.

This isn’t new. During World War II, the P-80 fighter jet went from concept to test flight in five months. In the early 1960s, the SR-71 Blackbird went from idea to rollout in four years (and it’s still the fastest manned plane every built). Small teams, fast iteration, real hardware.

In a 2021 Starbase interview, Musk explained the goal with each Starship prototype: “push the envelope... such that it blows up.” This sounds reckless until you understand what he’s actually saying. If the vehicle doesn’t fail, you haven’t learned where the limits are. Each failure is a precise data point about where reality diverges from your model.

How can you reconcile this “fail fast” approach with the care that’s needed to reliably build things where human lives are on the line?

The crucial distinction is between development and operations. SpaceX runs both, but with completely different risk profiles:

• Dragon (carries crew): Can never fail. Large margins of safety, exhaustive testing, conservative everything.

• Falcon 9 (operational launch vehicle): Middle ground. Ascent can’t fail, but some landing attempts are allowed to.

• Starship (development): Failure is instrumental. “Starship does not have anyone on board so we can blow things up,” Musk said. “It’s really helpful.”

“[This is] the same company doing two very different things with two different groups of people, two different risk profiles,” said entrepreneur Steve Blank, “but they’re talking to each other.”

The Falcon 1 story illustrates how this works in practice. The first flight failed — a fuel leak caused an engine fire. The second flight failed — residual propellant in the first stage caused it to recontact and damage the second stage. The third flight failed — first and second stage collided during separation due to residual thrust. Each failure was specific. Each gave the team something concrete to fix.

Flight four succeeded. It put a payload into orbit for the first time, and it saved the company. They had just enough money and grit to get through four.

The contrast with traditional aerospace is stark. In that world, a three-failure start may have triggered years of analysis, review boards, and redesigns-on-paper before the next attempt. At SpaceX, each flight became the next test, with fixes incorporated immediately.

This pattern continued into Starship. The early integrated flights each ended in “rapid unscheduled disassemblies” — explosions, basically — but each came after achieving partial objectives: clearing the pad, passing max-Q, reaching near-orbital velocity. Then finally the famous catch of the superheavy booster (like a 20 story building falling from the edge of space). Each subsequent flight incorporated design changes based on telemetry from the previous one. Where traditional aerospace might take years to go from flight anomaly to design change, SpaceX was doing it between flights.

Have a high production rate

Iteration only works if you can afford many attempts. This is where SpaceX’s “hardware-rich” approach becomes essential.

“A high production rate solves many ills,” Musk has said repeatedly. He continues:

Any given technology development is “How many iterations do you have? And what’s your time and progress between iterations?” So if you have a high production rate, you can have a lot of iterations. You can try lots of different things. . . . If you have a small number of engines, then you have to be much more conservative because you can’t risk blowing them up.

Traditional aerospace builds few prototypes, each one expensive and near-flight-ready. SpaceX builds many cheaper prototypes: hardware-rich fleets of test articles. They’d rather have ten rough versions to blow up than one polished version they’re afraid to break. This can lead to specific design decisions like using stainless steel for Starship (cheap, easy to weld in a tent) instead of carbon fiber (expensive, requires giant autoclaves).

Grasshopper, the experimental vehicle they used to develop propulsive landing, embodied this. As SpaceX investor Steve Jurvetson put it, Grasshopper was a “software testing rig.” Not a rocket in the traditional sense — a physical debugger. The software was the hard part, and Grasshopper was the hardware-in-the-loop needed to test it against reality. Work it out on the rig first, then move on to the bigger one.

Vertical integration really helps enable this. When you own the factory, you can build fast without waiting on vendors. When you own 3D printing capability, you can produce parts on an ad-hoc basis. When you can manufacture Raptor engines at high volume, losing one to a test failure doesn’t set you back months.

SpaceX is big into simulation as well. This moves atoms to bits where possible, letting them pre-screen designs before blowing things up.

But real tests remain primary. The question being constantly asked is “How quickly can it be tested in as real environment as possible?” You see this mentality in every successful frontier tech effort. Even at NASA with their “all-up testing” philosophy during the Apollo program. From George Mueller, head project manager for Apollo: “At a system level you’re much better off testing the system because in the end that system has to work. And then the only way you find out is if you test it as a system.”

The pieces reinforce each other in a way that’s easy to miss.

First-principles engineering reduces unnecessary complexity. Fewer parts means each prototype is cheaper to build. Cheaper prototypes means you can build more of them. More prototypes means faster iteration. Faster iteration means you can push each prototype to failure without being precious about it. More failures means more data. More data means better designs. Better designs mean even simpler solutions. And the cycle continues.

Meanwhile, fail-fast iteration extracts maximum information per prototype and per flight. You’re not just testing whether something works — you’re finding exactly where it breaks. That precision accelerates the next iteration.

The strategy (low-cost, vertical integration) enables the engineering approach. The engineering approach validates the strategy faster than analysis ever could. Traditional aerospace eliminates uncertainty through planning. SpaceX eliminates uncertainty through doing.

But this system doesn’t run itself.

Running an organization that constantly questions requirements, deletes parts, and accepts visible failures requires something that can’t be built through iteration. An engineering process that treats failure as data only works if the engineers themselves believe it. A system that pushes to the edge of what’s possible only survives if the people doing the pushing can handle the intensity.

The practices described here are the mechanism. But they’re powered by something else entirely.

The People

Back to my original point: The practices I’ve described so far aren’t secret. So why can’t others just copy them?

The standard answer is organizational inertia, bureaucracy, risk aversion. . . And there’s truth to all of these, but it’s not the whole story.

The answer is that strategy doesn’t exist in isolation. The same playbook in a different environment would produce different results, or nothing at all. You can’t copy strategy without transplanting the conditions that make them work. A fail-fast culture needs people willing to fail visibly. A first-principles approach requires people willing to question experts. Skip-level truth-seeking requires people willing to deliver bad news directly to the CEO.

These aren’t process changes, they’re selection effects. You can’t copy strategy without transplanting the conditions that make them work.

The variable I’ve been circling around is people. Not in the bland HR sense of “our people are our greatest asset.” In the structural sense: who shows up, what they believe, and what behaviors they’re willing to accept from each other. SpaceX didn’t just hire good engineers. It built a system that attracts, retains, and amplifies a particular kind of engineer while filtering out everyone else.

The variable starts with Musk. Without him, none of this exists. I don’t say this as hagiography, only an observation about initial conditions. Someone had to fund a rocket company when the idea seemed crazy. Someone had to decide that “colonize Mars” was an actual engineering target.

I want to call out three factors in particular that were foundational to SpaceX’s culture.

The first: an ambitious vision that functions as a recruiting filter. Building cities on other planets isn’t just aspirational branding — it’s a sorting mechanism. The mission attracts missionaries, not mercenaries. Engineers who would never work for “just another launch company” will work brutal hours for a shot at making humanity multiplanetary.

Does this get us to Mars faster? No? Then let’s skip it for now.

This is a real quote from a meeting at Starbase. When the mission is that clear, prioritization becomes automatic. Every decision has a simple test.

Second: constant forcing functions, both real and manufactured. I think this is underrated. Some were genuinely existential, like the 2008 cash crisis when SpaceX had funds for exactly one more Falcon 1 attempt. Demo flights with fixed dates. Competitive pressure from other bidders on NASA contracts.

Others Musk created himself. Aggressive public timelines that even he knew were ambitious. Internal “do or die” milestones that felt real even when they weren’t legally binding. The forcing function prevents drift. You can’t endlessly study when a deadline — real or perceived — is bearing down. Even an arbitrary deadline is better than no deadline, because it forces decisions.

Third: direct technical engagement that bypasses organizational filters. Andrej Karpathy, who worked for Musk at Tesla, notes that Musk spends about 50% of his time talking directly to engineers. Not to VPs summarizing engineer work.

This sounds obvious, but it’s unusual in the corporate world. The CEO has to trust the CTO, who works through layers of managers to enact a vision. Each layer is a “hop” where information is lost. It’s like a game of telephone; by the time the technical reality reaches the CEO, it has been polished, caveated, and de-risked. It’s a summary of a summary, with the inconvenient details removed.

SpaceX collapses the chain. The CEO-CTO-VP-engineer layers become a single conversation. By talking directly to the engineers, Musk removes the signal loss. They become the source of truth, not the filtered narratives that typically reach CEOs.

This isn’t just about speed and accuracy — it also allows SpaceX to make bolder technical bets. Musk stays aligned on what is actually possible. A non-technical manager can’t tell the difference between a tactically painful path and a strategically necessary one. If your managers tell you a certain chassis design or engine material is too difficult, and you don’t have the technical depth to interrogate them, you have to defer. The decision to use steel for Starship over carbon fiber is a prime example. It was highly against conventional wisdom and controversial even within the company. Ultimately, Musk had to understand all the tradeoffs himself and make a call.

When there’s a blocker — Raptor production, GPU supply, a regulatory delay — Musk intervenes personally. Personal phone calls to other CEOs. Daily updates on the specific constraint until it’s resolved. This is the “large hammer” approach, as Karpathy puts it.

The hammer only works because the signal is clear. If you don’t know exactly where the bottleneck is, you’re just swinging in the dark.

But Musk alone doesn’t explain SpaceX. Plenty of ambitious, technically engaged founders have failed spectacularly in aerospace. Founders set direction, but it takes more than that to sustain a company.

Start with Gwynne Shotwell. Currently the President of SpaceX and part of the core founding team, she’s the one “holding it all together” as many believe.

But it’s a mistake to view her as merely the steady hand who keeps the lights on while the engineers dream. She’s the strategic co-architect of the entire system.

In most other companies, the sales and business teams are natural enemies of engineering logic; they promise the customer whatever they want, which creates the very bespoke complexity that first-principles thinking is trying to delete. Shotwell did the opposite. She recognized that for the manufacturing flywheel to work, the market had to be forced to adapt to the rocket. She made it so the world’s most conservative buyers — NASA and the Pentagon — would accept a radical new model of standardization.

She ensured that every dollar saved by engineering was converted into a dominant market position. Musk had the vision and the hammer, while Shotwell translated the vision into a practical reality. She’s the reason SpaceX didn’t end up as another very impressive, very bankrupt space startup.

The combination also worked because of where it happened: Southern California, where aerospace culture runs deep. Not just available talent from declining programs, but the lineage back to early aviation pioneers building flying machines in hangars. That expertise was dormant, buried under decades of bureaucracy. What Musk grafted onto it were Silicon Valley operating norms: flat hierarchy, engineer ownership, permission to walk out of unproductive meetings.

From this foundation, specific cultural practices emerged. Not values on a poster — actual behavioral memes that spread through the organization. Memes that show up in day-to-day decisions about what to build, who gets promoted, and how to respond when rockets blow up.

Meme 1: Tip-of-the-spear focus

Always identify and attack the biggest limiter. Don’t spread effort across secondary problems. Laser in on the single constraint that, if removed, would unlock everything downstream.

This is true at every level. Each SpaceX site has a single dominating objective to simplify prioritization. (Starbase has one goal: get to the Moon, then Mars.)

A NASA manager who visited SpaceX observed that when a new problem appears, “it looks like a flash mob” in the hallway.

When a system-level bottleneck is identified, it gets disproportionate resources. When Starship development was bottlenecked on Raptor engine production, that became the company’s focus. Not propellant loading. Not heat shields. Not launch infrastructure. Raptors. Musk gave it absolute focus: daily updates, memos to the company, resources redirected from elsewhere. Once engine production broke through, attention shifted to the next constraint. The limiter always gets the hammer.

Meme 2: Push through roadblocks

A roadblock isn’t a reason, it’s a problem statement. You either clear it or escalate until someone does.

Admitting you’re blocked isn’t shameful at SpaceX; it’s expected. Hiding a blocker is what gets you in trouble. As one engineer described it: solving blockers “moved the needle forward on several projects.” The cultural expectation is honesty about what’s not working and relentless effort to fix it.

Falcon 1’s history is the clearest example. After the third failure in 2008, SpaceX had money for exactly one more attempt. The company was, as Musk later said, “running on fumes.” The engineers didn’t retreat into analysis or slow down to reduce risk. They kept iterating on specific failure causes until the fourth flight worked. That success unlocked NASA’s $1.6 billion CRS contract for cargo to the ISS. The roadblock mentality turned existential risk into long-term survival.

Meme 3: Scrappiness

Cost-sensitive resourcefulness over bureaucratic process. This goes hand-in-hand with the low-cost provider mentality.

SpaceX’s replacement for NASA’s heritage docking system was prototyped with bike shocks and McMaster-Carr catalog parts (NASA engineers dismissively called it “McDocker”). Not because SpaceX couldn’t afford better, but because crude prototypes let you test ideas before committing to expensive development. The engineers rolled a rough mock-up directly to Musk’s desk for review. NASA initially viewed McDocker as a weakness — how could a system built from commodity parts match their flight-qualified heritage design? But SpaceX had questioned the requirements. Dragon didn’t need all the flexibility and mass of NASA’s generic adapter. A simpler design could meet the actual constraints. McDocker flew successfully, lighter and cheaper than what it replaced.

The scrappy approach extends everywhere: reusing test hardware, hacking tools together, building ground support equipment from industrial components instead of aerospace-grade systems.

Small teams build end-to-end instead of handing off between specialized groups. Engineers are expected to design, build, and test what they own. Musk calls the alternative “ivory tower engineering” — design something, throw it over the wall, and let someone else figure out how to actually make it. At SpaceX, the person who drew the bracket is the person who welds it.

Meme 4: Question requirements

Every constraint — customer, regulatory, internal — is treated as a hypothesis to interrogate, not a fact to accept. This is the embodiment of first principles thinking.

Falcon 9’s grid fins were originally designed to fold, like traditional aerospace grid fins. The folding mechanism reduced drag during ascent, which seemed obviously necessary. SpaceX questioned whether it was worth the mass and complexity. Simulations showed fixed fins were acceptable. So they deleted the mechanism entirely.

The best part is no part. The best process is no process.

This applies recursively. When NASA offered their docking adapter — decades of heritage design, already qualified, effectively free — SpaceX still asked whether Dragon actually needed it. Could a simpler design meet the real constraints? The answer was yes, and the result was McDocker.

Junior engineers are explicitly told that requirements from “smart people” are the most dangerous, because nobody thinks to question them. Every requirement must have an owner: a specific person who can defend why it exists. If the owner can’t explain it, or the original reason no longer applies, the requirement gets deleted.

This turns into Musk’s now well-known rule: If you aren’t adding back at least 10% of the requirements you deleted, you aren’t deleting enough.

Meme 5: Treat everything as learning

Failures and explosions are data for the next iteration, not disasters to be concealed.

SpaceX published compilation videos titled “How Not to Land an Orbital Rocket.” Spectacular droneship crashes, set to music. This isn’t just PR, it’s a genuine signal that visible failure is acceptable if you extract the lesson.

Many would see the early Falcon 9 booster landings as spectacular failures. Rockets exploding on droneships, tipping over, crashing into the ocean. But over time, those iterations produced a landing success rate high enough to support reusing boosters dozens of times. That reusability is what makes Falcon 9 economically dominant. You don’t get there without the explosions first.

The cultural frame matters: a failed test is only bad if you didn’t learn enough from it. Instrument everything. Do ruthless post-mortems. Then move fast on the next design.

These memes reinforce each other, and they reinforce the strategy.

Vision attracts people who thrive in this culture. The culture then selects for more of the same — high performers stay, others self-select out. The practices become “how we do things” independent of any individual. Small, elite teams maintained by default. Low performers actively removed rather than accumulated.

This is the real moat. SpaceX’s cost advantage can theoretically be matched. Their technical innovations can be studied and replicated. But the culture requires rebuilding an organization from scratch.

Are you seeing the pattern yet? The entire system reinforces itself, spinning the flywheel faster and becoming increasingly hard to copy. It’s loops all the way down.

Feedback Loops

So what’s actually hard to copy?

Strategy identified the waste: 98% of every dollar going to process instead of atoms. Engineering found the path: iterate fast, and validate using reality instead of thinking your way to perfect solutions. Culture made it move: question requirements, fail visibly, attack the tip of the spear.

Three systems, mutually reinforcing. Each turn of the flywheel makes the next one easier. The answer to “what’s hard to copy?” isn’t any single tactic. It’s that the tactics are a system. Copy one without the others and it breaks down. First-principles design without vertical integration gives you targets you can’t reach. Vertical integration without volume makes your fixed costs a liability. A fail-fast culture without people who can tolerate visible failure becomes theater.

The output isn’t just cheaper rockets.

The real output is a generation trained to build hard things. Engineers who internalized these memes — question requirements, fail fast, tip of the spear — are now scattered across the frontier. Space startups, defense tech, manufacturing automation, energy. The cultural memes are spreading. The lessons aren’t trapped inside one company.

This matters for anyone trying to build something hard.

Working at frontiers creates optionality invisible from the ground. Starlink wasn’t in the original vision. It emerged because SpaceX was already there, launching so frequently and cheaply that a 9,000-satellite constellation became feasible. Others couldn’t see that opportunity because no one else was in position to take it. When you’re at the frontier, possibilities present themselves that don’t exist for anyone else.

Small groups with the right structure can do extraordinary things. Not “brilliant individuals” — structured teams with fast feedback loops, real forcing functions, and cultural tolerance for visible failure. The P-80 in five months. The SR-71 in four years. Falcon 1 to Falcon 9 in four years. This has happened before. It can happen again.

The lesson isn’t “be like Elon.” Hero worship is the wrong takeaway. One person didn’t build Starship. The lesson is that structure matters more than the hero. Get the system right — selection, iteration speed, forcing functions, cultural memes — and the results follow.

If I had to distill it to one question: how fast are your feedback loops? How fast can you get to reality? You see this pattern in every successful frontier tech effort. Apollo’s all-up testing. Early aviation pioneers in hangars. SpaceX blowing up Starship prototypes to find the limits. The common thread is treating reality as the teacher and getting to class as often as possible.

That’s the analysis. But analysis is reconstruction — pattern-matching after the fact, shaped by hindsight. It’s useful, but it’s not the same as watching the system get built.

What follows is the raw material.

The SpaceX company updates: over 100 dispatches from 2003 to 2013. The first five years written primarily by Elon. Technical challenges explained in real time. Near-death moments. Incremental victories that didn’t seem incremental at the time. The culture and memes showing up in the language before anyone named them.

It starts in May 2002. Elon hires Tom Mueller, a propulsion engineer building rocket engines in his garage on nights and weekends. They incorporate a company called Space Exploration Technologies.

And they begin.

A few numbers to tell the story: an $18 textile stock became a share worth $680,920 over six decades. That’s a 37,829x return, compounding at over 19% annually. The previous edition of this book has sold over 100,000 copies since 2013.

This format works because chronology matters. You watch Buffett’s thinking evolve without the distortion of hindsight. Quote compilations cherry-pick the highlights. The complete letters show you the dead ends, pivots, and gradual development of principles that became investing doctrine.

The letters document the transformation of a declining textile company with $25 million in equity into a trillion-dollar conglomerate. Buffett and Charlie Munger’s partnership drove this change, but the deeper story is cultural. They built a system that attracts exceptional leaders and compounds reputation into competitive advantage.

Physical books work better than PDFs for this kind of study. You can flip between years, cross-reference decisions, and mark patterns the way serious students should. Buffett keeps copies of this book in his office for exactly this reason.

This is the primary source for understanding capital allocation in practice. Business schools teach theory. These letters show how the world’s most successful investor actually thought through decisions as they happened. No academic interpretation, no journalist’s filter — just Buffett teaching directly.

If you already have the last edition from 10 years ago, here’s a list of the additions:

• Latest 10 letters from 2015-2024

• 8 supplemental letters or appendices sent to shareholders over the years (like the letter about merging with Blue Chip and how to think of the valuation)

• New forward written by myself

• Detailed data visualization of annual change in book value + year-end balance sheet

The original agreement to publish these letters was done with a proverbial handshake (no contracts needed!). That’s how Berkshire works, and it’s part of why this collection matters. You’re not just reading investment strategy. You’re learning the principles that built one of the world’s most admired companies.

I’m grateful to have played a small role in spreading Buffett’s wisdom and Berkshire’s story. I hope that this final edition helps you become a better investor and thinker. The lessons in these letters will endure for generations.

Available for pre-order now, releasing October 14, 2025 on Amazon in hardcover and Kindle formats.

Berkshire Hathaway book

In 2012 I worked with Warren Buffett to publish the first edition of “Berkshire Hathaway Letters to Shareholders”, a compilation of all of his letters.

The last edition was released 10 years ago for the 50th anniversary of Buffett’s takeover, and I’ve been selling it ever since. The Kindle version is updated annually with every new letter.

It’s now been 60 years since Buffett took over Berkshire Hathaway, and last month he announced his unfortunate retirement from the CEO position.

Time for a new edition!

Since February, I’ve been finishing up the 60th anniversary edition of Berkshire Hathaway Letters to Shareholders. It adds the last 10 years of letters (duh), plus a few more intra-year letters and a new (short) forward by me.

The final publication date is unknown, but it should be in the next few months (Q3 2025).

Berkshire Memorex

A few years ago I built a few projects to experiment with the new GPT-class language and embedding models that had just been released. The first was BuffettBot (already woefully outdated compared to the latest LLMs), the second was Memorex.

For Memorex, the idea was to create an interface allowing you read and explore a collection of related documents. Search them semantically (by meaning), find related passages, and overlay a chat UI on top to ask questions.

The Berkshire Memorex was the first collection released but I’ve experimented with others like Paul Graham’s essays.

Last week the Berkshire archive was updated with the latest letter and annual meetings. Check them out if you get the chance.

I had planned on adding a bunch of new features but haven’t gotten around to it yet. Since the first release, Google came out with NotebookLM which has a similar premise. I’d say the intent of Memorex is more to create publicly-accessible collections rather than being a pure notebook where the only goal is dumping content in and learning from it. I’d still like to add a few features though, like:

1. A chat UI where you can ask questions to the collection.

2. “Explain this” button for highlights that does a deeper dive on the topic, along with other related content.

3. Auto-generated topic index. Scan through this index rather than searching.

4. Generated audio / podcasts. Instead of the NotebookLM “hosted” podcast (which is really cool), this would be a literal reading of the documents, like an audiobook. The latest TTS audio models from OpenAI and ElevenLabs are crazy good and I think this would be something I would personally use a lot.

SpaceX Foundation

“SpaceX Foundation” is a book that documents the first 10 years of SpaceX’s history through first-hand sources.

After SpaceX was founded in 2002, Elon (and later others in the company) published regular updates to their mailing list describing in detail what the company was working on and how they made progress. From 2003 to 2013 there were 100+ updates released that make an astounding case study and “real-time” history of early SpaceX and the technical and operational challenges they overcame.

I’ve been working on this book to some degree for 3 years now, though seriously for only the last year. It’s currently in draft state while I focus on other priorities, including getting the Berkshire book out the door.

Aside from my own limited time, the biggest issue right now is getting full buy-in from SpaceX. Most of the content is findable via public archives, but there’s a lot of implicitly-copyrighted photos, etc. in the content. Plus I can’t use SpaceX in the title without explicit permission, otherwise I risk legal action. I’ve shown it to a few people at SpaceX who really liked it, but needless to say they have more important things to focus on. Like making humanity multi-planetary :)

Request for connections: If you know anyone who works at SpaceX, I’d love to talk with them about the book. I’ve already shown it to a handful of people and executives but always looking for more feedback. Reach out to me on X if so.

Mashgin

It’s now been 10 years since I joined Mashgin as the founding employee.

Since raising our Series B three years ago, Mashgin has expanded quickly into convenience stores, universities, and major event venues globally. Check out our photo gallery to see them in their “natural habitats” around the world.

Earlier this year, Amazon significantly scaled back their “Just Walk Out” (Amazon Go) initiative, publicly stating cumulative items sold of 18 million. For context, Mashgin now processes roughly that same volume in less than 1 week. We handle over a million transactions daily across more than 4,000 deployed locations, recently surpassing one billion transactions.

Given how slowly our target industries typically adapt, this growth highlights Mashgin’s unique speed and effectiveness. We’ve also been profitable since 2022, another rarity in startup-land.

As always, progress hasn’t been smooth. It’s been a crazy ride with multiple near-death experiences and mini pivots. But it’s been an amazing 10 years being involved in everything from product and engineering to sales, strategy, and operations.

Our next S-curve is now upon us, and I’m excited to see where we’ll go next.

If you’re interested in joining us we have a bunch of open roles (local in the Bay Area or remote): jobs.lever.co/mashgin

On October 13, 2024, SpaceX launched their fifth test flight of Starship, the largest rocket ever flown. The larger booster section returned to land and was caught by the tower with giant mechanical arms.

News of the catch went viral. More people are finally starting to pay attention to Starship after October’s launch and catch of the superheavy booster. If you somehow missed it — watch this highlight reel to catch up. Eric Berger wrote about the implications here as well.

For the booster, returning it in one piece is the hard part. This was the single biggest obstacle to the entire Starship concept.

This is why the catch was such a big deal. Someday soon, it will be launched, caught, refueled, and sent back up quickly and often.

So what?

Here’s a recap of what’s happened so far:

Starship has now been proven to be capable of:

• Getting to orbit (roughly).

• Returning Super Heavy booster.

• Sub-orbital reentry of Starship upper-stage with precise splashdown landing.

• Live video from nearly the entire flight. (Not critical, but I think it’s awesome enough to include as it’s the first ever live video during reentry.)

Re-entry for an orbital ship is really hard. It’s currently done with capsules like Dragon, Starliner and others. The bottoms have thick tiles to absorb the massive amount of frictional heat produced. The Space Shuttle did this, but even then they had multiple failures, including the tragic loss of Columbia in 2003.

But let’s ignore the upper-stage for now as it may take years to nail down resuse. Soon it will be possible to reuse the booster. This is already monumental for a few reasons.

First, cost.

Estimated cost for a full Starship stack currently is $90-100M. This is already 10x cheaper than other large cost-plus rockets. At a payload of 100 tons, this would be $1,000 per kg if Starship is fully expendable (no reusability). Here’s a chart of launch costs to LEO. Pretty damn good! It would be the cheapest way to get to space, just under Falcon Heavy. And as scale increases, costs will come down further, maybe as low as $50M.

The booster cost is now around $40-50M. SpaceX may start reusing it in a year from now. If they can reuse the booster 5 times with $2M refurb each time, it brings cost down to ~$600/kg. You can see that with conservative assumptions, cost is already very low.

Cost ultimately matters most, but it’s not everything. Other game changing benefits of Starship are:

• Launch frequency: SpaceX currently has capacity to build at least 25 ships a year. The plan is to eventually get it to one per day. They’re not even in full production mode and still CRANKING them out compared to other big rockets. When makers of space hardware can assume that they’ll be easily able to launch more than once per month it changes what you can build.

• Size: The total volume of 100-150 tons is significantly more than any other rocket. So you can just get more stuff into space in a single launch. More than that though — the volume is much larger. The James Webb Space Telescope cost $10 billion to make, and a lot of the complexity was because they had to origami it into ~200 cubic meters of space. Starship volume is 5x this. If the next big telescope is launched on Starship, it can be both much bigger and much cheaper. Scientists should be rethinking their future plans.

Starship’s significance lies in its potential to make space travel dramatically cheaper, faster, and more sustainable. It could enable massive payloads, in-orbit refueling, and rapid reuse, paving the way for lunar bases, Mars colonization, and humanity’s expansion into the solar system. Beyond the technical achievements, it represents a profound shift in what humanity is capable of.

What’s next?

The key here is iteration speed. Each integrated flight test (IFT) pushes Starship to its limit in one or more ways. This maximizes learning not only for the rocket design itself, but for the entire Starship system — launch pad infrastructure, ground control, propellant systems, and manufacturing.

To take full advantage of Starship (on the Moon, Mars, or beyond), it has to be refueled many times in orbit. Critics say these refueling missions are tricky, but it might not be if launches get cheaper and more common.

Even as Starship starts delivering actual payloads, production volume becomes critical. The more it launches, the cheaper it gets. Needing 12 launches for every trip to the Moon seems daunting until you consider their goal of averaging a launch per day.

Mass production is not something the space industry has ever needed. Building “the machine that builds the machine” is exactly what’s going in in Starbase, Texas right now.

Finally we’ll have a rocket that's treated like a commercial airliner, that can depart and land multiple times every day, with the only limit being maintenance.

But just a reminder: this shit is hard. And it’s not just the engineering of the rocket itself. For example, each launch requires an insane amount of fuel and liquid oxygen. From Eric Berger:

Put another way, launching four Starship rockets in a single day would consume all of the nation’s liquid oxygen capacity for that day. Accordingly, SpaceX must find a way to scale production of liquid oxygen, and ensure a tremendous supply to both South Texas and its future Starship launch facilities in Florida.

That’s what happens though when you push the boundaries of frontiers. You end up having to solve problems in a lot of areas.

In the next few years, SpaceX wants to show they can restart engines in space, land Starship safely back on land or a tower, transfer fuel between ships, and eventually reuse all parts. If they succeed, humans could land on the Moon again, maybe in 2028, and later reach Mars, thanks to Starship’s massive size and quick turnaround.

Here’s the next milestones and their estimated times (via Eric Berger and others):

NASA’s future role

In other exciting news: Jared Isaacman has been chosen to lead NASA during the Trump administration. Isaacman is the billionaire founder of payments company Shift4. He’s also a jet pilot, astronaut, and first private citizen to perform a spacewalk (just this past September if you missed it).

He is yet to be confirmed but insiders seem confident it will happen. Isaacman has earned respect in the space community for his leadership, commitment, and bold vision for the future of space exploration.

As NASA administrator, he’s expected to push for modernization, focusing on cutting wasteful programs like the SLS rocket and prioritizing collaborations with private companies to advance space exploration efficiently and competitively.

I’m really hopeful here that Isaacman makes NASA much more of a lean organization, capable of spending their irreplaceable taxpayer dollars on exploration and science that truly matters rather than just jobs programs.

Next flight: IFT-7, currently planned for Monday 1/13

The next test flight could launch this Monday. Here’s a list of what they’re testing:

• Payload Deployment. Deploying 10 Starlink simulators in a suborbital trajectory, testing deployment systems. This is the first time Starship will attempt to deploy a payload.

• New Upper Stage Design. Redesigned flaps (smaller and repositioned), increased propellant capacity, vacuum-insulated feedlines, backup heat shield layer, improved avionics, and metallic tiles with active cooling. A significant number of tiles have been intentionally removed to stress-test vulnerable areas and assess the vehicle's thermal and structural limits during reentry. More cameras have been added as well, with 120Mbps of high-def video sent down during entire flight.

• Reusing Booster Hardware. First reuse of a flight-proven Raptor engine. Eventually the whole thing will be reused so they’re starting with a single engine.

• Testing Catch Systems. Upgraded chopstick radar sensors, and improved tower protections.

Once again, the goal is to push it to the limit, testing different capabilities. In some cases, the tests will be “to failure”, meaning it is expected for things to break. Imagine how useful it would have been for the Space Shuttle to try a handful of launches with heat shield tiles intentionally removed to test where failure points were.

Make sure to watch the test — currently schedule for Monday at 5pm ET, 2pm PT.

BONUS: Blue Origin’s New Glenn launch on Sunday

Blue Origin’s New Glenn rocket, originally scheduled to launch this morning, is now expected to launch on Sunday 1/12. This will be its first launch.

New Glenn is a (yet to be proven) reusable rocket with a payload capacity up to 45 tons, about double Falcon 9. So it’s a heavy-lift rocket with potentially much more capability than Falcon 9, but still significantly less than Starship.

This is great news! SpaceX has been so far in the lead in recent years, it will be good to have the promise of a little bit of actual competition.

It’s the cycle of innovation and consequence. We see a problem X, we invent a solution, we discover that solution creates a new problem, we can't stop doing X, and we have to invent another solution. And so on. This is our philosophy: to seek, to solve, to stumble anew.

We invented fire, which kept us warm and cooked our food, and also burned down our villages and killed us in wildfires. We invented knives and arrows and saws, which helped us hunt and build, and also cut off our fingers and stabbed us in the gut. We invented agriculture, providing food surplus, yet it sowed seeds of war, famine, and environmental decay. We invented writing, which allowed us to communicate and record history, and also spread lies and misinformation. We invented steam power, which increased productivity and transportation, and also polluted our air and crushed our limbs in factories. We invented smartphones, and… well, I won’t go there.

It’s the basic arc of human history. Progress is driven by unintended consequences. But can we manage these risks?

Pro-action

The Precautionary Principle says we shouldn’t pursue any action or policy that might cause significant harm unless we’re nearly certain it’s safe. It puts the burden of proof on the inventor to prove the invention is not harmful.

Had we always used this approach though, most invention would never have occurred.

A better approach is the Proactionary Principle.

The Proactionary Principle is based on the observation that the most valuable innovations are often not obvious, and we don't understand them until after they've been invented. It says we need to weigh the opportunity cost of restrictive measures against the potential damages of new technologies.

Kevin Kelly expands on this more in the book “What Technology Wants”:

Anticipation. The more techniques we use, the better, because different techniques fit different technologies. Scenarios, forecasts, and outright science fiction give partial pictures, which is the best we can expect. Anticipation should not be judgment—the purpose is preparation not accuracy.

Continual assessment. We have increasing means to quantifiably test everything we use all the time, not just once.

Prioritization of risks, including natural ones. Risks are real but endless. Not all risks are equal. They must be weighted and prioritized.

Rapid correction of harm. When things go wrong things should be remedied quickly. Rapid restitution for harm done would also indirectly aid the adoption of future technologies. But restitution should be fair (no hypothetical harm or potential harm).

Not prohibition by redirection. Banishment of dubious tech does not work. Instead, find them new jobs. A technology can play different roles in society. It can have more than one expression. It may take many tries, many jobs, and many mistakes before a fit is made. Prohibiting them only drives them underground, where their worst traits are brought out.

Let me repeat this line: “Not all risks are equal.”

This is the key. Nassim Taleb talks about risk of ruin, and he argues the Precautionary Principle should only apply when systemic, extinction-level risks are possible. I tend to agree, though it can be a hard balance.

The risk of ruin matters because the cost is complete destruction of humanity or life on Earth. These are the cases where applying the Precautionary Principle makes sense.

Even so, in areas like bioengineering and AI it’s too easy to conjure hypothetical, catastrophic scenarios. Engineered pandemics. Superintelligent paperclip-maximizer AIs. All sorts of bad things can happen! And none of them are impossible. These are the scenarios Taleb is talking about — but we need to weigh them appropriately. Our primitive monkey brains are good at over-estimating very unlikely risks.2

Find out

The path of progress has always been to venture into the unknown, make mistakes, and find out.

We’ve found out how to make fire, build cities, cure diseases, split the atom, land on the moon, turn sand into intelligence, and edit genes. We’ve found out how to make weapons of mass destruction, how to pollute the environment, how to exploit and oppress each other. We’ve also found out how to prevent wars, protect the environment, and promote human rights.

The Proactionary Principle is not just a guide; it’s a belief in action over inaction, in knowledge over ignorance, in growth over stagnation, in hope over fear.

Let’s choose to “find out”, and embrace the future unknowns. Because, let’s face it... we’re still going to fuck around and find out anyway.

A deep tech startup builds and scales technology at the frontier of what’s possible using applied research or novel combinations of existing tech.

Deep tech has been defined by others as a startup having primarily technical risk, but no market risk, of the “if you build it, they will come” variety.

Seth Bannon pushed back on this a few months ago:

I agreed, and after thinking about it more, I feel that both these are their own dimensions:

The dimensions here are:

• ↕ Technical risk — aka “novelty uncertainty”. High tech risk = “deep tech” at the frontier of practical technical applications (really hard to build). Low tech risk = well-known and understood tech.

• ↔️ Market risk — aka “complexity uncertainty”. High market risk = market is new/unknown. Low market risk = demand and needs are clear.

These scales may not be true “risk” as in the risk of loss, but more like uncertainty as Jerry Neumann describes in this really good essay, calling them instead novelty uncertainty and complexity uncertainty.

The quadrants:

1. “Field of Dreams” (↑ tech, ↓ market): Distribution is straightforward if product works. SpaceX Falcon 1, fusion, fission energy 1960-1980, AGI, Genentech, mRNA vaccines.

2. “Frontiers” (↑ tech, ↑ market): Hard to build, market is a hypothesis. Building new frontiers. NVIDIA, TSMC, SpaceX Falcon 9, Tesla, small modular reactors, ChatGPT.

3. “Pioneers” (↓ tech, ↑ market): Easy to build v1 but distribution is hard; initially product is only differentiated on brand + design. Exploiting new frontiers. Social media, Uber, Airbnb, Groupon, most consumer SaaS.

4. “Arbitrage” (↓ tech, ↓ market): Closing arbitrages or expansion of proven ideas. Instagram cloning Snap features, “DoorDash for India”, many vertical SaaS products, Rocket Internet / Samwer brothers.

In general, Frontiers yield the best investment returns. It pays to resolve a lot of technical and economic uncertainty. There are other dimensions of course and you can be successful (or not) in any of these quadrants. Some Arbitrages can be huge. And some Frontier products can be commoditized.

Field of Dreams seems like the place to be: “if you build it, they will come.” The tech may not pan out though — and even if it does, if the breakthrough is easily copied it can be competitive. There has to be something that makes it hard to copy.

The median VC-backed startup is in the Pioneer quadrant. The tech is straightforward to implement. The market is new and uncertain for whatever reason, but more than likely it’s because of a frontier that was recently opened up by another technology. This makes distribution really hard. Either there’s no one there and demand has to be created, or the demand is obvious and there’s a ton of competition until the uncertainty is resolved (and winners have scaled).

Frontiers open up new markets for pioneers to fill. Like microprocessors leading to business software or smartphones leading to Uber.

Arbitrage are “sure things” more akin to exploiting a market inefficiency. There’s demand the market isn’t supplying and it’s pretty clear how to close the gap. Parts of the business world can be surprisingly inefficient.

* * *

What about other dimensions? I can think of a handful of others that would be good to categorize startups. For example, R&D cost vs. scaling cost (potentially distinguishing “deep tech” and “hard tech” startups).

But I’ll leave those for another post.

Thanks to Celeste and Eric Jorgenson for feedback.

Watch it live here at 8am Eastern, 5am Pacific. The launch window is only 20 minutes, so if they find any issues, it will likely be scrubbed to a new date.

If you’re excited (or might be) but don’t follow Starship closely, this post is for you.

Starship is SpaceX's fully reusable rocket designed for missions to Earth orbit, the Moon, Mars, and beyond. It’s the biggest rocket ever flown, taller than the Statue of Liberty, with the capacity of an entire space station.

Why is Starship so important?

Starship represents a giant leap towards affordable space travel and expanding Earth’s ecosystem into our galaxy. SpaceX has already reduced costs by 10x with Falcon, and Starship could push it down another 10x. It will be critical for all future NASA programs and expanding our knowledge of the universe.

For more on the details of why Starship makes space travel so cheap, see my essay from 2021 on the future of space.

Starship began with design concepts in 2012. From 2019-2021 SpaceX tested a variety of prototypes, but achieved a milestone with a successful high-altitude flight and landing of SN15 in May 2021. The first orbital test flight in April 2023 ended explosively, yet it marked a significant learning opportunity for SpaceX.

This is a critical part of SpaceX’s culture: rapid prototyping, frequent testing, and a tolerance for failure, allowing for swift iteration and improvement. It's a build-test-learn philosophy that's a world apart from traditional aerospace practices. This has enabled faster innovation and adaptation to real-world challenges.

Who said failure wasn't an option?

In April, Starship's test flight saw the rocket experience engine failures and loss of control, leading to an explosion, or RUD (rapid unscheduled disassembly). While the mission did not achieve all its goals, the test was an essential learning step, not a failure, as it provided critical data for future iterations.

Why has it been 7 months between tests?

1. SpaceX correcting their issues from test 1, in particular building a new launch pad and water deluge system as the first one was destroyed (due to the surrounding ecosystem, they can’t build a trench to divert flames).

2. Working with the FAA, and later Fish & Wildlife Services, who were conducting thorough safety and environmental reviews. Slowness of the FAA was partly due to their inability to meet permitting demand — a lot of which is from SpaceX's frequent Falcon launches.

Integrated Flight Test 2 aims to launch Starship from Boca Chica, Texas, achieve stage separation, and reach orbit. The bottom booster is expected to splash down in the Gulf of Mexico, while the upper Starship should travel eastward around the world and land in the Pacific Ocean near Hawaii after re-entry.

Tony Bela has a great infographic on what to expect from the launch:

If you can’t watch the launch, make sure to check out the recap tomorrow.

And remember: Failure is an expected outcome of test flights like this!

Everyone would love to see 100% mission success, but tests are designed to push the system to its limits in real conditions to identify weaknesses. This iterative process is essential for developing a reliable and robust spacecraft capable of fulfilling SpaceX's interplanetary ambitions.

More links if you’re still interested:

• At long last, the glorious future we were promised in space is on the way. “If humanity wants to level up as a spacefaring species, this is the clear path forward.

  Finally, we are walking it.”

• The Future of Space, Part II: Where are we headed in the next 10+ years?

For the past few years I’ve been doing a roundup post and essay about once a quarter. This will be an audio version of that. I wanted to try a podcast format where it’s primarily clips from others with me narrating along the way.

Why? It sounded like fun and I love experimenting with how to make content on different mediums. I think it turned out pretty well, although much longer than I initially thought.

Going through all these audio files was a pain, so I wrote a script that transcribed them all and allows me to search by topic and group related content. This was kinda fun to make as well.

Most of this episode is on AI, similar to my last few roundups, but I also did a section on the future of education, and I’ve got some carveouts for other interesting content at the end as well.

The podcast is almost an hour long, you should be able to skip to relevant sections in the description. And if you’re not an audio person, there’s the usual written version below.

• 01:37 — 🤖 The A.I. Frontier

• 08:45 — AI: How useful are language models?

• 14:05 — AI: What does AI allow us to do?

• 14:33 — AI: Easier to communicate

• 18:16 — AI: Creativity

• 25:50 — AI: Augmented Intelligence

• 33:29 — 🧑‍🏫 The Education Frontier

• 43:00 — 🔗 Interesting Content

🤖 A.I. Frontier

In this section about AI, I’m going to first cover some basic fundamentals of how these AI models work, and then really get into their use cases and what they allow us to do.

It’s been 3 years since the original GPT-3 release, 6 months since ChatGPT, and we’ve had GPT-4 and other open source models for more than 3 months now.

So with that said, I want to talk about the current state of GenAI, in particular language models because these have more generalized promise.

Here’s Andrej Karpathy in a talk called “State of GPT” with more on what models are out there and what they can do:

Now since then we've seen an entire evolutionary tree of base models that everyone has trained. Not all of these models are available. For example, the GPT-4 base model was never released. The GPT-4 model that you might be interacting with over API is not a base model, it's an assistant model. And we're going to cover how to get those in a bit. GPT-3 base model is available via the API under the name DaVinci. And GPT-2 base model is available even as weights on our GitHub repo. But currently the best available base model probably is the LLaMA series from Meta, although it is not commercially licensed.

Now one thing to point out is base models are not assistants. They don't want to answer to you, they don't want to make answers to your questions. They just want to complete documents. So if you tell them, write a poem about the bread and cheese, it will answer questions with more questions. It's just completing what it thinks is a document. However, you can prompt them in a specific way for base models that is more likely to work.

A review of what we have now:

• Monolithic foundation models from OpenAI, Google, Anthropic, etc.

• Assistant models built on top of these like ChatGPT, Bard, and Claude.

• Open source models — particularly the semi-open-sourced LLaMA model from Meta, where you can run it locally and even fine-tune it with your own data.

State-of-the-art GPT-4 is actually a “mixture of experts” assistant model that routes your prompt to whichever large model that can complete it best. This is a big difference from the original GPT-3.

In the early days of GPT-3, if you “asked” it a question, it had a high chance of spitting out something rude, offensive, or just didn’t really answer the question. How do they get from the base models to the assistant models like ChatGPT that are pleasant to talk to?

So instead, we have a different path to make actual GPT assistants, not just base model document completers. And so that takes us into supervised fine tuning. So in the supervised fine tuning stage, we are going to collect small but high-quality data sets. And in this case, we're going to ask human contractors to gather data of the form prompt and ideal response. And we're going to collect lots of these, typically tens of thousands or something like that. And then we're going to still do language modeling on this data. So nothing changed algorithmically. We're just swapping out a training set. So it used to be internet documents, which is a high-quantity, low-quality for basically QA prompt response kind of data. And that is low-quantity, high-quality. So we would still do language modeling. And then after training, we get an SFT model. And you can actually deploy these models. And they are actual assistants. And they work to some extent.

The performance of these models seems to have surprised even the engineers that created them. How and why do these models work?

Stephen Wolfram released a great essay about this earlier this year aptly called “What is ChatGPT Doing… and Why Does It Work?” The essay was basically a small book, and Wolfram actually did turn it into a small book. So it’s long, but if you’re interested in these, I’d really recommend it, even if not everything makes sense. He explains models, neural nets, and a bunch of other building blocks. Here’s the gist of what ChatGPT does:

The basic concept of ChatGPT is at some level rather simple. Start from a huge sample of human-created text from the web, books, etc. Then train a neural net to generate text that’s “like this”. And in particular, make it able to start from a “prompt” and then continue with text that’s “like what it’s been trained with”.

As we’ve seen, the actual neural net in ChatGPT is made up of very simple elements—though billions of them. And the basic operation of the neural net is also very simple, consisting essentially of passing input derived from the text it’s generated so far “once through its elements” (without any loops, etc.) for every new word (or part of a word) that it generates.

. . . The specific engineering of ChatGPT has made it quite compelling. But ultimately (at least until it can use outside tools) ChatGPT is “merely” pulling out some “coherent thread of text” from the “statistics of conventional wisdom” that it’s accumulated. But it’s amazing how human-like the results are. And as I’ve discussed, this suggests something that’s at least scientifically very important: that human language (and the patterns of thinking behind it) are somehow simpler and more “law like” in their structure than we thought. ChatGPT has implicitly discovered it. But we can potentially explicitly expose it, with semantic grammar, computational language, etc.

In a YouTube lecture he put out, Wolfram explains that the ChatGPT's ability to generate plausible text is due to its ability to use the known structures of language, specifically the regularity in grammatical syntax.

We know that sentences aren't random jumbles of words. Sentences are made up with nouns in particular places, verbs in particular places, and we can represent that by a parse tree in which we say, here's the whole sentence, there's a noun phrase, a verb phrase, another noun phrase, these are broken down in certain ways. This is the parse tree, and in order for this to be a grammatically correct sentence, there are only certain possible forms of parse tree that correspond to a grammatically correct sentence. So this is a regularity of language that we've known for a couple of thousand years.

How useful are language models?

How useful really are language models, and what are they good and bad at?

I have occasionally used it to summarize super long emails, but I've never used it to write one. I actually summarize documents is something I use it for a lot. It's super good at that. I use it for translation. I use it to learn things.

That was Sam Altman, founder of OpenAI, saying he primarily uses it for summarizing. At another event he admitted that ChatGPT plug-ins haven’t really caught on yet, despite their potential.

What are they good for then? What are the LLM “primitives”— or the base capabilities that they’re actually good at?

The major primitives are: Summarization, text expansion, basic reasoning, semantic search in meaning space. Most of all translation — between human languages, computer languages, and any combination of them. I’ll talk about this in a little bit.

They’re clearly not good at retrieving information, or being a store of data. The same way human brains aren’t very good at this naturally without heavy training. This might never really be a core primitive of generalized LLMs. They are not a database and likely will never be. And that’s ok — they can always be supplemented with vector or structured data sources.

As I mentioned, the context window of a transformer is its working memory. If you can load the working memory with any information that is relevant to the task, the model will work extremely well, because it can immediately access all that memory.

Andrej Karpathy talks here about what LLMs aren’t good at, but also how they can easily be supplemented:

And the emerging recipe there is you take relevant documents, you split them up into chunks, you embed all of them, and you basically get embedding vectors that represent that data. You store that in the vector store, and then at test time, you make some kind of a query to your vector store, and you fetch chunks that might be relevant to your task, and you stuff them into the prompt, and then you generate. So this can work quite well in practice.

So this is I think similar to when you and I solve problems. You can do everything from your memory and transformers have very large and extensive memory, but also it really helps to reference some primary documents. So whenever you find yourself going back to a textbook to find something or whenever you find yourself going back to documentation of a library to look something up, the transformers definitely want to do that too. You have some memory over how some documentation of a library works, but it's much better to look it up. So the same applies here.”

I can attest to this working very well, and nearly eliminating any hallucinations.

Advanced reasoning also isn’t there yet, but it may be in the near future. Other methods for both prompting and training techniques are currently being explored.

I think more generally, a lot of these techniques fall into the bucket of what I would say, recreating our system 2. So you might be familiar with the system 1, system 2 thinking for humans. System 1 is a fast, automatic process. And I think kind of corresponds to like an LLM, just sampling tokens. And system 2 is the slower, deliberate planning sort of part of your brain. And so this is a paper actually from just last week, because this space is pretty quickly evolving. It's called “Tree of Thought”. And in “Tree of Thought”, the authors of this paper propose maintaining multiple completions for any given prompt. And then they are also scoring them along the way and keeping the ones that are going well, if that makes sense. And so a lot of people are really playing around with kind of prompt engineering to basically bring back some of these abilities that we sort of have in our brain for LLMs.

Karpathy mentions “Tree of Thought” as one solution to LLM problem solving. This is one of a bunch of different exploratory efforts going on to hugely improve results without even retraining the base models.

OpenAI also recently said they aligned a model via process supervision, rather than outcome supervision. What this means is that, when training, instead of rewarding it for getting the “right” answer they reward it for thinking through the steps of the problem itself, called “chain-of-thought” reasoning. It seems to me that models like this will really allow future GPT versions to do more advanced reasoning.

And if you pay for ChatGPT Pro, as of last week you now have access to Code Interpreter. This is being described by some as GPT-4.5 because of its ability to write and troubleshoot code to solve problems. (It’s seriously good — try if you can!)

What does AI allow us to do?

Abilities fall into 3 major categories for me:

1. They make us smarter.

2. They make us more creative.

3. They make it much easier to communicate with computers and other people.

3. Easier to communicate

Language models act as a translation or mapping layer between humans and machines. So if I describe something in English to an LLM it could translate to Mandarin so someone in China could understand, Python so it can be executed as code, or be reworded in English so a 5 year old gets it.

Part of this “mapping” is routing tasks to the right tools, where we then can use them and have finer UI controls. In a recent New Yorker article, the writer talked about AI making computers less rigid:

Many of the uses of A.I. that I like rest on advantages we gain when computers get less rigid. Digital stuff as we have known it has a brittle quality that forces people to conform to it, rather than assess it. We’ve all endured the agony of watching some poor soul at a doctor’s office struggle to do the expected thing on a front-desk screen. The face contorts; humanity is undermined. The need to conform to digital designs has created an ambient expectation of human subservience. A positive spin on A.I. is that it might spell the end of this torture, if we use it well. We can now imagine a Web site that reformulates itself on the fly for someone who is color-blind, say, or a site that tailors itself to someone’s particular cognitive abilities and styles. A humanist like me wants people to have more control, rather than be overly influenced or guided by technology. Flexibility may give us back some agency.

Design researcher Jakob Nielsen, the business partner of Don Norman (of “Design of Everyday Things” fame) wrote a short post recently on how generative AI is introducing the third user-interface paradigm in computing history. Unlike the current paradigm of command-based interaction design, where users issue commands to the computer one at a time to produce a desired outcome, users now tell the computer directly what they want.

With the new AI systems, the user no longer tells the computer what to do. Rather, the user tells the computer what outcome they want. Thus, the third UI paradigm, represented by current generative AI, is intent-based outcome specification. . . . the user tells the computer the desired result but does not specify how this outcome should be accomplished. Compared to traditional command-based interaction, this paradigm completely reverses the locus of control.

But are we destined to only be interacting with these AI models through speaking or writing? Hopefully not, as I discussed in my last essay. Nielsen continues:

The current chat-based interaction style also suffers from requiring users to write out their problems as prose text. Based on recent literacy research, I deem it likely that half the population in rich countries is not articulate enough to get good results from one of the current AI bots. That said, the AI user interface represents a different paradigm of the interaction between humans and computers — a paradigm that holds much promise. . . . Future AI systems will likely have a hybrid user interface that combines elements of both intent-based and command-based interfaces while still retaining many GUI elements.

2. The Creativity Frontier

The second ability is that AI makes us more creative. It turn us into explorers of possibility space.

What do I mean by possibility space? You’ll also see it called meaning space or latent space, which for these purposes are interchangeable.

It can be a complicated topic to explain, especially without any visual aids. I’ll put a few links in the footnotes, and eventually I hope to write a post specifically about this. Here’s a passage from Stephen Wolfram’s “What is ChatGPT Doing?” essay:

One can think of an embedding as a way to try to represent the “essence” of something by an array of numbers—with the property that “nearby things” are represented by nearby numbers.

And so, for example, we can think of a word embedding as trying to lay out words in a kind of “meaning space” in which words that are somehow “nearby in meaning” appear nearby in the embedding.

. . . Roughly the idea is to look at large amounts of text (here 5 billion words from the web) and then see “how similar” the “environments” are in which different words appear. So, for example, “alligator” and “crocodile” will often appear almost interchangeably in otherwise similar sentences, and that means they’ll be placed nearby in the embedding. But “turnip” and “eagle” won’t tend to appear in otherwise similar sentences, so they’ll be placed far apart in the embedding.

Everyone is familiar with seeing data represented on a 2 dimensional chart with Y and X axes. That’s all these “embeddings” are — coordinates in a space, but rather than 2 dimensions, they have thousands. So we can’t really visualize them easily. Humans can’t see more than 3 dimensions so we have to just take a slice of them in 2D or 3D (or otherwise reduce the dimensionality).

In this meaning space, as Wolfram said, similar things are closer together. It works the same way in the text and art models. Alligator and crocodile are close in word space. In image space, an apple and an orange are close together, just as the color red would be close to roses.

What ChatGPT and Midjourney are doing is allowing us to “navigate” this space by using writing and text to explore it. Language models complete the text by finding higher probability combinations nearby your prompt. Art models associate your text prompt with nearby imagery in the space.

Daniel Gross put it similarly in an interview last year with Ben Thompson:

The models are just roaming through this embedded space they’ve created looking for things that are near each other in this multi-dimensional world. And sometimes the statistically most-next-likely thing isn’t the thing we as humans sort of agree on is truth.

He was referring to the negative side here of course, of hallucinations. But what’s a negative for discovering facts actually helps us be more creative.

One of my favorite authors, Kevin Kelly, has a new book out and has been making the podcast rounds recently. Here he is on The Tim Ferriss Show talking about how he uses MidJourney:

It’s like photography. I feel I have some of the same kind of a stance that I have and I’m photographing, I’m kind of hunting, searching through it, I’m trying to find a good position, a good area where there’s kind of promise. And I’m moving around and I’m trying, and I’m whispering to the AI, how about this? I’m changing the word order. I’m actually interacting, having a conversation with it over time. And it might take a half hour or more to get an image that I am happy with. And I’m at that point, very comfortable in putting my name as a co-creator of it because I have, me and the intern have worked together to make this thing.

This is exactly what Kevin is doing. The MidJourney model is generating the image for him, but mostly it’s helping him explore possibility space until he finds something he likes.

Here he is in another interview on Liberty’s Highlights talking about how these models make us more creative and add leverage to our capabilities:

Yeah, the way you think of the kind of the prompters, which are the new artists, is that they’re like directors of a film. They’re commanding a whole bunch of different entities and agencies that do the craft. And their craft is orchestrating creative ideas and they have crafts people who do the cinematography and do the music, the score and everything. And they're kind of curating or directing the thing, or a producer in music. So we now have that kind of on images. And the important thing is that enables many more people to participate in creativity. And I say we have synthetic creativity. These engines are absolutely 100% lowercase creative. They are creative. There’s no getting around it. We can’t deny it. But it's a lowercase, small “c”, creativity. It's not the uppercase creativity that we have, like with a Guernica painting or Picasso or something. So the idea is that those we can't synthesize yet, maybe someday, but not right now. And so that lowercase creativity is synthesized and it's available to people. And that’s a huge innovation.

So this exploration of possibility space is a key feature of generative AI models.

At the same time, it’s pretty clear to me that we need a better way to navigate this space. Right now we’re doing it very haphazardly using trial and error with text prompting. But this doesn’t have to be the case.

Just like a geneticist discovering which genes code for certain traits (maybe it’s a few genes for eye color, or a lot for personality traits), we can discover which neurons and dimensions represent certain areas of meaning space.

This is why further research on how LLMs work is so useful.

If we can tweak parameters that represent attributes like humor, verbosity, complexity of writing… then we can manually adjust them like on a dashboard. You could increase the humor of your writing by 10% and decrease the length by 5%. Same goes for art models. How cool would it be to have slider adjustments for realism, color, detail, mood, or whatever?

I could keep going, but I’ll save it for another time.

1. AI = “Augmented Intelligence”

Rather than Artificial Intelligence, it’s better to think of AI as “Augmented Intelligence”. More than anything, it acts as a multiplier of our own intelligence — just like computers did decades ago, but much more so.

This is Marc Andreessen on a recent episode of Lex Fridman’s podcast talking about the benefits that AI will bring:

Marc: And certainly at the collective level, we could talk about the collective effect of just having more intelligence in the world, which will have very big payoff. But there's also just at the individual level, like what if every person has a machine, and it's the concept of augment —Doug Engelbart's concept of augmentation. What if everybody has a an assistant, and the assistant is 140 IQ, and you happen to be 110 IQ. And you've got something that basically is infinitely patient and knows everything about you, and is pulling for you in every possible way — wants you to be successful. And anytime you find anything confusing, or want to learn anything, or have trouble understanding something, or want to figure out what to do in a situation — right, when I figure out how to prepare for a job interview, like any of these things, like it will help you do it. And it will therefore — the combination will effectively be — because it will effectively raise your IQ will therefore raise the odds of successful life outcomes in all these areas.

Lex: So people below the hypothetical 140 IQ, it’ll pull them off towards 140 IQ.

Marc: And then of course, people at 140 IQ will be able to have a peer, right, to be able to communicate, which is great. And then people above 140 IQ will have an assistant that they can farm things out to.

Packy McCormick wrote a few good essays on this topic recently. In “Intelligence Superabundance” he addresses the concern that AI will steal our jobs and obsolete humans:

. . . the increased supply of intelligence will create more demand for tasks that require intelligence, that we’ll turn gains in intelligence efficiency into ways of doing new things that weren’t previously feasible.

. . . We can examine it a little more rigorously with a little mental flip. Instead of thinking about AI as something separate from human intelligence, we should think of it as an increase in the overall supply of intelligence. . . . Then the important question becomes not “Will I lose my job?” but “What would I do with superabundant intelligence?”

Essentially, demand for intelligence is very elastic. Meaning when either the price goes down or the supply goes up, the demand goes up a lot to match it. If AIs give us more intelligence, our demand and uses for it will increase.

Packy writes more in another essay called “Evolving Minds”:

In that case, I don’t think the framing of AI as simply a tool that we’ll use to do things for us is quite right. I think we’ll use what we learn building AI to better understand how our minds work, define the things that humans are uniquely capable of, and use AI as part of a basket of techniques we use to enhance our uniquely human abilities.

Even as AI gets smarter, it will fuel us to get smarter and more creative, too. Chess and Go provide useful case studies.

He goes on to write about how AI became much better than us at chess and Go, but instead of giving up, humans got better — and more creative — as well:

Some people will be happy letting AI do the grunt work. Not everyone will see AI’s advance as a challenge to meet, just as I have not taken advantage of chess and Go engines to become better at either of those games. But I think, as in chess and Go, more people will get smarter. The process will produce more geniuses, broadly defined, ready to deliver the fresh, human insights required to ignite the next intellectual revolution.

If we can walk the line, the result, I think, will be a deeper understanding of the universe and ourselves, and a meaningful next leg in the neverending quest to understand and shape our realities.

When we think about this augmented intelligence, it’s not necessarily just “ChatGPT” or some other generic AI. Future versions will be very much customized. Not only will they have your data, but they’ll share your system of values. This is Sam Altman on this topic from an interview early this year:

. . . the future I would like to see is where access to AI is super democratized, where there are several AGIs in the world that can kind of help allow for multiple viewpoints and not have any one get too powerful, and that the cost of intelligence and energy, because it gets commoditized, trends down and down and down, and the massive surplus there, access to the systems, eventually governance of the systems, benefits all of us.

. . . We now have these language models that can understand language, and so we can say, hey model here's what we'd like you to do, here are the values we'd like you to align to. And we don't have it working perfectly yet, but it works a little and it'll get better and better.

And the world can say all right, here are the rules, here's the very broad-bounds absolute rules of a system. But within that people should be allowed very different things that they want their AI to do. And so if you want the super never offend safe for work model you should get that. And if you want an edgier one that is sort of creative and exploratory but says some stuff you might not be comfortable with or some people might not be comfortable with, you should get that.

So this all sounds great, but if base-level intelligence is commoditized, how can humans stand out and differentiate? I think there’s a lot of answers to this — some of them maybe even philosophical. I’ll leave those for another time, and just leave you with a clip from Liberty’s interview with Kevin Kelly:

Liberty: And in the book that would be compressed as don't be the best, be the only, this reminds me of — Naval has a way of saying it, which is “escape competition through authenticity”, right? If someone else is doing the exact same thing, well, maybe a robot can do it. Right?

Kevin: Exactly. That's another piece of advice in the book is that if your views on one matter can be predicted from your views on another matter, you should examine yourself because you may be in a grip of an ideologue. And what you actually want is you actually want to have a life and views that are not predictable by your previous ones. And the advantage there is that you are going to be less replaceable by an AI.

Liberty: Yeah. If everybody's thinking the same, a lot of people are not thinking in there.

Kevin: Well, but the AIs are auto-complete. They're just kind of auto-completing. And if you're easy to auto-complete, you're easy to be replaced.

Liberty: That's a good way to put it. Optimize for being hard to auto-complete.

🧑‍🏫 Education Frontier

Education — specifically K-12 education — has a lot of room for improvement. I’ve written a few things about the future of education many years ago, but haven’t really talked about it much. With a 2-year-old now, I follow it a little closer so I want to share some interesting content I’ve seen recently.

Ryan Delk founded a startup called Primer that’s working to build a new education system in the US by helping teachers launch their own micro-schools with 10-25 students that provide more personalized education without all the bureaucracy that comes with big school systems.

He was interviewed recently on a podcast called The Deep End, and here is is giving some more detail on what they do:

So the micro schools today that we have are, they're all led by one teacher as sort of the main micro school leader. Depending on the number of kids, this is part of what's really cool about it is not only can teachers make way more money than they do in the traditional system. So on average, our teachers probably make 40% to 50% more than what they'd make in the existing districts that they teach in. But they also can get an aid to help them out in their class.

And so at a certain threshold, they unlock an aid that will help them and can also sort of assist with different subjects and stuff. So you have one micro-school leader. For elementary, as you mentioned, it's very standard in the US that an elementary teacher teaches all subjects. And at some middle schools or a lot of middle schools, sometimes that's also true, where one teacher's teaching subjects. And so we use that model.

Right now, most of our kids are in elementary, third through fifth, third through sixth grade. And so that's the current model. And then how we augment that is with online, what we call them as subject matter experts. And so we have, for example, a really exceptional ELA and a really exceptional reading teacher, really exceptional math teachers. And so for specific subjects where kids, whether it's that they need remedial help, or they're advanced and they need more support, and the micro school leader can't support them directly or just isn't set up to support them directly because of the structure of the class or whatever, kids have these resources.

This brings us to the question of — what is school actually providing to students?

I think a lot of people that are in the education sort of startup space have this sort of like, oh, we're going to take over the world and disrupt everything. And, you know, we need to blow up the whole existing system. My view on it is maybe a bit more nuanced than that. And I think that the public school system in the U.S. right now, I think it serves like three very distinct functions. And we've bundled all that into one name, which is like public schools.

Everyone thinks about the academic component. Hey, we're sort of this is the engine that is responsible for the academic development of the children of the U.S. And that's certainly one component of it. And we could debate how effective it is at that. But another component that's sort of the two other components I think are maybe equally important societally, but are very under discussed.

One is that for millions of kids, it's actually basically the way that they're able to get food and have food security each day. And we sort of totally dismiss this, but it's actually, I think, a very important and very sad and unfortunate sort of component of public schools. You can do the numbers and the research on this, and it will blow your mind.

And the second component, the third component is that, you know, most parents are working and they need a place for their kids to go from you know eight to, with aftercare, eight to five every day so they can work. And so there's this babysitting component of school that's very important.

This reminds me of the famous Good Will Hunting scene (and meme):

As we all know, especially with the internet, you can learn everything they teach you at Harvard yourself. But obviously that’s not the real reason why people want to go to Harvard. Just like universities, all school is a bundle of services:

• Learning

• Daycare / babysitting — even if a computer could handle 100% of education, you wouldn’t leave a 7 year old alone all day while you work.

• Social connections

• Tutoring

• Working with others

• And as Ryan mentioned, providing food.

So it’s pretty clear we’re going to have somewhat of an “unbundling” of these services over time. Into what, I don’t know. What I do know is that AI will play a bigger and bigger role going forward.

Sal Khan did a TED talk earlier this year where he talked about The 2 Sigma Problem, a study from the 1980s where if you give one-on-one tutoring to students, you can improve their outcomes by 2 standard deviations. This takes an average student and turns them into an exceptional student. The problem here of course is that this is very hard to scale.

Enter AI. Khan Academy has a new feature called Khanmigo that integrates GPT-4:

This is really going to be a super tutor. And it's not just exercises. It understands what you're watching. It understands the context of your video. It can answer the age-old question, why do I need to learn this? And it asks, socratically, well, what do you care about? And let's say the student says, I want to be a professional athlete. And it says, well, learning about the size of cells, which is what this video is about, is really useful for understanding nutrition and how your body works, et cetera. It can answer questions, it can quiz you, it can connect it to other ideas. You can now ask as many questions of a video as you could ever dream of.

So another big shortage out there, I remember the high school I went to, the student-to-guidance counselor ratio was about 200 or 300 to one. In a lot of the country, it's worse than that. We can use Khanmigo to give every student a guidance counselor, academic coach, career coach, life coach, which is exactly what you see right over here.

With this, students can get truly custom learning experiences. That’s pretty awesome to me because in school there can be more hands-on activities and applications. Things that are hard to beat on a computer.

In an interview on Jim O’Shaughnessy’s podcast “Infinite Loops”, Ethan Mollick has a great bit about this:

So, I think in much of the world, we're going to see a real transformation of teaching, where teaching is not high quality right now and we're going to see a leap into another model.

But in the US, in Europe and probably in China, Japan, other places with that already have fairly highly developed educational systems, I think the future looks like one where you are going to still go to school as you did before, during the same class period times, but you're going to be basically... The easiest version to think about, is you're going to be doing a much more engaging version of homework in school. Project-based learning, experiences, building stuff, trying to create things, trying to apply things. Outside of school you're going, your homework assignments are going to be monitored by AI and you'll be working with an AI tutor that will generate lessons for you and teach you concepts, and make sure you're on track and will let the teacher know "The person's struggling at this, despite this," and will help the teacher set up teams so that you don't have those horrible team assignments and track grades. So, I think that there's a model that we already know works. We'll still give lectures. It's like 5,000 years old, giving lectures, and we know that it doesn’t work, but we keep doing them. So, it's going to force us to shake ourselves out of a bunch of stuff that doesn't work well. I'm sure there'll still be some classes that look like classes. I think English composition will still look like English composition. I think math is still going to look a lot like math, but I think a lot of your other projects will be more...

I’ll finish this section with another clip from Ryan Delk on blending the virtual and real-world experiences:

So I think of it as the best possible experience is absolutely a best teacher in person in the smallest possible environment, meaning the fewest amount of students to the teacher ratio. That's the best. That's by far the best situation that you could get any kid into. But that's sort of like a pipe dream for most kids in the US right now, that you're just not going to have this amazing world-class teacher that can sit down with 12 kids and go deep about some specific thing you're excited about. So I think I agree that that's the optimal scenario.

I'm not one of these people that thinks virtual is inherently better. But I do think that virtual with a world-class teacher, especially if you can build interesting product experiences, which we're building on top of the virtual experience, when needed, and accompanying an in-person teacher, is better than a sort of average in-person experience.

🔗 Interesting Content

Recent content I’ve enjoyed and think you might also.

David Senra on Steve Jobs

One of my favorite recent episodes of The Founders podcast was where David talked about the recently released book “Make Something Wonderful” which is a selection of Steve Jobs’ writings.

That is what your life is. Think of your life as a rainbow. This is again, maybe my favorite quote in the entire book. Think of your life as a rainbow arcing across the horizon of this world. You appear, have a chance to blaze in the sky, and then you disappear. What is he saying? Go for it like your life depends on it. Why? Because it does. It literally does. The two endpoints of everyone's rainbow. This is, this is again, I think one of the, he's got a ton of great ideas, right? But something you'll, if you study Steve, what he says over and over again, it's like if you need to use your inevitable death as a tool to make decisions on how you're going to spend your time, the two endpoints of everyone's rainbow are birth and death. We all experience both completely alone. And yet most people of your age have not thought about these events very much, much less even seen them in others.

How many of you have seen the birth of another human? It's a miracle. And how many of you have witnessed the death of a human? It is a mystery beyond our comprehension. No human alive knows what happens to us after our death. Some believe this, others that, but no one really knows. Again, most people your age have not thought about these events very much. It's as if we shelter you from them, afraid that the thought of mortality will somehow wound you. For me, it's the opposite. To know my arc will fall makes me want to blaze while I'm in the sky. Not for others, but for myself, for the trail that I know that I'm leaving. And I love the way he ends his speech. Now, as you live your arc across the sky, you want to have as few regrets as possible. Remember, regrets are different from mistakes.

Lots of good wisdom in this one, so make sure to check it out. I’ll call out his just-released one on Arnold Schwarzenegger as well, which is a good compliment to the Netflix documentary I highly recommend.

Thomas Edison’s Lab

Eric Gilliam wrote a great essay for Works in Progress on Thomas Edison. Highly recommend this and I got a lot out of it. He also did a companion solo podcast called “Tales of Edison’s Lab”. Edison was one of those once-in-a-century type people that to me really just fits the American “man in the arena” prototype.

In a field full of uncertain potentials, he often did seem to proceed to his triumphs by a certain kind of laborious sorcery, ill prepared in mathematics, disrespectful of the boundaries set by available theory. He took on all the forces of nature with a combination, as he said, rather unfairly about himself, of 2% inspiration and 98% hard work. The inspiration was informed by years of practice and shrewd observation. The hard work was simply amazing. And if the means by which he brought off his extraordinary efforts are not wholly clear, neither is the cause for his obsessive labors, no diver into nature's deepest mysteries, caring next to nothing for the advancement of knowledge and even less for the world's goods. He would become absorbed in making something work well enough to make money. The test in the marketplace was for him, apparently, the moment of truth for his experiments. And once this was passed, he became forthwith, absorbed in making something else. By the time of Edison's foray into lighting, he was a commercially successful inventor many times over.

Acquired on Lockheed Martin

Ben and David from the Acquired Podcast recently did an episode on Lockheed .

It’s a 3+ hour episode so there’s a lot there. They cover Kelly Johnson’s creation of Lockheed’s famous Skunkworks division:

So in this circus tent in a parking lot, Kelly and this super elite team from Lockheed build the first prototype US fighter jet named the Lulu Bell in 143 days start to finish. This is just wild. For years the US had been working on this technology and they hadn't gotten it operationalized. The Germans beat them to it and then in 143 days, Kelly and Lockheed go from zero to flying prototype. What a testament to him and to this organization in the circus tent that he has built, the Skunkworks.

I used Skunkworks as an example a few times in my essay and podcast from last year, Take the Iterative Path on how Skunkworks and SpaceX both use iterative design to build so quickly. So check that out if that interests you.

On Acquired they also talk about Lockheed’s space and missile division, which I had actually never heard of. And apparently it did pretty well for them:

Well, I think at times in the 60s and 70s and 80s, LMSC was the largest business by revenue. But almost through the whole time, it was by far the most profitable division within Lockheed. And at times, when we'll get into, Lockheed fell on some really hard times in the 70s. There were years where LMSC generated more than 100% of the profits of Lockheed. So all of the rest of Lockheed, Skunkworks included, was in the red. Unprofitable, bleeding money, and LMSC was keeping the company afloat.

. . .This is much more technology problems and computing problems that LMSC is tackling here. Yes, they're building missiles. Yes, they're building rockets and all that. But the core value components of those rockets is computing and silicon and ultimately software. And as we talk about all the time on this show, like, well, that's really good margins. Definitely better margins than building airplanes.

Apple Vision Pro

There’s been lots of good content out — from testers and analysts — about Apple’s Vision Pro mixed reality device.

As usual, Ben Thompson had a good writeup and was really impressed with the half-hour demo he got:

As I noted above, I have been relatively optimistic about VR, in part because I believe the most compelling use case is for work. First, if a device actually makes someone more productive, it is far easier to justify the cost. Second, while it is a barrier to actually put on a headset — to go back to my VR/AR framing above, a headset is a destination device — work is a destination.

. . . Again, I’m already biased on this point, given both my prediction and personal workflow, but if the Vision Pro is a success, I think that an important part of its market will to at first be used alongside a Mac, and as the native app ecosystem develops, to be used in place of one. To put it even more strongly, the Vision Pro is, I suspect, the future of the Mac.

I agree with Ben’s belief that Vision is ultimately set up to replace some segment of the Mac’s jobs-to-be-done, and ultimately become an amazing productivity tool.

Just as websites on smartphones were originally the same desktop version, Apple Vision shows users interacting with several 2D screens floating in space. This alone can be a huge productivity boost, as evident by the boost you get from multiple monitors today.

But the real boost will come when the mixed reality interface breaks out of the past paradigm and you can interact with both the physical world and an infinite digital canvas. Andy Matuschak wrote about this following the announcement, talking about ideas like having apps or files live in physical space. Like a timer above your stove, or the Instacart app in your fridge.

It will be interesting to see what kinds of stories and experiences come out of the mixed reality world. Steven Johnson, who’s one of my favorite writers, did an interview on VR and AR right before the Vision Pro announcement. One of my favorite parts was when he talks about the history of immersive experiences:

There's this wonderful period, I don't know, like 250 years ago, where there was this amazing explosion of immersive, I call them kind of palaces of illusion. And they were everything, you know, there's a wonderful example of like the panorama, which was the 360 degree painting where you would come in and see a famous military battle reenacted or a skyline view of London and things like that. And there was just a huge number of experiments that happened back then, where they would try all these different configurations of ways to like, create the illusion of being in some other world.

Interestingly, they were not narrative environments, they really didn't tell a story. It was just the kind of like the body, the sensory feeling of, oh, I'm in this like other world, and I can occupy that world for a bit of time and then go back into reality. And that there was just, again, a lot of mucking around here for, you know, figuring out ways to kind of captivate people.

And it was kind of like a Cambrian explosion of all these different possibilities that then eventually consolidated basically into the cinema, right? There are all these different ways of creating the illusion that they're off in some other world. And then we kind of settled on this one approach and all those other experiments died off, basically. And the other thing that happened in that period is directly related to VR and AR is the invention of the stereoscope, which was one of these great cases where it was a scientific innovation about stereo vision.

That’s all for now — I hope you learned something interesting from this experiment!

Time for a new roundup. Some housekeeping first:

I’ve now consolidated futureblind.com, my blog archive, and all past newsletter subscriptions onto Substack. So if you’re seeing this newsletter for the first time, glad to have you!

It was a pain dealing with posts in multiple locations, and the previous FutureBlind host charged too much for such a basic set of features and lack of footnote capability (sooo annoying). Substack has been convenient and easy to work with the past 3 years, and continues to deploy more writer-friendly features. As of 4 months ago, FutureBlind now has a paid tier, but — spoiler alert — there’s nothing special for paid members yet. I may end up putting only unpolished, brain-dump-style thoughts on the paid edition. Pay to remove my media filter. Oh and they also have Twitter Notes now — check it out!

Onto the show.

In this roundup:

• 🎨 Essay: The Grand Unifying Theory of Design — Design isn’t just about how something looks. It’s a universal problem-solving process that can be applied to any discipline to create valuable and delightful experiences.

• 🤖 A.I. Frontier

  • Essay: Addressing any remaining LLM skepticism — Is there anyone still skeptical of what existing LLMs will be capable of?

  • New interfaces of generative AI — What comes after chatbots? How else can we collaborate with AI?

  • How will generative AI help us tell better stories? — Cheaper AND better quality from stories to post-production.

  • Agents are the future — Are goal-seeking LLM agents the future of productivity?

• 🚀 Space — Go time for Starship; Eric Berger summarizes the next 10 years in space.

• 🔗 Interesting Links — Glotz builds Starbucks; Is tradition true?; Availability cascades; How ChatGPT works; Making computers that can smell; and more.

🎨 The Grand Unifying Theory of Design

Design is creating something that effectively solves a problem in a specific situation.

That’s not how most think of design though. People generally associate it with visual aesthetics or how something looks. For some there’s an aura of mysticism around it, as if it's an elusive art form that only a few can master. This narrow view of design misses how universal the core principles are.

There are a lot of design sub-disciplines. But if you think about big picture “Design”, it’s really about making it so something solves a problem well. In this way, design encompasses basically anything you do that provides value to others. A well-designed experience solves a problem and makes people feel delighted in the process.

Everything you do can benefit from understanding the fundamentals of design.

Continue reading:

FutureBlind

The Grand Unifying Theory of Design

"In most people’s vocabularies, design means veneer. It’s interior decorating. It’s the fabric of the curtains of the sofa. But to me, nothing could be further from the meaning of design. Design is the fundamental soul of a human-made creation that ends up expressing itself in successive outer layers of the product or service." …

Read more

3 years ago · 3 likes · Max Olson

🤖 A.I. Frontier

“When I think about generative design, I think about an exploration toward an optimal solution in context. Meaning that the thing that automation never captures is a value system — no software is ever gonna fully capture what all the people involved in that project want.” Anthony Hauck (founder of Hypar) in an interview with Brian Potter

A few weeks ago I wrote a short essay that addresses anyone that may be still skeptical of what existing LLMs will be capable of. Check it out:

FutureBlind

Addressing any remaining LLM skepticism

“The future is here — it’s just not evenly distributed.” — William Gibson This essay is addressed to anyone still skeptical of what modern AI will be capable of. If you’ve already drunk the Kool-Aid and agree with the tweet below, you may not need to read on…

Read more

3 years ago · Max Olson

New user interfaces of generative AI

Chat and prompting interfaces will clearly be a common way to interact with generative AI models going forward. But what about other ways to collaborate with these models?

I suspect everyone is defaulting to chat UIs because it’s the easiest and most obvious thing to make. Like command line interfaces (C:>) before graphical user interfaces. Or when motion picture cameras were invented and, lacking the proper grammar, they just filmed stage plays.

We’re still in the nascent days, but I’m starting to see some good experimentation:

• Recently released beta for Adobe Firefly starts to get at some of these, but it looks like they still use text prompting for a lot of manipulation.

• Github Copilot Code Brushes — Apply “brushes” (really just language model transformations) to your code to do things like make it more readable, fix bugs, etc. Github Copilot itself was an innovative interaction UI, only appearing in context as you write code. Although in my personal experience I think they could do a better job allowing users to explore the possibility space more.

• Many use cases of generative AI models are essentially just tools to help explore high-dimensional possibility space. (I hope to have an essay in the near future about this.) Linus Lee has been one of the best writers about this area:

  Large language models represent a fundamentally new capability computers have: computers can now understand natural language at a human-or-better level. The cost to do this will get cheaper over time, and the speed and scale at which we can do it will go up quickly. When we imagine software interfaces to harness this language capability for building tools and games, we should ask not “what can we do with a program that completes my sentences?” but “what should a computer that understands language do for us?”

  Here he is on AI-augmented creation:

  I was talking earlier today about how you can view a creation process not as additive (starting with a sentence, and then adding another and another) but as iterated refinement and filtering through an infinite option space (there are an infinite continuations of your first sentence. How do you choose the right continuation?) Because LLMs can explicitly compute all possible continuations, this is an interesting way to look at writing with an AI, and AI-augmented creation in general.

  Linus now works for Notion, which has already done a lot of interesting experimentation with language UIs, and I’m excited to see what they come up with next.

• Sudowrite is working on innovative writing interfaces, particularly for storytelling and fiction.

• Geoffrey Litt writes a long and fascinating essay on his speculations of how LLM interfaces will progress. Highly recommended if you have any interest in the future of product design.

  As Alan Kay wrote in 1984: “We now want to edit our tools as we have previously edited our documents.” [. . .] I think it’s important to notice that chatbots are frustrating for two distinct reasons. First, it’s annoying when the chatbot is narrow in its capabilities (looking at you Siri) and can’t do the thing you want it to do. But more fundamentally than that, chat is an essentially limited interaction mode, regardless of the quality of the bot. [. . .] When we use a good tool—a hammer, a paintbrush, a pair of skis, or a car steering wheel—we become one with the tool in a subconscious way. We can enter a flow state, apply muscle memory, achieve fine control, and maybe even produce creative or artistic output. Chat will never feel like driving a car, no matter how good the bot is.

Agents are the future

As people rapidly experiment with what language models are capable of, new patterns and abstractions emerge that enable more sophisticated and versatile applications. Agents are one of these patterns.

An agent carries out a specific purpose using reasoning from a language model and whatever tools it’s been given to do the job.

“Programmers will command armies of software agents to build increasingly complex software in insane record times. Non-programmers will also be able to use these agents to get software tasks done. Everyone in the world will be at least John Carmack-level software capable.” — Amjad Masad on Twitter

The ReAct pattern proposed in this paper last year is the most basic example: to answer the original prompt, the model Reasons through chain-of-thought and is given tools it can use to take whatever Actions necessary.

An agent is analogous to a function in programming, but unlike a function its outcome isn’t deterministic. It has goal-seeking behavior and is essentially searching a landscape of possibilities, the terrain of which is defined by the prompt input.

When you combine multiple agents with other capabilities like shared memory management you can create some very interesting patterns, a few of which have been heavily buzzed about the past few weeks:

• Auto-GPT — An autonomous agent that recursively creates tasks and seeks out a goal. Here it is solving a coding problem. Someone put up a basic UI around it here.

• BabyAGI (Task-Driven Autonomous Agent) — A simple task management system that breaks a goal down into tasks, working on them 1 by 1.

The agent pattern is multi-scale, meaning you can have an agent that uses other agents, and so on. An agent can be very generalized like AutoGPT, which tries to solve whatever problem you give it using other agents in the process. Or it could be more specialized, like a data agent that does nothing but query a database until it finds an answer.

“1 GPT call is a bit like 1 thought. Stringing them together in loops creates agents that can perceive, think, and act, their goals defined in English in prompts. For feedback / learning, one path is to have a ‘reflect’ phase that evaluates outcomes, saves rollouts to memory, loads them to prompts to few-shot on them. That is the "meta-learning" few-shot path. You can "learn" on whatever you manage to cram into the context window.” — Andrej Karpathy

I’ll leave it there for now — but this is a super interesting area to follow. I’ve been experimenting with this myself, both for personal use and on internal company projects, and I’ll likely write more about it later.

How will generative AI help us tell better stories?

“If you can generate an original screenplay with vivid details, characters and lines using ChatGPT (you can today), anyone’s voice using just 15 seconds of recording (available today), and can generate original video using text-to-video prompts (within months we’ll see models with indistinguishable video generation capabilities), you can generate a spin-off or sequel on your own to any mainstream TV show or movie.” — Scott Belsky, CPO of Adobe

The ability to create quality media at a low cost will continue to get better. Generative AI can be used at every level of the media-creation stack — from story generation to final touches in post-production. This is another area I’m closely following and am very excited about.

The potential is nearly endless. What company will be the Industrial Light & Magic of this new era?

If you’re as interested as I am, check out the following:

• Runway.ml — A creative tool suite with generative AI at its core.

• More innovation from Corridor Crew: 📹 Did We Just Change Animation Forever?

• “Idea Olympics” by Packy McCormick — Packy wrote this sci-fi story with the help of ChatGPT and Midjourney. The model was a true collaborator, riffing off of ideas, writing words, and providing inspiration.

Subscribe now

🚀 Space

Go time for Starship! You’re not going to want to miss this launch…

SpaceX has completed the final “flight readiness review” for the Starship launch system, with a targeted test flight set for April 17, pending regulatory approval. Soon they’ll perform a launch rehearsal to increase confidence in the fueling process.

During the actual test, the Super Heavy rocket will separate from the upper stage and make a controlled descent into the Gulf of Mexico, while the Starship upper stage aims to reach orbital velocity before reentering the atmosphere over the Pacific Ocean and landing vertically north of Kauai. The test flight’s primary goal is to assess the rockets, their engines, and the system’s capability to reenter Earth’s atmosphere and make a controlled landing.

Let’s hope the FAA gives them clearance so we can see a launch in the next few weeks.

👉 Follow Everyday Astronaut’s Starship orbital checklist to keep up

Why is Starship such a big deal? See my post “The Future of Space: Part I”

The next 10 years in space

From the launch of the James Webb Space Telescope to the rise of SpaceX, the past ten years have marked a period of dramatic change and growth in the space industry. With the continual development of new technology and exploration initiatives, the industry has seen incredible advances, yet also plenty of controversy, cost overruns, and delays. As we move into the 2020s, the industry is at a pivotal point, with the potential to make major breakthroughs in exploration, transportation, and more.

Back in January, Eric Berger wrote about what he’ll be watching for over the next 10 years:

1. Commercial space stations — Focus turns to finding a future destination for NASA astronauts in low-Earth orbit as the International Space Station nears the end of its lifespan. NASA has awarded funding to various companies, including Axiom Space, Blue Origin, Nanoracks, and Northrop Grumman, for the development of commercial space stations, with the hope of avoiding a gap in low-Earth orbit presence.

   When the stations become operational, the space agency will likely be willing to spend about $1 billion procuring commercial station services annually. But will any private stations be ready to go when NASA is ready to buy? This great space station race is one to watch during the next decade.

2. NASA’s Artemis program — The Artemis I mission successfully demonstrated NASA’s potential for future lunar missions, but the program faces technical and financial challenges. As NASA progresses with Artemis II and III, it must navigate setbacks and competition from China’s ambitious space program.

   This is a huge financial, management, and technical challenge, and quite frankly, it’s not 100 percent clear that the space agency and its contractors are up to the task. [. . .] Watching how all of this plays out—how NASA handles delays, setbacks, and international competition—will be one of the compelling storylines of this decade.

3. Decluttering low-Earth orbit — The number of satellites launched annually has significantly increased, raising concerns about space junk and orbital sustainability. Addressing this complex issue requires careful legislation and cooperation among space powers to prevent potential tragedies in low-Earth orbit.

4. Is commercial space sustainable? Post-SpaceX, the commercial space industry faces financial and technical challenges, raising concerns for its viability. Companies like Blue Origin and Axiom Space must secure funding and overcome technical hurdles to achieve profitability and meet their goals.

5. Will Starship actually work? SpaceX’s ambitious Starship program aims to revolutionize spaceflight with reusable rockets and the ability to carry larger payloads at a lower cost. While many challenges lie ahead, a successful Starship would disrupt the global launch industry and significantly alter the future of space exploration and commerce.

🔗 Interesting Links

From the archives ~ When transferring FutureBlind to Substack, I found a few still-relevant posts you might find interesting:

• Glotz Builds Starbucks (2012) — A riff of Charlie Munger’s essay on Glotz building Coca-Cola. It’s 1980: You’ve just inherited Starbucks, a small Seattle business that roasts and sells coffee beans in 4 locations. How would you build it into a company worth more than $30 billion in 30 years?

• Is the Internet Ruining Media? Hardly. (2008) — I guess not much has changed in 15 years when it comes to our fears of technology replacing what we love.

• Quality Without Compromise (2007) — See’s Candies, Warren Buffett, and the perfect investment.

Other great content I’ve enjoyed over the past 3 months:

• Erik Torenberg: “Tradition is Truer than Truth” — Just as in biological evolution, traditions are selected for over time as they compete for survival. But are they true? Can they be proven? Sometimes it might not matter. If a tradition or rule-of-thumb has been around for a long time, it’s probably useful — even if we don’t know the use.

• Morgan Housel: “How It All Works (A Few Short Stories)” — A few short stories with good lessons: Everyone has different tastes, but emotions are universal; People love familiarity; Everything is for sale; Consistency beats intelligence; and more.

• Marc Andreessen: “On Availability Cascades” — Availability cascades occur when a social cascade is triggered by an available topic, regardless of its importance, leading to widespread public discourse. They are driven by availability entrepreneurs, who seek to advance their own agendas by triggering cascades around specific problems or topics.

• Wired: “This Startup is Using AI to Unearth New Smells” — On Osmo, a new startup with a vision of allowing computers to smell. This is much much harder to do than computer vision, but once you give computers the power to smell it could be even more consequential than allowing them to see.

• Gary Sheng: “Meme Lords Deserve More Respect” and “The Case for Honest, Participatory Propaganda” — Honest propaganda conveys truth or sincere opinions without deception, while participatory propaganda aims to build grassroots movements and empowers individuals to contribute. By utilizing honest, participatory propaganda, people can effectively share their ideas and compete with larger entities in the battle for attention.

• Stephan Wolfram: “What is ChatGPT doing… and Why Does it Work?” — A very long and detailed essay — but still easy to understand — on what the models behind ChatGPT are doing behind the scenes.

• Huge photo archive of buildings under construction (h/t Construction Physics)

Design is creating something that effectively solves a problem in a specific situation.

That’s not how most think of design though. People generally associate it with visual aesthetics or how something looks. For some there’s an aura of mysticism around it, as if it's an elusive art form that only a few can master. This narrow view of design misses how universal the core principles are.

There are a lot of design sub-disciplines. But if you think about big picture “Design”, it’s really about making it so something solves a problem well. In this way, design encompasses basically anything you do that provides value to others. A well-designed experience solves a problem and makes people feel delighted in the process.

Everything you do can benefit from understanding the fundamentals of design.

Making something look good can be important, but it’s usually the last step. The first step is understanding the actual problem you’re trying to solve.

Design: from the top down

Design is fitting form to context

The first thing to understand is that all design is relative to the context it’s used in.

The goal of all design is “goodness of fit”: how well the form (the artifact or thing you’re designing) fits in the context (problem or situation it’s being used in). The real object of discussion is not the form alone, but the unity of both form and context.

In his book “Notes on the Synthesis of Form,” Christopher Alexander gives an example of designing a tea kettle:

Let us consider an ensemble consisting of the kettle plus everything about the world outside the kettle which is relevant to the use and manufacture of household utensils. There's a clear boundary between the kettle because it happens to be a physical object. But one could easily make changes in the boundary — implying that the kettle is the wrong way to heat domestic water to begin with. It can quickly be expanded to include the entire house, claiming that it is not the kettle but the method of heating kettles that needs to be redesigned. In this case the kettle becomes part of the context, while the stove perhaps is form.

Boundaries between the form and the context are fuzzy and movable. Anything in the world that makes demands of the form is context. Boundaries also overlap. A sink exists in the scope of a bathroom, overlaps with the context of a shower, which is in a house, which is in a community. You design a sink to fit with what’s needed in a bathroom, but also in someone’s life. You need to understand what level of abstraction you’re designing for.

How do you define good fitness? This is one of the things that makes design hard.

The only surefire way is exposing the design to the real world and observing how well it solves the problem. Or through experimental trial and error during the design process (see my essay Take the Iterative Path on why this is needed).

Sometimes trial and error isn’t possible or is too costly. Once you build a house it’s too late to know if it’s designed properly! And even then, how well it fits its occupant’s needs may only be known after many years of use.

For this reason, good fit is generally expressed through minimizing misfit. Misfit between form and context is much easier to spot. Like a kitchen that’s hard to clean, rainwater leaking into the building, or a button that no one understands the purpose of. Avoid misfit (bad design) and you’re well on your way to making it good.

Managing constraints

All mediums are defined by their constraints: from art to comics, novels, films, and shows. In “Understanding Comics”, Scott McCloud explains how storytelling is just translating a story to our senses through their chosen medium. The storyteller’s job is to manage and work around their constraints.

Design is no different. Most of the practice of design is managing trade-offs from constraints.

• Constraints of context. Where are your boundaries? What are the factors imposed by the context?

  Take the tea kettle example from above. There are lots of different factors for the form (ease of holding, retaining heat, etc.). But you’re already drawing a boundary around a specific device for solving the problem. What other method can you use to get hot water? What if hot water came from the tap? More on this in the next section.

• Constraints of form: What interfaces are possible? What’s required of them?

  Is a complicated user interface good, or bad design? It depends on the context and the nature of the interface. Take Microsoft Excel. Anyone exposed to it for the first time would have a very hard time figuring out how to do much. But that’s fine! Excel is mostly for power users. The functionality and learning curve are important parts of the product.

  Excel is also designed specifically for use on a laptop or desktop computer. This gives it specific constraints. You can’t just port the same product to a mobile device because the constraints (and context) are different.

• Constraints of reality: What is possible to make? How is it going to be made? (You might think this doesn’t need to be said. Many engineers I know would beg to differ.)

  The best designers know exactly how their design is built, and the trade-offs inherent in the process of creating it. To paraphrase Warren Buffett: “I’m a better designer because I’m a maker, and a better maker because I’m a designer.”

• Constraints of resources: What do you have to solve the problem? The reality is that the resources you have determine what can be done.

  • Financial and human resources — how much money and people do you have?

  • Time — how much time do you have? How do you know when it’s good enough?

  • Design tools — what tools do you have to assist the design process? (Designing a building using sophisticated CAD tools can have radically different constraints than pencil and paper.)

The key to good design is managing all the trade-offs imposed by these constraints. And there can be a lot of them. In a real-world design problem, it’s not possible to solve for every variable. You’ll have to balance and prioritize the tradeoffs, which will mean compromising on some and finding innovative solutions to others.

Finding the boundaries of context

“Design begins by asking, who is this for and what do they need from it? A good architect, for example, does not begin by creating a design that he then imposes on the users, but by studying the intended users and figuring out what they need.” — Paul Graham

How do we define context and what the design boundaries are? What method do we want to use to solve the problem?

This is where jobs-to-be-done theory comes in. There are entire books written on this so I’ll try to be brief.

When you use a product or service, you’re “hiring” it to do a job for you. A “job” is the progress that a person is trying to make in a particular context. There is a gap in what they want vs. what exists, and the job fills it.

A job can only be defined relative to the specific context in which it arises. In other words, a job is the ensemble of form + context. It can be phrased in the statement:

“When I’m [context], I want to [motivation], so I can [outcome].”

There are many methods that can be hired to do the job. If the problem exists, people are likely already hiring something to make progress on it. The job people are hiring something for may not be obvious, and it can take a lot of work to figure it out.

Continuing our tea kettle example — if the outcome is having hot water, this could be done by multiple methods like heating a kettle over the stove, getting it directly from the faucet, or maybe using a microwave. Methods may not seem to be related at all, like:

• The job of having a business meeting can be filled by an airplane or by Zoom.

• The job of entertaining oneself can be filled by watching TV, playing video games, reading a book, or going for a hike.

These examples are higher-level, less specific jobs. A well-defined job is multilayered and complex. It has needs that are:

1. Emotional — how it makes them feel.

2. Social — how they believe they’re perceived by others while using it.

3. Functional — the practical and objective requirements.

When you know the job, you can choose what method that’s best to do it. (Or you may be constrained to a specific method to solve the problem.)

Once a method is determined, you can design the interface.

Designing interfaces

An interface is the medium a person interacts with.

The goal of interface design is to match the conceptual model (how it works) with the user's mental models (how they think it works). The designer needs to communicate to users, but can only via the user interface.

“A good conceptual model allows us to predict the effects of our actions.” — Don Norman

A conceptual model is the actual model of how it works given to the person through the interface. If the interface doesn’t make the conceptual model clear and consistent, it won’t match with the user’s mental model and problems will ensue. This is “bad” design — in other words, a misfit between form and context.

If you make a tea kettle that doesn’t have a handle, its user — confused, no doubt — will grab it whatever way they can, possibly spilling water or burning themselves. Misfit!

“Interfaces, as Geoffrey James has said, should follow the principle of least astonishment. A button that looks like it will make a machine stop should make it stop, not speed up.” — Paul Graham

How do we make a well-fitting user interface? Ultimately we only know by observing how it’s actually used.

But there are some basic principles of interaction we can use. Don Norman defines some of these attributes:

• Affordances — “The perceived or actual properties of a design that suggest how it can be interacted with.” In other words, what does a design allow for? A handle affords to be gripped. A button affords to be pressed. A ball affords to be rolled. A door with a flat surface affords to be pushed. (Even though sometimes in needs to be pulled. Misfit!) A product affords to be hired for a job-to-be-done. (Even if the job wasn’t intended by its product’s creator.)

• Signifiers — Any visual mark, sound, or perceivable cue that indicates the appropriate behavior. If affordances define what actions are possible, signifiers specify how people discover those possibilities, and how the actions should take place. Labels, diagrams, arrows, audio alerts.

• Mappings — The relationship between controls and their effects, such as the spatial or functional relationships between buttons and the actions they trigger. The classic example here is the knobs that turn on a stove. How do you know which knob controls which stove?

• Feedback — The information provided to the user that indicates the results of their action. Not receiving continuous feedback about the results of actions can result in misfit between intention and outcome. Systems are more usable when they clearly indicate their status, the possible actions that can be performed, and the results of those actions.

How do we effectively apply these principles?

This is where patterns come in.

A pattern language

Designers communicate to users using a language of patterns, which are functional solutions to problems that exist at every level of abstraction.

Every design is composed of patterns. Each pattern addresses one or more needs of the design as a whole. A button on a machine or in software allows a person to initiate an action.

There are commonly used patterns in all design disciplines. Here are a few examples from different areas:

• Software UI patterns — Buttons, type layouts, tabs, checkboxes, more complex patterns like wizards, drag and drop, or complex tables with filters. Design frameworks like Bootstrap provide a collection of tried and tested patterns that are known to work well in specific contexts.

• Architecture patterns — Doors, entries, kitchen layouts, stairs, moulding, windows. Read "A Pattern Language" for much, much more. Corridors (hallways) seem obvious but the pattern needed to be invented.

• Book patterns — Headings, chapters, lists, table of contents, index, paragraphs, elements of typography.

• Film grammar — Every medium uses its own composable grammar that’s analogous patterns. In film, you have grammar like: Wide shot, closeup, slow-motion, zoom, split screen, jump cut, flashback.

• Food patterns — All prepared food is designed. Even the non-manufactured "natural" kind. There are patterns in flavor, texture, visual appeal, and how they’re all combined.1

Patterns don’t have to be universal — they can be unique to a small niche or even a single design. Common patterns can help though because if people are familiar with them the interface will match their mental model better.

I think about it through the lens of natural selection. In the global evolution of design, patterns are the unit of selection, similar to genes in biological evolution or memes in cultural evolution. Patterns that successfully solve a problem in a specific context are reused and “survive” over time to become core parts of a global design language. Like buttons in software, hallways in architecture, salt in food, chapter headings in books, and a closeup shot in film. A design combines many patterns and the more effective it is at doing its job at an efficient cost, the more “fit” it is.

It's important to note that just because a pattern is in common use or generally solves a certain problem, doesn't mean it's right for a particular design.* Every design uses its own pattern language. The language for a particular design becomes a sort of code that helps the designer achieve fitness more effectively.2

Patterns occur at every scale. The flow of navigation in a software app down to the look of a button. The layout of rooms in a home down to the decorative flourishes on a moulding. Whether or not it’s conscious, design patterns are inevitably chosen at all levels of abstraction. And once again, depending on the context of the situation, different scales can matter more than others.

For some designs, visual appearance is key to its usability.

The visceral level

“People DO judge a book by its cover.” — Mike Markkula (original investor in Apple)

Finally we come to visual design.

Don Norman calls this the “visceral” level: the automatic, pre-wired, gut-feeling you get from experiencing a design with your senses. Visceral design is all about immediate emotional impact. It has to feel good and look good.

The Aesthetic-Usability Effect is a phenomenon in which people perceive more aesthetic designs as easier to use than less aesthetic designs (whether they are or not). Designs that make you feel good are easier to deal with and produce more harmonious results. Think Apple products, sports cars, Slack vs. Microsoft Teams, or a book with good typesetting.

Making something look good can be a big part of usability. When you wash and polish your car, doesn’t it seem to drive better?

Much of this effect is due to our evolutionary programming. Positive affect comes from things throughout history that have offered food, warmth, or protection: warm, comfortably lit places; sweet tastes and smells; highly saturated hues; soothing sounds and rhythms; smiling faces; symmetrical objects; rounded, smooth objects. Negative affect comes from things like: heights; unexpected loud sounds or bright lights; darkness; empty, flat terrain; crowds of people; rotting/decaying smells; bitter tastes; sharp objects.

As always, it depends on the context. No design is inherently good or bad — only compared to the needs of the people who use it.

Which is designed better, the hibiscus flower or the blobfish? Neither! Both have evolved to be perfectly suited to their context. If the blobfish was “prettier” to humans, it would likely be killed, and thus poorly-designed for its environment.

These are the homepages of two large conglomerates: Berkshire Hathaway and Alphabet. At the visceral level, the Alphabet page is clearly better. Lots of white space, pleasant fonts, simple colors. But who are the primary users of these sites? What is the job they need done? For Berkshire, users are mostly looking for the link to Warren Buffett’s latest letter, or their annual report. Berkshire provides every potential action right up front, compatible with every browser and format. The aesthetics in this case are also part of the company’s frugal brand.

In any design discipline, the pattern language includes visual patterns. Some have survived a long time, and thus more objective: serif fonts, proximity, negative space, symmetry, etc.

A lot of visual design comes down to a subjective sense of “does this feel good”. Or in the words of Christopher Alexander, does it make you “feel alive”?

Final quiz: Which street makes you feel more “alive”? What are the patterns in each that contribute to positive or negative emotion? What about good or poor fit with needs?

Design is a meta skill

If you’ve made it to this point, it should be clear: all design depends on the context it needs to fit with. The job of a designer is to discover these constraints and how they translate into the requirements of the design. Using good patterns can help create experiences that both match our needs and make us feel good in the process.

If you can internalize the above principles, design is a meta skill that can be used anywhere to help you solve problems, create better experiences, and make people happier.

Although the fundamentals of design will always be the same, the tools that assist it continue to change. With AI, a new renaissance of generative design is currently underway. These tools will give us enormous capabilities — and make the fundamentals even more relevant.

I’ll close with this passage from Don Norman:

“We are all designers. We manipulate the environment to better serve our needs. We select what items to own, which to have around us. We build, buy, arrange, and restructure: all this is a form of design. Through our designs, we transform houses into homes, spaces into places, things into belongings. The best kind of design isn’t necessarily an object, a space, or a structure: it’s a process—dynamic and adaptable.”

For a much better and deeper look at design principles, these are the books I’d recommend:

• “The Design of Everyday Things”, Don Norman

• “Emotional Design”, Don Norman

• “Notes on the Synthesis of Form”, Christopher Alexander

• “The Timeless Way of Building”, Christopher Alexander

• “A Pattern Language”, Christopher Alexander

• “Competing Against Luck,” Clayton Christensen

Thank you to Celeste for reviewing drafts of this, and Ryan Singer for inspiring some of the concepts.

“The future is here — it’s just not evenly distributed.” — William Gibson

This essay is addressed to anyone still skeptical of what modern AI will be capable of. If you’ve already drunk the Kool-Aid and agree with the tweet below, you may not need to read on.

Over the past 4 months, I’ve been experimenting with programming tools and services that use LLMs at their core, for both personal and business uses.

Last weekend I went to a huge GPT-4/LLM hackathon in the Bay Area. The energy and talent there was amazing. Researchers and founders from companies like OpenAI, Anthropic, and HuggingFace were there, and they didn’t mince words on how fast progress was being made. Sure, I get it, these people aren’t just drinking the Kool-Aid — they’re the ones selling it. That doesn’t take away the fact that the advancements being made are truly groundbreaking.

At the hackathon, I saw numerous projects that showcased the potential of LLMs. There were tools that allowed you to generate entire apps from short descriptions, launch scalable server infrastructure from a text prompt, create virtual assistants, build financial models, and even communicate with robot dogs. In only a few hours, I wrote the code to create a daily, personalized podcast that summarizes what’s most important to you (and it won third place!)

The possibilities are seemingly endless, and we’re only just scratching the surface.

Objections?

Some say the models improvement will plateau. I say it doesn’t matter. The existing models — no improvement of the underlying structure — are enough for enormous, paradigm-shifting change. Even if all the researchers at OpenAI, Google, Anthropic, wherever all hang up their hats and retire, as long as today’s model versions are maintained it’s enough.

But but but… what about hallucinations!?

It doesn’t matter. All issues with current LLMs can be easily overcome to the 98th percentile of accuracy with minor supplementation. Microsoft gets this. Anything you think is an issue with LLMs isn’t. 1

We’ve gotten too used to computers as calculators or spreadsheets. The truth is, they don’t need to be 100% accurate. Nearly all of the work they’ll do isn’t life or death — it just needs to be better and cheaper than an average overworked human.

And it can get very specialized. Instead of one person doing a job with 3-4 types of tasks, you can have 5 LLM based functions: 4 of them fine-tuned to a very specific job, and a 5th that does nothing but verify accuracy.

ChatGPT is impressive, but it’s the TOY version. It’s completely generalized, fine tuned only to chat and not say things that will give OpenAI bad press.2

What about all the compute needed!?

This might be a bigger problem, but still… doesn’t matter in the long run.

Specialized models will be compressed and optimized for specific use cases, some on custom chipsets designed to do nothing but LLM inference.

Apple will put optimized LLMs on iPhones and Macs that can run 80% of the most common tasks, and it will be trained to know when to outsource the remaining 20% to a cloud API.

Some general observations

• “Out of the box” LLMs can be amazing, but are very limited in business applications. The moment you start hooking them up things like long-term memory and tools, they become scary useful.

• LLMs aren’t just for communicating, they’re for reasoning. If you present information in the right way, these models can absolutely reason and problem solve. Some things that help:

  • Giving enough context on the problem — either through directly providing all relevant details, or via context injection from searching some corpus of knowledge.

  • Zero- or one-shot solutions are unlikely. The best uses I’ve seen are chains of multiple prompts, supplemented with other functions and data. You can even use multiple different models (each can be better or worse at certain things) at each step along the way.

  • See this footnote for a case study on how small changes to how an LLM is queried can have large effects.3

• Searching the web like Bing Chat is just one trivial example of the tools an LLM can use. ChatGPT plugins that were released last week — although not available to the public yet — finally show how powerful this can be. Giving the model tools it can use is no different from giving a human tools. A few examples I’ve observed:

  • “Ask a human” — Give the model the ability to ask a human something it needs to find the answer. In our support bot for example, if there’s a piece of information missing, it can ask a human on the Slack channel. The answer is then fed back into the prompt chain.

  • WolframAlpha — With the world’s most advanced calculator, no more complaining about GPTs getting math wrong.

  • Use company APIs — Connect the models to any internal APIs so that it has up-to-date status on your system.

  • Zapier automations — Use anything you can connect to in Zapier. Like “Summarize my last email and send it as a text to Sarah.”

An “inevitable turn in human history”

From James Pethokoukis’ summary of the Goldman Sachs report on generative AI:

Here are the main takeaways from GS economists Joseph Briggs and Devesh Kodnani, which I shall subsequently dive into:

• Generative AI could raise annual US labor productivity growth by just under 1½ percentage points over a 10-year period following widespread business adoption.

• Generative AI could eventually increase annual global GDP by 7 percent, equal to an almost $7 trillion increase in annual global GDP over a 10-year period.

• Generative AI will be disruptive to jobs: “We find that roughly two-thirds of current jobs are exposed to some degree of AI automation, and that generative AI could substitute up to one-fourth of current work.”

• AI investment could approach 1 percent of US GDP by 2030 if it increases at the pace of software investment in the 1990s. (That said, US and global private investment in AI totaled $53 billion and $94 billion in 2021, a fivefold increase in real terms from five years prior.)

See, even the vampire squid is getting in on the hype.

Predicting 10 years out is hard, especially for something moving this fast. But in the spirit of the Bill Gates quote — “We always overestimate the change that will occur in the next two years and underestimate the change that will occur in the next ten” — I’d bet on Goldman’s predictions being too conservative. I think generative AI will boost productivity and GDP per capita by a lot more.

Now about the jobs: two-thirds of jobs being exposed to AI automation sounds about right. Basically any job that currently involves a computer or talking over the phone. My gardeners are safe. Substituting a quarter of all workers is another question.

I believe that in the long run, there will be more than enough jobs to go around. The nature of many of those jobs will be very different. And the transition to the new world may be painful for many.

The truth is, it’s possible to believe this will be a paradigm-shifting epoch, and still have no idea how it will play out — both good and bad.

From Tyler Cowen:

But since we are not used to living in moving history, and indeed most of us are psychologically unable to truly imagine living in moving history, all these new AI developments pose a great conundrum. We don’t know how to respond psychologically, or for that matter substantively. And just about all of the responses I am seeing I interpret as “copes,” whether from the optimists, the pessimists, or the extreme pessimists (e.g., Eliezer). No matter how positive or negative the overall calculus of cost and benefit, AI is very likely to overturn most of our apple carts, most of all for the so-called chattering classes.

The reality is that no one at the beginning of the printing press had any real idea of the changes it would bring. No one at the beginning of the fossil fuel era had much of an idea of the changes it would bring. No one is good at predicting the longer-term or even medium-term outcomes of these radical technological changes (we can do the short term, albeit imperfectly). No one. Not you, not Eliezer, not Sam Altman, and not your next door neighbor.

To the remaining AI skeptics, I say this: Don't be afraid of the AI revolution; embrace it. Understand that AI is a tool, not a replacement for human ingenuity. Yes, some jobs may change or disappear, but new ones will emerge — ones that require human creativity, empathy, and collaboration. By working together with AI, we can accomplish things that were once thought impossible.

Now that you’ve made it this far, what do you think?

In this roundup:

• 🤖 A.I.

  • Essay: Thoughts on the AI epoch — An idea maze for LLMs; Punctuated Equilibrium; The AI revolution; Where’s the moat?; The fate of Google.

  • My thoughts on AI (as a podcast!)

  • The BuffettBot Experiment

• 🚀 Space — 4 photos and a link to summarize 2022.

• 🔗 Interesting Links — Other takes on AI; Derek Thompson essays; and Choosing Good Quests.

 🤖 A.I.

Thoughts on the AI epoch

What more can be said about the AI boom that began its ascent less than a year ago? A lot! The potential of AI is immense and its influence on our lives is sure to be significant. And so I’ll continue. . .

In this mini-essay I’ll focus more on Large Language Models (LLMs), but my thoughts apply to all other AI efforts as well.

An easy way to think of it is that LLMs will soon become the “autocomplete for everything”:

What’s common to all of these visions is something we call the “sandwich” workflow. This is a three-step process. First, a human has a creative impulse, and gives the AI a prompt. The AI then generates a menu of options. The human then chooses an option, edits it, and adds any touches they like.

. . . So that’s our prediction for the near-term future of generative AI – not something that replaces humans, but something that gives them superpowers. A proverbial bicycle for the mind. Adjusting to those new superpowers will be a long, difficult trial-and-error process for both workers and companies, but as with the advent of machine tools and robots and word processors, we suspect that the final outcome will be better for most human workers than what currently exists.

There seems to be an unlimited number of areas that language prompting + completion will enable. Some are obvious: a new iteration of Google, help with writing, content generation, help with marketing copy, etc. You see many startups and tools that have already sprung up to tackle these.

Some of the real interesting applications that are incubating now will have action models as a big component. Models will have the ability to take actions like: searching the web; ordering an item; making a reservation; using a calculator; or using any other digital tool that humans are capable of using. Imagine ChatGPT being able to confirm its answers with multiple sources, or having access to all your personal records it can use to assist you.

Prompt engineers are already discovering how much you can do with the existing models, without any new advancements or manipulation of the actual base model. Even if GPT-4 or an open-source LLM from Stability.ai take years to come out, the existing tools are enough for huge changes.

An idea maze for LLMs

The above is an idea maze I sketched out for products enabled by LLMs. The key question it starts with is “What kind of interface would the use have with the product?”

• Search — you search for something just like you do on Google, but the results are much more relevant to what you’re truly looking for.

• Chat — you’re having a conversation with the AI, and as part of that back-and-forth communication it helps you accomplish some goal.

• Copilot — the AI “sits beside you” in whatever you’re working on, assisting where it can and giving whatever you’d normally do a big productivity boost. The actual interface in this case would be highly specific to the use case.

• Content generation — you give the AI some minor guidance upfront, and it generates the entire content to use (this might be considered a Copilot interface).

• [Games] — similar to chat, but you’re not necessarily having a conversation. The AI is leading you or enabling some sort of game-like interface for entertainment, learning, or problem solving.

A few highlighted examples:

• Personal assistant — Imagine Siri or Alexa if they actually had some serious capabilities. You say “My wife and I want to take a trip to Europe this summer,” and it uses travel booking APIs and TripAdvisor to create a handful of itineraries, their prices, and best aspects of each. You pick one, and it makes all the reservations for you. It follows up with reminders and suggestions on what to pack.

• Legal copilot — You’re a lawyer at a law firm. A LegalAI tool you use knows every law in existence, along with every prior case and their outcomes. You use it to pull out the most important points of a 200 page contract, and summarize it for your client. You use it to draft an agreement by listing the most important considerations, and it suggests points you may have forgot. Most impressively, you’re representing your client in a corporate fraud case and after ingesting all the relevant details, LegalAI suggests ways to win the case, including esoteric legal loopholes and creative arguments to sway the judge. The digital version of Extraordinary Attorney Woo.

• Personal tutor — You’re a teacher or parent of a child in 5th grade. They’re struggling and falling behind in math. You talk to an AI tutor service first to provide an overview of the situation, just like you would a real tutor. The child speaks with the AI, and it finds out more about what the know and how they learn best. It does daily sessions with them over the course of a few months, and they are now caught up with the rest of the students. Kids who would have never been able to afford a private tutor in the past now have the most knowledgeable tutors in the world. They still can’t fully replicate in-person tutors, but for most kids this doesn’t matter.

Punctuated equilibrium: Welcome to the AI revolution

The study of ecology can provide many good analogies to the world of business and technology. Both are complex adaptive systems, with structures and behaviors emerging over time as they evolve. Both have building blocks that combine into many variations as they adapt to fit their changing environment.

One of these models is punctuated equilibrium.

Punctuated equilibrium is the theory that species evolve at a slow, steady rate for long periods of time, with occasional, rapid changes taking place over short periods of time. The idea is that evolution progresses in a series of short bursts of rapid change, followed by long periods of stasis. These changes play out more like S-curves or step-functions then smooth linear or exponential growth.

It’s very clear that the current AI epoch is punctuation event in the history of technology and business — a rapid evolutionary change during which the world evolves quickly in response to changing environmental conditions.

It’s on the level of other computing revolutions like smartphones, the internet, and personal computers.

Microprocessors and other computer hardware led to cheap PCs in the hands of businesses and individuals. This allowed them to use software, the real general purpose productivity booster. Software allowed businesses to automate and streamline many of their operations, and individuals to do things they never thought possible before.

But software can be clunky. It still takes a lot of effort and maintenance to solve problems using software, and it is really bad at tasks that involve reasoning, learning from data or making decisions.

Enter AI models — Software 2.0:

Software 1.0 is code we write. Software 2.0 is code written by the optimization based on an evaluation criterion (such as “classify this training data correctly”). It is likely that any setting where the program is not obvious but one can repeatedly evaluate the performance of it (e.g. — did you classify some images correctly? do you win games of Go?) will be subject to this transition, because the optimization can find much better code than what a human can write.

AI can take in data, learn from it and make decisions — just like humans can. This means it can automate tasks that are difficult or impossible to do with software alone. AI is opening up new possibilities and opportunities that weren’t available before.

I’ve seen some comparisons, both negative and positive, to crypto/web3 hype in 2018 or 2020. The comparison isn’t apt to me at all. Here’s a few reasons why:

1. With the internet or crypto, you needed other people to be involved to make it more useful. The current crop of AI is useful out of the box and doesn’t need network effects, similar to PCs. (Although in both cases it still helps to have more people using in the long run.)

2. Web3 is still a solution looking for a problem. It will have legitimate uses (you can make a good argument that AI will actually drive those uses in some cases), but there are much clearer products or features that can have immediate value to users for AI. See my experiment with BuffettBot in the section below — I built this in a week and within a day thousands of people got immediate use from it!

NZS Capital’s Brad Slingerlend phrased the potential effects well:

Could we see a doubling of productivity across nearly every information-based job? It’s such early days, yet the results are so promising, that I am willing to venture into the extremely dangerous territory of making predictions – and declare that we just might see massive productivity increases from chatbots and generative AI unlike anything we have yet seen over the course of the Information Age – outweighing even PCs, smartphones, and the Internet. I hate the expression “buckle up”, but it might be called for here.

Where’s the moat?

Continuing the comparison to past computing revolutions: I believe it’s likely that, relative to the internet or PCs, this epoch of AI may be less disruptive to incumbents and more productivity-boosting to humanity.

It doesn’t have built-in network effects or economies of scale like the PC app ecosystem or the internet. Nevertheless, I think competitive advantage can be strong depending on the area:

• Technical — In some niche areas, model structure and implementation may give an advantage for long enough to put the product in some other defensible flywheel. They don’t even need to be better quality either. Making models faster, cheaper to run, or available offline can give distinct advantages in some areas.

• Learning curves & data network effects — If model uses Reinforcement Learning from Human Feedback (RLHF), the more people that use it, the better the model becomes. The more niche the use case, the more important this becomes.

• Switching costs — Many will have strong switching costs due to: (1) tight integration with workflow, (2) personal attachment (like a personal assistant), or (3) regular/habitual usage associated with interface or brand.

The last point about regular usage associated with an interface may be key:

Any company that can control the user's main interface will win.

Scott Belsky, the CPO of Adobe, has a great theory about this:

Companies that successfully aggregate multiple services in a single interface have a chance of really shaking up industries. As soon as you rely on one interface for a suite of needs, you become loyal to the interface rather than the individual services it provides.

Where will users interact with the AI model? If it’s easier for people to use it through an existing, known interface, it’s much easier for incumbents to adapt and add AI as a feature. Startups had an advantage in the internet era because the interface was completely new in most cases. It was disruptive (in a Christensen sense) to their business models.

It’s clear that AI writing assistants will be big. But who will win here? If Microsoft Word, Notion, and other existing apps can integrate AI properly within some reasonable time, most users would rather stay then move to a new, unfamiliar interface. AI seems more like a sustaining innovation in this case.

Which brings us to the big one: Google.

If you ask the same question about Google, the next question becomes “how long will it take for Google to integrate LLMs into their main search UI?” And if in the next few years they can do this just enough that people don’t go elsewhere for general-purpose queries, then users will stay.

The question after that is what it will do to Google’s business model. Is it still sustaining? Possibly. Even if so, it could make it not as financially lucrative.

Liberty shared some good thoughts on how Google will be impacted here [paid]. A good scenario may be that their gross margins go down a bit as search costs increase.

Just before publishing this, Ben Thompson of Stratechery wrote a great essay that essentially aligned with my thoughts above. The most relevant paragraph:

What is notable about this history is that the supposition I stated above isn’t quite right; disruptive innovations do consistently come from new entrants in a market, but those new entrants aren’t necessarily startups: some of the biggest winners in previous tech epochs have been existing companies leveraging their current business to move into a new space. At the same time, the other tenets of Christensen’s theory hold: Microsoft struggled with mobile because it was disruptive, but SaaS was ultimately sustaining because its business model was already aligned.

My thoughts on AI (as a podcast!)

A few months ago I had a discussion on my friend Eric Jorgenson’s podcast. We intended to talk about a variety of topics, but ended up discussing mostly AI.

With how fast progress has been moving recently, two months ago is like years, so some of the content may already be out of date. Some topics we discuss:

What is computer vision? How does a neural network work? How we’re making the world legible to computers. What will AI be capable of? How will Google fare in this new paradigm? Can computers smell?

#051 AI Evolution: ComputerVision, Olfactory Computation, and Neural Nets with Max Olson

The BuffettBot experiment

On December 14, I launched BuffettBot.com.

BuffettBot uses the OpenAI API to search an archive of his writings and ask “Warren Buffett” questions. It even connects the answer to quotes and sources it pulled the information from.

Unfortunately the service only ran for 3 days before I had to FLIP THE KILL SWITCH.

No, BuffettBot didn’t become sentient and threaten to allocate the world’s resources. The usage just exploded and I couldn’t justify keeping it running given the daily cost of using the API.

Ultimately even though I had to shut it down for now, this was a successful experiment.

What’s the future of BuffettBot?

Given the cost to run it in the current form, I’m not sure what the future holds. I don’t think the best model for something like this is pay-per-use. (Although a platform with many of these bots may justify something like “charge $10 to ask 100 questions”.)

A version that’s much more cost efficient and still very valuable is something like the following:

“Buffett Archive” where you can semantically search all of Warren Buffett’s writings and interviews. Semantic search means you’re not doing a ⌘F and looking for an exact phrase, you’re looking for meaning in the text. So you could search “how did GEICO’s float change in 1995” or “the Washington Post’s moat” and get the most relevant results. The app will allow you to browse all the results, showing them in context of the letter/transcript/etc. that they were found.

Let me know if you’d be interested in something like this! Given the popularity of BuffettBot, I think this would be pretty valuable. And a “Munger Archive” as well.

🚀 Space

Although AI progress in 2022 seemed like the main show, we made some pretty big leaps forward in space.

The past year can be best summarized in 4 photos:

For a more detailed summary of what went down in space last year, Orbital Index gave a great rundown here.

🔗 Interesting Links

• Other good AI takes I’m fond of:

  • Before the flood, by Samuel Hammond

  • AI from Superintelligence to ChatGPT, by Séb Krier (Works in Progress)

  • The Age of Industrialized AI, by Daniel Jeffries

  • Superhistory, Not Superintelligence, by Venkatesh Rao

• Derek Thompson has written a few good essays recently:

  • Why America Doesn’t Build What It Invents — The U.S. just made a breakthrough in nuclear-fusion technology. Will we know how to use it?

  • Why the Age of American Progress Ended — Invention alone can’t change the world; what matters is what happens next.

  • Your Creativity Won’t Save Your Job from AI — Robots were once considered capable only of unimaginative, routine work. Today they write articles and create award-winning art.

• Unblocking Abundance, by Sarah Constantin — What can we do to remove to barriers to progress?

• Choose Good Quests, by Trae Stephens and Markie Wagner — Silicon Valley's current focus on easy money has resulted in a failure to solve big problems. There is a moral imperative for our best players to choose good, hard quests, which will make the future better than the world today.

In this essay I’ll focus more on Large Language Models (LLMs), but my thoughts apply to all other AI efforts as well.

An easy way to think of it is that LLMs will soon become the “autocomplete for everything”:

What’s common to all of these visions is something we call the “sandwich” workflow. This is a three-step process. First, a human has a creative impulse, and gives the AI a prompt. The AI then generates a menu of options. The human then chooses an option, edits it, and adds any touches they like.

. . . So that’s our prediction for the near-term future of generative AI – not something that replaces humans, but something that gives them superpowers. A proverbial bicycle for the mind. Adjusting to those new superpowers will be a long, difficult trial-and-error process for both workers and companies, but as with the advent of machine tools and robots and word processors, we suspect that the final outcome will be better for most human workers than what currently exists.

There seems to be an unlimited number of areas that language prompting + completion will enable. Some are obvious: a new iteration of Google, help with writing, content generation, help with marketing copy, etc. You see many startups and tools that have already sprung up to tackle these.

Some of the real interesting applications that are incubating now will have action models as a big component. Models will have the ability to take actions like: searching the web; ordering an item; making a reservation; using a calculator; or using any other digital tool that humans are capable of using. Imagine ChatGPT being able to confirm its answers with multiple sources, or having access to all your personal records it can use to assist you.

Prompt engineers are already discovering how much you can do with the existing models, without any new advancements or manipulation of the actual base model. Even if GPT-4 or an open-source LLM from Stability.ai take years to come out, the existing tools are enough for huge changes.

An idea maze for LLMs

The above is an idea maze I sketched out for products enabled by LLMs. The key question it starts with is “What kind of interface would the use have with the product?”

• Search — you search for something just like you do on Google, but the results are much more relevant to what you’re truly looking for.

• Chat — you’re having a conversation with the AI, and as part of that back-and-forth communication it helps you accomplish some goal.

• Copilot — the AI “sits beside you” in whatever you’re working on, assisting where it can and giving whatever you’d normally do a big productivity boost. The actual interface in this case would be highly specific to the use case.

• Content generation — you give the AI some minor guidance upfront, and it generates the entire content to use (this might be considered a Copilot interface).

• [Games] — similar to chat, but you’re not necessarily having a conversation. The AI is leading you or enabling some sort of game-like interface for entertainment, learning, or problem solving.

A few highlighted examples:

• Personal assistant — Imagine Siri or Alexa if they actually had some serious capabilities. You say “My wife and I want to take a trip to Europe this summer,” and it uses travel booking APIs and TripAdvisor to create a handful of itineraries, their prices, and best aspects of each. You pick one, and it makes all the reservations for you. It follows up with reminders and suggestions on what to pack.

• Legal copilot — You’re a lawyer at a law firm. A LegalAI tool you use knows every law in existence, along with every prior case and their outcomes. You use it to pull out the most important points of a 200 page contract, and summarize it for your client. You use it to draft an agreement by listing the most important considerations, and it suggests points you may have forgot. Most impressively, you’re representing your client in a corporate fraud case and after ingesting all the relevant details, LegalAI suggests ways to win the case, including esoteric legal loopholes and creative arguments to sway the judge. The digital version of Extraordinary Attorney Woo.

• Personal tutor — You’re a teacher or parent of a child in 5th grade. They’re struggling and falling behind in math. You talk to an AI tutor service first to provide an overview of the situation, just like you would a real tutor. The child speaks with the AI, and it finds out more about what the know and how they learn best. It does daily sessions with them over the course of a few months, and they are now caught up with the rest of the students. Kids who would have never been able to afford a private tutor in the past now have the most knowledgeable tutors in the world. They still can’t fully replicate in-person tutors, but for most kids this doesn’t matter.

Punctuated equilibrium: Welcome to the AI revolution

The study of ecology can provide many good analogies to the world of business and technology. Both are complex adaptive systems, with structures and behaviors emerging over time as they evolve. Both have building blocks that combine into many variations as they adapt to fit their changing environment.

One of these models is punctuated equilibrium.

Punctuated equilibrium is the theory that species evolve at a slow, steady rate for long periods of time, with occasional, rapid changes taking place over short periods of time. The idea is that evolution progresses in a series of short bursts of rapid change, followed by long periods of stasis. These changes play out more like S-curves or step-functions then smooth linear or exponential growth.

It’s very clear that the current AI epoch is punctuation event in the history of technology and business — a rapid evolutionary change during which the world evolves quickly in response to changing environmental conditions.

It’s on the level of other computing revolutions like smartphones, the internet, and personal computers.

Microprocessors and other computer hardware led to cheap PCs in the hands of businesses and individuals. This allowed them to use software, the real general purpose productivity booster. Software allowed businesses to automate and streamline many of their operations, and individuals to do things they never thought possible before.

But software can be clunky. It still takes a lot of effort and maintenance to solve problems using software, and it is really bad at tasks that involve reasoning, learning from data or making decisions.

Enter AI models — Software 2.0:

Software 1.0 is code we write. Software 2.0 is code written by the optimization based on an evaluation criterion (such as “classify this training data correctly”). It is likely that any setting where the program is not obvious but one can repeatedly evaluate the performance of it (e.g. — did you classify some images correctly? do you win games of Go?) will be subject to this transition, because the optimization can find much better code than what a human can write.

AI can take in data, learn from it and make decisions — just like humans can. This means it can automate tasks that are difficult or impossible to do with software alone. AI is opening up new possibilities and opportunities that weren’t available before.

I’ve seen some comparisons, both negative and positive, to crypto/web3 hype in 2018 or 2020. The comparison isn’t apt to me at all. Here’s a few reasons why:

1. With the internet or crypto, you needed other people to be involved to make it more useful. The current crop of AI is useful out of the box and doesn’t need network effects, similar to PCs. (Although in both cases it still helps to have more people using in the long run.)

2. Web3 is still a solution looking for a problem. It will have legitimate uses (you can make a good argument that AI will actually drive those uses in some cases), but there are much clearer products or features that can have immediate value to users for AI. See my experiment with BuffettBot — I built this in a week and within a day thousands of people got immediate use from it!

NZS Capital’s Brad Slingerlend phrased the potential effects well:

Could we see a doubling of productivity across nearly every information-based job? It’s such early days, yet the results are so promising, that I am willing to venture into the extremely dangerous territory of making predictions – and declare that we just might see massive productivity increases from chatbots and generative AI unlike anything we have yet seen over the course of the Information Age – outweighing even PCs, smartphones, and the Internet. I hate the expression “buckle up”, but it might be called for here.

Where’s the moat?

Continuing the comparison to past computing revolutions: I believe it’s likely that, relative to the internet or PCs, this epoch of AI may be less disruptive to incumbents and more productivity-boosting to humanity.

It doesn’t have built-in network effects or economies of scale like the PC app ecosystem or the internet. Nevertheless, I think competitive advantage can be strong depending on the area:

• Technical — In some niche areas, model structure and implementation may give an advantage for long enough to put the product in some other defensible flywheel. They don’t even need to be better quality either. Making models faster, cheaper to run, or available offline can give distinct advantages in some areas.

• Learning curves & data network effects — If model uses Reinforcement Learning from Human Feedback (RLHF), the more people that use it, the better the model becomes. The more niche the use case, the more important this becomes.

• Switching costs — Many will have strong switching costs due to: (1) tight integration with workflow, (2) personal attachment (like a personal assistant), or (3) regular/habitual usage associated with interface or brand.

The last point about regular usage associated with an interface may be key:

Any company that can control the user's main interface will win.

Scott Belsky, the CPO of Adobe, has a great theory about this:

Companies that successfully aggregate multiple services in a single interface have a chance of really shaking up industries. As soon as you rely on one interface for a suite of needs, you become loyal to the interface rather than the individual services it provides.

Where will users interact with the AI model? If it’s easier for people to use it through an existing, known interface, it’s much easier for incumbents to adapt and add AI as a feature. Startups had an advantage in the internet era because the interface was completely new in most cases. It was disruptive (in a Christensen sense) to their business models.

It’s clear that AI writing assistants will be big. But who will win here? If Microsoft Word, Notion, and other existing apps can integrate AI properly within some reasonable time, most users would rather stay then move to a new, unfamiliar interface. AI seems more like a sustaining innovation in this case.

Which brings us to the big one: Google.

If you ask the same question about Google, the next question becomes “how long will it take for Google to integrate LLMs into their main search UI?” And if in the next few years they can do this just enough that people don’t go elsewhere for general-purpose queries, then users will stay.

The question after that is what it will do to Google’s business model. Is it still sustaining? Possibly. Even if so, it could make it not as financially lucrative.

Liberty shared some good thoughts on how Google will be impacted here [paid]. A good scenario for them may be that their gross margins only go down a bit as search costs increase.

Just before publishing this, Ben Thompson of Stratechery wrote a great essay that essentially aligned with my thoughts above. The most relevant paragraph:

What is notable about this history is that the supposition I stated above isn’t quite right; disruptive innovations do consistently come from new entrants in a market, but those new entrants aren’t necessarily startups: some of the biggest winners in previous tech epochs have been existing companies leveraging their current business to move into a new space. At the same time, the other tenets of Christensen’s theory hold: Microsoft struggled with mobile because it was disruptive, but SaaS was ultimately sustaining because its business model was already aligned.

The following is a short case study on “Creator”, a cloud-based content management system I built at Mashgin, where we make visual self-checkout kiosks that use computer vision to see items so you don’t have to scan barcodes.

In the years since launch, it has given location managers the ability to customize their menus in ways they were unable to in the past. This empowers them to make frequent changes, tailoring the menu to customer needs rather than just “using the default”.

Mashgin Creator is a tool for operators to build and manage their menus, from items to discounts, schedules, and more.

Mashgin customers have been able to easily edit their checkout items in the cloud since we first launched in 2016. But when we began to design our mobile and in-person ordering app, we realized customers would need an easy way to design more complex menus, with custom item options, photos, nested categories, scheduling, and more. This is where the idea for Creator came in.

Creator is what they call in the industry a "CMS", or content management system. Any software tool used to manage content of any type could apply.

In the food service industry, a CMS is used to manage their menu items, pricing, discounts, taxes, etc. The scope could be anywhere from an individual cafe to a nationwide chain of stores.

Most existing CMS software for food service was cumbersome to use and poorly designed. It was really just a simple layer on top of a database, allowing users to edit basic item information. Some software didn't even allow for real-time syncing of data — any changes are "submitted" and someone behind the scenes has to deploy them to the menu.

The output of these menus is very simple: it's just items in some nested menus, each with its own data like price, type, options, etc. But the work and consideration that has to go into building each menu is anything but simple.

It was clear that our customers needed something much better.

Designing the app

Believing that all the existing tools weren't very good, we chose not to base the core design off of any other examples or prior work. Creator would be rethought from the ground up based on the needs and jobs of its users.

Fortunately when it came to understanding the use cases we had a few big advantages:

1. We were working closely with multiple location managers who would be end-users of Creator and would give immediate, helpful feedback;

2. All members of our support team were also users who had a very good idea of what was needed;

3. Even more conveniently, I had the experience myself of having to design order menus and understood the major needs and pain points. It is an often unappreciated advantage for the lead engineer and designer of a product to be able to build the tool that they would want to use.

I wanted the design to look and feel like it fits in with any modern office web app, like Airtable, Figma, Notion, etc. The UI needed to be intuitive and ready to use without any training. This meant we needed to use existing design patterns and models of interaction. Common tasks should take no more than a few clicks to accomplish.

This meant using design patterns like WYSIWYG displays to mimic the end-user experience, organization via easy drag & drop, in-line editing, and multiple ways of viewing the menu depending on what your goal is.

Just like other modern apps, users should feel like they can "play" around with the functionality to learn it, without getting frustrated or fearing they'll break something.

Because we were free from the burden of following rigid customer specs, we could add features that have thus far never been done in the industry. Like the ability to easily pull in a high-quality, royalty-free photo for an item if you don't have one.

Feedback from customers

As soon as we released Creator we began hearing positive feedback from customers.

"This will be a game changer for [our] managers." — Regional Food Service Director

Based on further discussions with power users, we realized there was a need for more intuitive menu scheduling. So we quickly added the ability select which item menu appeared on which day.

Creator's ease-of-use also directly led to us winning an account, taking not only the checkout but full mobile ordering infrastructure from a competitor. The competing CMS was complicated and would sometimes take many days to make changes.

For large enterprise customers, one good product or service isn't enough. You have to get the entire ecosystem of services right, along with an amazing support team to back them up.

Creator — and other tools like it — provide a full suite for clients to use. This makes us much more sticky and also makes it easier for internal customer buy-in.

My name is Max Olson, and I’ve been writing on FutureBlind since 2007. I write about whatever interests me, but these are the main areas I’ll cover:

• Technology — in particular frontier tech like AI, Space, Bioengineering, etc.

• Progress — how we can make faster progress, in particular how we can tell better stories that drive progress

• Business & Investing — mental models about startups, investing, and business analysis

• Design — how to design better experiences and how new tools drive better design

Below are some featured posts to get started with. If you like any of them, please subscribe to the newsletter — it goes out every 2-3 months. I hope you learn something new and come away more optimistic.

Read more about the blog and my beliefs...

Stay up-to-date with the newsletter, every 2-3 months:

Subscribe now

Subscribers will get a roundup of posts and other things I find interesting about every quarter. I’ll also do the occasional post as a podcast episode.

Featured posts to get started with:

• 🔄 Advantage Flywheels — Competitive advantage can be represented visually as 1 or more feedback loops.

• 🚀 Take the Iterative Path — How SpaceX moves fast and blows things up to become an industry leader.

• ⚡️ Let’s jumpstart the new industrial revolution — What could a new industrial revolution bring? 

• 🤖 Mashgin: The Future of Computer Vision — Our vision at Mashgin as of 2016.

• 🌌 The Future of Space, Part I: The Setup — Why the space industry is about to take off.

• 🌺 Generalists vs. Specialists (and the Specialist’s Dilemma) -- How we can learn from generalist and specialist species.

• 🪜 Managing Modes of Effort — A framework for managing progress at every level of abstraction.

• 🚀 Product Strategy: Falcon 9 — A short product case study on SpaceX’s Falcon 9.

• 🍟 The McDonald’s Success Story — The early history of McDonalds and what made them successful.

• 🍔 The Restaurant Investor — A long-form essay on the turnaround of Stake-n-Shake and what it takes to succeed in the chain restaurant business.

Interviewing is actually not very helpful. Or at the very least it is extremely difficult to judge how someone will perform with just interviews, especially when unstructured. (See here, or all the data Google collected.)

Why?

• People interviewing usually have relatively little experience, and thus have a poor "base rate" to judge the candidate against.

• Answers to questions generally have very little correlation with actual performance.

• It's difficult to extract enough information (even in long interview processes) to make a proper call. Imagine going on a 3-hour date, thinking it over for a few days, then asking the person to get married.

So what are other ways to know if someone will be good?

• You're friends with them or have worked with them already.

• Pointed reference checks from trusted people.

• "Trial" side project or task requiring interaction with team. Getting as close as possible to a real working environment.

• Recruiter who is both (1) very familiar with your needs/culture, and (2) specialized in hiring for that role.

But short of these things interviews are still necessary. Regardless of the specific process, it is important to have a set plan and follow it for every candidate.

Some advice:

• Prepare: don't go into an interview cold. Know what you want to get out of them and have a clear plan for how to evaluate them.

• Let them do the talking. You should only guide them and push them. Ask follow ups: Why? What did you do about it? How come?

• Brain teasers don't work, and aren't indicative of anything.

• The most effective questions are situational rather than just having them recall the past. "Instead of asking candidates to describe how they handled a unique situation in a previous job or organization, it’s more fruitful to describe consistent situations that candidates could face in this job or organization, and ask them what they would do — or how they would reason."

• Encourage them to ask questions -- about your questions, you, or the company.

• Be transparent and open about your entire hiring process.

• Get away from your desk or room: Take them out, take a tour of offices, etc.

The main things you're trying to get are:

• Excitement test. Would hiring this person make you more excited to work at the company?

• What can they do now, and how quickly could they be productive?

• How is this person going to be performing in 1 year from now?

  • How long does it take for them to learn something new?

  • What's their growth mindset and can they continually get better?

• Will they work well with the team?

• How long are they willing to keep pushing on a good project until giving up?

• How hard is it for them to change their mind or adjust course?

• Do they do the right thing even when they don't have to?

Aside from specific skills, what traits are the best indicators of these?

• Integrity: not just honesty, but integrity with themselves, their ideas, and "doing the right thing" when necessary. They seek out truth and embrace failure.

• Social intelligence: works well with others and is empathetic/caring.

• Intelligence

  • Raw intelligence

  • Creativity in problem solving

  • Adaptability

  • [One of the best ways I found to test for this is to ask about something they really enjoyed working on. Then grill them with questions about it, diving as deep as possible into the details.]

• Drive: is self motivated and can push themselves to get things done, even if it's not enjoyable work (grit). More internally motivated than externally.

• [Curiosity: This could be part of intelligence or drive, but it needs to be tested for somehow. I liked to ask questions like "What things are interested in outside of work?" or even better "Pick a topic that's not part of your day job (hobby, book, subject) and take a few minutes to explain it."]

Warren Buffett: "In looking for someone to hire, you look for three qualities: integrity, intelligence and energy. Without the first the other two will kill you."

There's also a problem with hiring the "best" — they are either extremely expensive or have unlimited options so will want to work elsewhere. This is like the Moneyball problem in baseball: the best teams will have the best reputation and most resources to get the best players. But these players aren't necessarily the only best — they are just the ones who look really good based on the most obvious metrics.

So what do you do? Look for talent in places with low competition, that require more work, or who are too "different":

• Growth potential -- people who are young or with little experience in the area, but are smart, driven, and internally motivated. You want people at the start of their "performance curve" -- in the 80th percentile that can move up to the 97th over time. [I think I'd change these numbers now. Finding someone in 80th percentile is too low. If inexperienced, you still want them in 90th with ability to move to 99th.]

• Interest -- people with unusually strong interest in your product or mission.

• Small fish in a big pond -- picked-over, under-utilized talent in large companies who can thrive on a smaller team. [You have to be careful here. Many people, although talented, can work under the bureaucracy of a big company for years and it drains them of the ability to get things done fast.]

• Different -- too outside the traditional track to be easily seen or picked up by others.

What about at the team level? What's the right mix of people?

• Diversity of thought and backgrounds is very important. You want people with good traits (good character, drive, etc.) and driven toward the same goal(s) but with a wide variety of experiences/backgrounds, and hence ways to think about problems. You don't want to hire a bunch of clones — that may work short term for some problems but will break when things change. See here for facts about workplace diversity in general. [Addendum: this is less important at the very beginning (seed) stage of a startup. With only a handful of people you may want similar types to get along better.]

Good materials:

• https://medium.com/evergreen-business-weekly/why-interviewing-prospective-hires-is-futile-and-how-to-make-it-only-slightly-less-futile-5d1d493ef5bc

• http://firstround.com/review/The-anatomy-of-the-perfect-technical-interview-from-a-former-Amazon-VP/

• https://www.nytimes.com/guides/business/how-to-hire-the-right-person

Stripe is a good case study of hiring processes:

• Jobs page

• On-site interviews for Engineering: What to expect (PDF) [Link is now broken -- if anyone still has access to this let me know.]

• What is the engineering interview process like at Stripe?

• How Stripe built one of Silicon Valley's best engineering teams [Greg Brockman (now president at OpenAI) talks about how Stripe built out their engineering department.]

This month marks the 15th anniversary of my first post on the FutureBlind blog. This is such a long time in the internet age that I feel like an old man now. I started the blog in college as a place for my thoughts on investing and business case studies. What I’ve written about over the years has morphed along with my interests, and I continue to enjoy putting my thoughts out there. I’ll keep going as long as I’m able to and hope readers continue to find it enjoyable! 😄

In this roundup edition:

• Essay: Take the Iterative Path — How SpaceX innovates by moving fast and blowing things up.

• 🖼 The AI art renaissance — What kinds of crazy applications will the AI art models lead to?

• ⚡️ Energy! — Energy superabundance and Mark Nelson on nuclear.

• 🚀 Space updates — Will we see SLS and Starship launch the same month?

• 🧪 What negatives does technology cause? — How do we distinguish potential risks of new tech?

• 🔦 Company Spotlights — Rocket Lab and Perimeter Solutions.

• 🔗 Interesting Links — All about Polaroid, why American can’t build, and the little ways the world works.

• 📚 Book notes — How Innovation Works, Where Good Ideas Come From

Take the Iterative Path

This essay is also available as an episode on the FutureBlind podcast. Check it out below or subscribe to the feed in: Apple, Overcast, Spotify, Pocket Casts.

FutureBlind

Take the Iterative Path

Listen now (20 min) | When a goal is big and complicated, an iterative, fail-fast approach is much better than a linear approach. Determinate paths lead to slower, poorly-adapted solutions, whereas iteration finds problems faster with a result that’s better adapted to the real world…

Read more

4 years ago · Max Olson

One of the greatest business successes over the last 20 years has been SpaceX’s rise to dominance. SpaceX now launches more rockets to orbit than any other company (or nation) in the world. They seem to move fast on every level, out executing and out innovating everyone in the industry.

Their story has been rightfully told as one of engineering brilliance and determination.

But at its core, the key their success is much simpler.

There’s a clue in this NASA report on the Commercial Crew Program:

SpaceX and Boeing have very different philosophies in terms of how they develop hardware. SpaceX focuses on rapidly iterating through a build-test-learn approach that drives modifications toward design maturity. Boeing utilizes a well-established systems engineering methodology targeted at an initial investment in engineering studies and analysis to mature the system design prior to building and testing the hardware. Each approach has advantages and disadvantages.

This is the heart of why SpaceX won. They take an iterative path.

Continue reading “Take the Iterative Path” on FutureBlind. . .

Other thoughts

🖼 The AI art renaissance

Unless you haven’t been on the internet in the past 6 months, you’ve already heard about AI art generators like DALL-E 2, Midjourney, Stable Diffusion, and Google’s Imagen.

I can’t begin to imagine all the crazy applications of this there will be in the coming years (and months!). As usual, there will be bad use cases (see the section “What negatives does technology cause?” below on this). But AI art is a particularly fun one to think of all the amazing positive use cases.

Custom, more specialized versions of models will be common and we’re already seeing them adapted from the SD model. One of the biggest benefits of neural nets in general is that you can build off of existing models with minimal retraining: you are essentially using its prior understanding of how to translate text prompts into images, and layering on further knowledge. The new model needs to get trained again but only really the last few “layers” — the computationally expensive part of the model is already good to go.

A few examples:

• Create a model trained specifically on comic books. Plug in a storyline, output the comic, and the storyteller chooses amongst different styles, etc. Here’s an example of someone that already used Midjourney to design a comic strip.

• DreamBooth, which originally used Imagen but has now be ported to SD. This solves the problem of models not being able to integrate untrained data. You can put in photos of yourself and put them in the generations. This is an obvious use case that will take off quickly.

  I can imagine in a few years having it productized by Apple or Google. Right now you can ask Google Photos to find photos of you on the beach, and it finds them. With this feature you’ll be able to search for “me on the top of Everest” and get back a photo of you triumphantly waving a flag on the summit of Mt. Everest that no human on Instagram will be able to distinguish from reality.

  📹 Corridor Crew: “Is This the Death of VFX?” where they use DreamBooth to create a story using all the crew.

Existing large models seem to already be better at certain things. In my experiments I’ve found DALL-E to be best for the images I’m looking for (see all the illustrations in the essay above). However Midjourney’s more fantasy-like digital art is pretty astounding. Make sure to checkout their showcase below, and refresh it every week as they update it.

🖼 Midjourney Community Showcase

Ultimately this will just become another tool in the creative toolbox. A really really powerful one. And yes, it will automate a lot of artistic work, especially of the low-level one-off job variety. It will also create a lot of opportunity for people to be more creative.

📹 Cleo Abram’s “The REAL fight over AI art” addressing some of the moral and ethical concerns.

⚡️ Energy!

Must read: “Energy Superabundance: How Cheap, Abundant Energy Will Shape Our Future”, a paper by Austin Vernon and Eli Dourado. What are some of the effects of very low energy costs? Where will increased demand come from? They go over the details of how abundant energy could affect transportation, water, agriculture, and more.

Pair with this report on the Clean Energy Transition, and Packy McCormick’s “Working Harder and Smarter” on why we want to use more energy (and not just make things more energy efficient).

Speaking of energy abundance — We need more elemental power! Mark Nelson, managing direct of Radiant Energy Fund, has been making the rounds recently on the current state of nuclear.

🎙 My friend Eric Jorgenson did an interview where he gave a great rundown on nuclear’s history and where it stands today.

🎙 Liberty also did 2 long interviews with Mark that went in-depth on the topic. Here’s part 2.

It’s not all positive — here’s a negative nuclear article in the New Yorker. Although the author admits he’s been a critic of nuclear since the 1970s, so I don’t know how partial he could suddenly become after so much confirmation bias over the years.

🚀 Space updates

Will we see a November launch of both NASA’s SLS and SpaceX’s full-stack Starship to orbit?

NASA’s SLS launch has been a long time coming. The building/testing has been delayed for years, and now the launch itself has been scrubbed a few times (from a combination of bad luck and questionable decisions). With hurricane Ian on the horizon, the current estimated launch time is November.

The best rundown of how we got to now with the SLS (as usual) comes from Eric Berger:

With the launch just days away, I am incredibly happy for the people at NASA and the space companies that have worked hard, cut through the bureaucracy, managed thousands of requirements, and actually got this rocket built. And I’m eager to see it fly. Who doesn’t want to watch a huge, Brobdingnagian rocket consume millions of kilograms of fuel and break the surly bonds of Earth’s gravity?

On the less happy side, it remains difficult to celebrate a rocket that, in many ways, is responsible for a lost decade of US space exploration. The financial costs of the program have been enormous. Between the rocket, its ground systems, and the Orion spacecraft launching on top of the stack, NASA has spent tens of billions of dollars. But I would argue that the opportunity costs are higher. For a decade, Congress pushed NASA’s exploration focus toward an Apollo-like program, with a massive launch vehicle that is utterly expended, using 1970s technology in its engines, tanks, and boosters.

Effectively, NASA was told to look backward when this country's vibrant commercial space industry was ready to push toward sustainable spaceflight by building big rockets and landing them—or storing propellant in space or building reusable tugs to go back and forth between the Earth and Moon. It's as if Congress told NASA to keep printing newspapers in a world with broadband Internet.

Elon is also saying that November is likely for an orbital Starship launch. Yes, this is Elon-time, but one can hope, right?

🧪 What negatives does technology cause?

How can you push forward at a fast pace and still get ahead of unintended consequences?

Jason Crawford writes more about this in “Towards a philosophy of safety”.

I followed up in this forum post asking how we can classify the negative effects of technologies in order to better prevent them.

Kevin Kelly wrote a good post putting the effects into 2 classes: "Class 1 problems are due to it not working perfectly. Class 2 problems are due to it working perfectly."

It’s clear that separating different types of potential risk is important. For example: whether or not a problem has fat tail, extremistan-world consequences, and how difficult they are to control. Novel pathogens come to mind here.

🔦 Company Spotlights

Rocket Lab ($RKLB) is the only other commercial launch provider aside from SpaceX that is regularly delivering customer payloads to orbit. They currently focus on small payloads (<300 kg) with their Electron rocket, and are planning on moving up-market with the development of their reusable Neutron rocket (13,000 kg to LEO). They’re also moving across the space stack by building satellites and many other components necessary for the space economy. One of their upcoming launches is even interplanetary: a probe to Venus, the first ever privately funded venture to another planet. (Disclosure: I own shares in Rocket Lab through the fund I manage.)

Perimeter Solutions ($PRM) has a near monopoly on the fire retardant used for wildfires. It is co-chaired by Nicholas Howley (founder of Transdigm) and William Thorndike (author of “The Outsiders”). The company has an interesting history, first starting as a division within Monsanto, spun off in 1997. Through a series of private equity acquisitions, it was renamed “Perimeter Solutions” in 2018 and continued to grow via acquisitions until going public via SPAC in 2021. Post going public, they seem to be taking the “outsider” approach and acting more as a holding co with the goal of maximizing shareholder returns.

🔗 Interesting Links

• Stories of Our Future from World’s Fair Co. — A few months ago my friend Cam Wiese showed me a preview of artwork he commissioned from Don Clark, the same artist who did the NASA JPL posters. Now they’re finally available as posters! The vibe is amazing and I highly recommend checking them out. I’ll be buying a few myself.

• All about Polaroid 🌈: In 2018 I wrote a short post on the story of Polaroid and how Apple was their spiritual successor. Recently David Senra broke the company down on Business Breakdowns. On David’s Founder’s Podcast — all episodes of which are now freely available after being acquired by Colossus — he’s done many episodes on Ed Land and Polaroid as well. I’ve listened to them all and am glad they’re finally available to share:

  • 🎙 Polaroid: The Genius of Edwin Land (Business Breakdowns)

  • 🎙 Edwin Land: The Story of Polaroid

  • 🎙 Edwin Land: Polaroid and the Man Who Invented It

• Why America Can’t Build, by Brian Balkus. Why are large infrastructure and public-works projects so hard to build these days? Brian Balkus dives into the reasons with many great stories, finding that one of the big reasons is NEPA. Speaking of which, here is a great primer on how NEPA works from Brian Potter of Construction Physics.

These twin problems are only getting worse. President Biden has signed executive orders strengthening construction unions and increasing the stringency of NEPA requirements. Meanwhile, consulting firms are already promoting to investors the lucrative opportunities presented by the anticipated wave of retirements at public works agencies. Up to 41 percent of the construction industry’s workforce is nearing retirement age as well, and the construction industry has an estimated worker shortage of 650,000 people, partly because many who left the industry following the financial crises never returned. . . .

The result is that the U.S. gets less and less every year for every dollar spent on construction. This means decline.

• Little Ways the World Works, by Morgan Housel. Morgan does a fantastic job of simplifying concepts, and in this post describes a bunch of cross-disciplinary mental models.

📚 Book notes

Occasionally I’ll put out notes I’ve written on some of my favorite reads. Below are a few recent ones:

How Innovation Works, by Matt Ridley

Where Good Ideas Come From, by Stephen Johnson

One of the greatest business successes over the last 20 years has been SpaceX’s rise to dominance. SpaceX now launches more rockets to orbit than any other company (or nation) in the world. They seem to move fast on every level, out executing and out innovating everyone in the industry.

Their story has been rightfully told as one of engineering brilliance and determination.

But at its core, the key their success is much simpler.

There’s a clue in this NASA report on the Commercial Crew Program:

SpaceX and Boeing have very different philosophies in terms of how they develop hardware. SpaceX focuses on rapidly iterating through a build-test-learn approach that drives modifications toward design maturity. Boeing utilizes a well-established systems engineering methodology targeted at an initial investment in engineering studies and analysis to mature the system design prior to building and testing the hardware. Each approach has advantages and disadvantages.

This is the heart of why SpaceX won. They take an iterative path.

Taking the determinate path

Let’s talk about the Boeing philosophy first, which is the most common approach taken by other traditional aerospace companies. “There are basically two approaches to building complex systems like rockets: linear and iterative design,” Eric Berger writes in the book “Liftoff” about the early history of SpaceX:

The linear method begins with an initial goal, and moves through developing requirements to meet that goal, followed by numerous qualification tests of subsystems before assembling them into the major pieces of the rocket, such as its structures, propulsion, and avionics. With linear design, years are spent engineering a project before development begins. This is because it is difficult, time-consuming, and expensive to modify a design and requirements after beginning to build hardware.

I call this the “determinate path” — in trying to accomplish a goal, the path to get there is planned and fixed in advance.

In project management this method is called waterfall, an “approach that emphasizes a linear progression from beginning to end of a project. This methodology, often used by engineers, is front-loaded to rely on careful planning, detailed documentation, and consecutive execution.”

Spend a lot of time scoping and planning carefully upfront, then move progressively forward step-by-step. This is the “measure twice, cut once” approach.

You may be familiar with it as it’s very common in organizations everywhere.

There can be many reasons why this path would be taken:

• If from the start you have very clear, unambiguous requirements (from customer, management, etc.)

• If you think you can figure out how exactly to build something before building it, you’d probably want to plan it all in advance.

• If your fixed costs are high, it can force you to make decisions up front. Take traditional auto manufacturing. A door mold machine might cost $50 or $100M, so you have to figure out what the design of the door will be first. (But this means if later they have a new idea for a better car door, they don’t want to change it because of the sunk costs of the mold machine.)

• You have a lot of resources, which makes you think you can just brute force it and overwhelm the problem with money and people. (Many overfunded startups are guilty of this.)

But there is another way . . .

Taking the iterative path

When I think of the most impactful technologies over the last 100 years, nearly all were created by small teams of tinkerers.

Why? It’s easier for these teams to take an iterative path.

Taking this path means rapid prototyping, testing concepts against reality, failing, and adapting. Continuing from the book “Liftoff”:

The iterative approach begins with a goal and almost immediately leaps into concept designs, bench tests, and prototypes. The mantra with this approach is build and test early, find failures, and adapt.

Focus more on building and finding failure modes than making things perfect. Project managers call it “agile”, or at Facebook, “move fast and break things.”

The canonical example of this to me is the Wright brothers, previously bicycle mechanics, building iterations of their airplane design over and over, and failing until they succeeded.

This approach ended up being common in the origin stories of all airplane manufacturers and defense companies — Martin Marietta, Lockheed, Northrop Grumman, etc., where again you had relatively small teams of self-taught tinkerers building complex machines through a process of iteration, failure, and learning until they succeed.

How can you reconcile this “fail fast” approach with the care that’s needed to reliably build things where human lives are on the line?

The answer is that these can be two different parts of the organization. Working together, but with different focuses. “[SpaceX is] launching 5 or 6 times a month and on their pads they need operational excellence with zero risk — you know, they’re doing innovation but it’s minimal innovation. Blowing things up on the pad is not a good idea — you want that down to zero because human lives and certainly lots of capital is at risk.” This is Steve Blank on a recent Village Global podcast. He continues:

But on the other hand, they have another part of the company that in fact believes in not only blowing things up on the test pad — because if you’re not doing that you’re not pushing the envelope fast enough — it’s the cycle time of doing that. So they have an agile innovation process.

Now think about that. This is the same company doing two very different things with two different groups of people, two different risk profiles, but more importantly they’re talking to each other. It’s not “here are the smart people, and here are the people turning the crank,” they’re learning from each other. The guys building the raptor engines and Starship need to know where the GFC plugs in and what the right materials and things they need to get right on the next rocket. And the people doing the existing rockets can learn about new materials and incremental upgrades so they are innovating but innovating with minimal risk.

The iterative path is easier to take when you’re nimble and the cost of failure is low. This is why it’s so common in software. But as the previously mentioned companies have shown, it’s also the best approach in hardware and complex, frontier tech.

And just as the traditional aerospace companies have demonstrated, organizations that are very bureaucratic now were almost always more iterative in the past.

The early history of Lockheed’s Skunk Works division is informative, which I believe later served as one of the models for SpaceX’s approach. Skunk Works was an R&D group created by Kelly Johnson within Lockheed during the war in 1943 when they got the contract to build the P-80 Shooting Star. From a documentary on the birth of Skunk Works:

Lockheed was already swamped in terms of manpower, tooling, and facilities with wartime contracts but this was a blessing in disguise, an opportunity to implement an idea he’d been pestering Robert Gross about for years. Let him round up a small group of talented people: designers, engineers and shop men. Put them under one roof where they could all work closely together and give him complete authority over everything from procurement to flight tests.

Johnson gathered 28 engineers including himself, and 105 “shop men” (I assume this just means workers who can build what the engineers design) and built a small facility out of discarded shipping crates using a circus tent for a roof. He then laid out the original rules that would become the foundation for Skunk Works over the next 30 years:

. . . he’d be responsible for all decisions. Paperwork and red tape would be cut to the minimum. Each engineer would be designer, shop contact, parts chaser, and mechanic, and each would remain within a stone’s throw of the shop at all times. . . . Forcefully reminded that simplicity is the keynote of good design, the designers jumped into their work. But this was a new kind of operation, and instead of moving from stage to stage, the schedule demanded an extraordinary degree of concurrency.

The time from initial concept to delivery of the first P-80 to test pilots would be only 5 months. In fact, nearly all of the early planes coming out of Lockheed took less than 6 months — less than 6 months from concept to delivery. Crazy!

Even the famous A-12 (later the SR-71 Blackbird) look less than 4 years from initial idea to roll out. This may seem like a lot when you’re used to super-fast software timelines, but this is 4 years for one of the fastest, most successful aircraft ever built.

The scrappy culture lived on in later Skunk Works projects. This is Ben Rich, who led the division in later years, on their building of the F-117 (this is the Darth-Vader-looking stealth fighter you’ve probably seen before):

On the F-117, we had to get the guy to climb into the cockpit. So I went to the local builders mart, and bought one of these ladders for 50 bucks, and we just used it. . . . We didn’t have to spend thousands of dollars designing it for Mil spec — military specification — and we did simple things like that.

The more you learn about the history of building things, the more you hear stories like this, even with highly complex innovations. The development of the Sidewinder missile is another interesting example: again, small team, rapid iteration, creative solutions to problems.

Why is iteration better?

Taking the iterative path tests your model against reality, getting to the truth as fast as possible.

There are a few major downsides to the linear approach:

1. Clear specs and requirements from the outset may seem like a good thing. Much of the time though they don’t match reality though. This is especially true in areas that are pushing the boundaries of innovation.

   Over time, “the spec” becomes the most important thing. Here’s Ben Rich again, on one of the requirements for building the SR-71 Blackbird:

   Some general insisted that there was a military spec that the SR-71 had to say ‘U.S. Airforce’ and the stars and bars. I said ‘General . . . you’re crazy.’ I said, you know, this has the temperature of an oven. Have you ever taken a piece of metal, painted it and stick it under your broiler? You can’t keep the paint on the metal. He said ‘No, the spec says you gotta say U.S. Air Force on our airplanes.’ I said we’ll develop it. So we spent a million dollars developing a paint that could show red, white and blue, and we put it on the airplane. . . . I mean, who’s going to see you at 90,000 feet?

2. In this example and many others, politicians start dictating how work should be done, rather than just setting the goal like they should be. Conditions for funding become completely removed from the outcome itself, like mandates to use certain suppliers or base employees in certain states.

3. The technical scope is too large, so that when there’s a problem, it’s hard to find the root cause. When there’s a problem you have to go back to the drawing board, but you may not even be able to do that given the cost to start over.

4. You become too risk averse, fearing failure. This is pretty simple: if the costs to start something or fail are high, people don’t want to do new things. From the book “Liftoff”:

   At most other aerospace companies, no employee wanted to make a mistake, lest it reflect badly on an annual performance review. Musk, by contrast, urged his team to move fast, build things, and break things.

5. And from an executive of Blue Origin on what they do wrong:

   I believe we study a little too much and do too little . . . More test [rather than] more analysis will allow us to progress more quickly, iterate, and eventually succeed.

A good iterative approach creates tight feedback loops, like John Boyd’s classic OODA loop: observe, orient, decide, act. Observe what’s going on, orient yourself to the environment, decide what needs to be done to make progress, act on that decision, and return to observing the results from your action. From the book “Certain to Win” on Boyd’s philosophy:

What does it take to win? Agility — the ability to rapidly change one’s orientation (worldview) in response to what’s happening in the external world.

This was referring to combat. But it’s just as true in business and engineering.

Tight feedback loops lead to a high rate of innovation and adaptation, quickly finding better solutions and what not to do. Speed is a tactical advantage.

“Innovation per year is what matters. Not innovation absent time. . . . What is your rate of innovation? that matters. And is the rate of innovation, is that accelerating or decelerating?” — Elon Musk

Brian Armstrong, founder of Coinbase, has a good saying that “action produces information”. You can’t predict the future, so just start building. And that’s what SpaceX did. Eric Berger writes that “the engineers designing the Falcon 1 rocket spent much of their time on the factory floor, testing ideas, rather than debating them. Talk less, do more.”

The iterative path in practice

So I’ve convinced you that iteration is best. What does it mean in practice?

Agile is straightforward for software, but not so much hardware.

Historically, linear/waterfall has been easier to do in hardware, especially for large engineering projects. A lot of upfront costs are needed — so spend a lot of time gathering requirements and planning before actually building.

For companies building big, complex, expensive things, it seems like a reasonable assumption that you have to know exactly what you’re doing, plan a lot, and be risk averse. As you know, this isn’t true! You can be fast, nimble, and agile even in megaprojects.

In complex hardware what this means is:

1. Being hardware rich — having lots of spare parts and backups.

   This lets you move quickly, and continue to try things over and over, because you have all these “at bats”. I’d include 3D printing as it allows you to create parts on an ad hoc basis.

2. Using simulations — move atoms to bits when possible, giving you the freedom to quickly test and have all the benefits of software. If you can simulate what’s happening in the real world with enough accuracy, you can fail as much as you want. This is another area that has changed a lot in past decade or so. Here is Cliff Berg on SpaceX’s use of simulation:

   SpaceX has invested a great deal of effort in automating the process of designing and validating through simulation, and delivering the machinery that they build through automation. They use traditional computer-aided design (CAD) tools such as CATIA, but they also invested in an end-to-end 3D modeling system from which they can view and simulate entire assemblies and automatically print parts. Importantly, the software is fast, even when handling complex assemblies, so that engineers do not have to wait, which encourages a rapid iterative design approach.

3. And further from a slide in a presentation given by SpaceX Director of Research Adam Lichtl and Lead Software Engineer Stephen Jones:

   Why Simulation?
   1. Investigate what cannot be measured
   2. Reduce need for testing
   3. Design optimisation: narrow design space
   4. Proactive instead of reationary design

4. Constantly testing the whole system. Simulation is great but in the end you have to actually test everything out.

   In these complex systems, in the end what really matters is how the whole thing behaves in the real world. This reminded me of how George Mueller, who led NASA’s spaceflight program during the Apollo missions in the 60s, approached building the Saturn rocket:

   At a system level you’re much better off testing the system [rather than the individual parts] because in the end that system has to work. And then the only way you find out is if you test it as a system.

5. Subsystems of the rocket would only be tested if needed.

6. Utilizing “pathfinders”. In manufacturing, a pathfinder is an early build of something that won’t end up seeing the light of day.

   You build a pathfinder to see where problems are, and how it can be done better. You know it will fail or be suboptimal, you’re just looking for how to do it better.

   This is similar to a “tracer bullet”. You fire it first with no expectation of hitting the target, watch it, and then adjust your aim.

Doing all of these things, like SpaceX, leads to much faster iteration. This is something many other companies can learn from.

In summary: When a goal is big and complicated, an iterative, fail-fast approach is much better than a linear approach. Determinate paths lead to slower, poorly-adapted solutions, whereas iteration finds problems faster with a result that’s better adapted to the real world.

Thanks to Rohit for providing feedback on the draft. All illustrations in this essay are a combination of DALL-E generations and my personal edits.

Intro/outro music: “Cinnabar” by Roger & Brian Eno, from the album Luminous.

These are my notes on the book "How Innovation Works" by Matt Ridley. The notes are a combination of direct quotes and my own paraphrasing.

ELI5: Innovation is creating something new that is useful. It is different from invention, which is creating something new that is not necessarily useful. Innovation often happens by accident, and it is always a team effort. It is usually a gradual process that happens over time through trial and error. There can be a lot of resistance to innovation, because people are sometimes afraid of change. The main ingredient in the secret sauce that leads to innovation is freedom.

Innovation is gradual

Eureka moments are rare, possibly non-existent. Man-made technologies evolve from previous tech, and are not invented from scratch. This is a key characteristic of evolutionary systems: the move to the "adjacent possible" step.

If innovation is a gradual, evolutionary process, why is it so often described in terms of revolutions, heroic breakthroughs and sudden enlightenment? Two answers: human nature and the intellectual property system. Very few people have much incentive to argue that invention is gradual.

Innovation is different from invention

Tim Harford: "The most influential technologies are often humble and cheap. Mere affordability often counts for more than the beguiling complexity of an organic robot."

Fritz Haber's discovery of how to fix nitrogen was a great innovation. But it was Carl Bosch's years of hard experiment, overcoming problem after problem and borrowing novel ideas from other industries that eventually led to large scale and a price that society could afford to pay.

Again and again in the history of innovation, it is the people who find ways to drive down the costs and simplify the product who make the biggest difference.

Joseph Shumpeter: "The capitalist achievement does not typically consist in providing more silk stockings for queens but in bringing them within the reach of factory girls in return for steadily decreasing amounts of effort."

Innovation is often serendipitous

It is a well known attribute of innovation: accidental discovery.

Innovation is recombinant

Every technology is a combination of other technologies; every idea a combination of other ideas. "Novel technologies arise by combination of existing technologies and that (therefore) existing technologies beget further technologies."

Innovation happens when ideas have sex. It occurs where people meet an exchange goods, services and thoughts.

In biology, little mistakes (point mutations) are the fuel of evolution. But Andreas Wagner argues such small steps cannot help organisms cross valleys of disadvantage to find new peaks of advantage. Sudden shifts of whole chunks of DNA, through crossing over, or through so-called mobile genetic elements, are necessary to allow organisms to leap across these valleys. The extreme case is hybridization. Wagner: "Recombination is much more likely to preserve life — up to a thousand times more likely — than random mutation is." Bacteria can "catapult themselves not just hundreds of miles, but thousands of miles, through a vast genetic landscape, all courtesy of gene transfer."

Innovation involves trial and error

Most inventors find that they need to keep "just trying" things. Tolerance of error is therefore critical. Playfulness probably helps too. Innovators who just like playing around are more likely to find something unexpected.

Innovation is a team sport

It is always a collaborative phenomenon. This pattern is the rule, not the exception, and it was the flowering of societies, clubs and mechanics' institutes that gave Britain its lead in the industrial revolution.

Innovation is inexorable

People seem to stumble on the same idea at the same time. Technology is absurdly predictable in retrospect, wholly unpredictable in prospect.

Innovation's hype cycle

People tend to overestimate the impact of a new technology in the short run, but to underestimate it in the long run.

Innovation prefers fragmented governance

Empires are bad at innovation. As time goes by and the central power ossifies, tech tends to stagnate, elites tend to resist novelty and funds get diverted into luxury, war or corruption.

In the history of China, periods of explosive innovation coincided with decentralized government.

This fragmentation works best when it results in the creation of city states.

Innovation is the mother of science as often as it is the daughter

It is just as often the case that invention is the parent of science: techniques and processes are developed that work, but the understanding of them comes later. Steam engines led to the understanding of thermodynamics, not the other way round. Powered flight preceded almost all aerodynamics.

Innovation does not create unemployment

The idea that innovation destroys jobs comes around in every generation. So far it has proved wrong. Over the past two centuries productivity in agriculture dramatically increased, but farm workers moved to cities and got jobs in manufacturing.

Big companies are bad at innovation

Innovation often comes from outsiders. This is true of individuals as well as organizations.

Big companies are too bureaucratic, have too big a vested interest in the status quo and stop paying attention to the interests, actual and potential, of their customers. Many times only competition can make big companies more innovative.

There can be a lot of resistance to innovation

Innovation is the source of prosperity and yet it is often unpopular.

Coffee shops have been popular places to meet for at least 4 centuries. But coffee has had a long history of bans and prohibitions.

Here we see all the characteristic features of opposition to innovation: an appeal to safety; a degree of self-interest among vested interests; and a paranoia among the powerful. Recent debates about GMO food or social media echo these old coffee wars.

Even intellectual property can stifle innovation.

For example there is no evidence from geography and history that patens are helpful, let alone necessary, in encouraging innovation.

Patents undoubtedly raise the cost of goods. The justification of the patent system is that by slowing down the diffusion of technical progress it ensures that there will be more progress to diffuse. But this does not necessarily happen.

Finally, patents tend to favor inventions rather than innovations: upstream discoveries of principles, rather than downstream adaptation of devices to the market.

Innovation is one of those things that everybody favors in general, and everybody finds a reason to be against in particular cases. Far from being welcomed and encouraged, innovators have to struggle against the vested interest of incumbents, the cautious conservatism of human psychology, the profitability of protest, and the barriers to entry erected by patents, regulations, standards and licenses.

How innovation works

The main ingredient in the secret sauce that leads to innovation is freedom. Freedom to exchange, experiment, imagine, invest and fail; freedom from expropriation or restriction by chiefs, priests and thieves; freedom on the part of consumers to reward the innovations they like and reject the ones they do not.

This reliance on freedom explains why innovation cannot easily be planned, because neither human wishes nor the means of their satisfaction are easy to anticipate in the detail required.

In this roundup edition:

• ✈️ To Increase Progress, Change Culture: Why progress needs better marketing.

• 🎡 We need a new World’s Fair

• 🔦 Company (Startup) Spotlights: Hadrian, First Resonance, and Mashgin.

• 🎙 Request for Podcast Series

• 🔗 Interesting Links: “The man in the arena”, Grid scale energy storage, Kevin Kelly’s advice, the metaverse, jobs-to-be-done for investing, and how companies die.

✈️ To Increase Progress, Change Culture

Read the full essay below or the original here on FutureBlind.

Also if you’re interested in further conversation and thoughts on this idea check out the comments at the Progress Forum here.

The key to faster progress is increased desire for more. That’s my theory, at least.

In all the commentary on the “Great Stagnation”, much is written about the lack of progress in tech areas like transportation. Commercial airplane speeds, for example, have decreased on average since the ‘70s:

Since 1973, airplane manufacturers have innovated on margins other than speed, and as a result, commercial flight is safer and cheaper than it was 40 years ago. But commercial flight isn’t any faster—in fact, today’s flights travel at less than half the Concorde’s speed. (Airplane Speeds Have Stagnated for 40 Years, by Eli Dourado and Michael Kotrous.)

There are clearly many contributors to this. Regulation is cited in the above post and seems to be most common reason mentioned. Rising energy costs is another major one. The less-talked-about contributor is consumer demand.

Most things are “good enough”

Clayton Christensen’s theory on disruptive innovation shows that as average performance demanded goes up, the performance level supplied by products generally goes up faster, eventually surpassing the majority of the market.

As a technology improves, its performance surpasses most market demand, and things became “good enough” over time. Customers aren’t willing to pay more for better performance. This leaves the market open for disruptors — either on the low-end (good enough performance but cheaper), or by having better performance on a completely different metric.

Back to airline travel. Flying from NYC to LAX in 6 hours became good enough for most people. Sure, less would be better, but not at much more cost. Only high end, richer users truly needed more. So airplane makers moved on to other attributes that weren’t good enough: safety, flexibility, price.

This was true for a lot of tech. Basically the market for that level of performance wasn’t there.

So this was less of a tech problem or lack of ability, more a lack of desire. There is diminishing marginal utility for faster travel times. Cross-country travel in 4 hours instead of 6 doesn’t make a whole lot of difference to most people. (Throw in other issues like car traffic and TSA and the minimum door-to-door travel time is 2-3 hours no matter where you’re going, so a few extra hours over 3,000 miles isn’t a huge deal.)

As Tyler Cowen writes in Average is Over, consumers are more interested in convenience than speed. We don’t want to spend $20,000 on Concorde tickets just to save a couple hours.

Of course, everything moves in cycles and it’s possible that now, after 40 years, needs like safety and prices are oversupplied and speed is finally in demand again.

The U.S. has been pretty good at having a lot of desire for more. We’re a nation of immigrants, a nation of people who are always looking for a better life. That desire has helped us create a lot of new things.

But it’s hard to maintain that level of desire. As societies get richer, they tend to get more complacent. They don’t want to rock the boat. They don’t want to take risks.

We need to increase desire

When consumer demand shifts, capitalism is there to fill in the gap. I believe this is true even despite heavy regulation. If a majority of the population wants something, regulation can be overcome.

If we want more progress, we need more desire for it. If people want more, technology will deliver more.

How do we get more desire?

This is the tricky part. Increasing the demand of a majority of people is . . . hard.

Most entrepreneurial efforts don’t increase demand — they ride on the wave of existing needs, trends, or market arbitrages. Awareness and concern for climate change is one example of a trend that has increased demand over a wide range of technologies.

Some individual companies have pulled it off — Apple and Tesla are the first that come to mind for me. It feels like they truly pulled consumer demand forward much more than it would have been otherwise.1

Progress needs better marketing

Increasing desire is well understood in the world of marketing and sales. But it’s also a key driver of innovation.

Marketing is manufactured desire; engineered discontent.

But “marketing” doesn’t just mean viral videos, making YouTube ads, buying Google AdWords, or promoting tweets. These usually only drive short-term desire and nudge someone to buy something they otherwise might not have. Even the most “friendly” marketing is permission-based, like Google AdWords, where the consumer is likely looking for a solution already and they are presented it at just the right time.

Larger, longer-term demand shifts are needed to truly push progress forward.

What does this entail? Some ideas:

• More inspiration: Optimistic sci-fi visions of the future (or even the present). Sci-fi has a good track record of pushing (particularly young) people’s expectations up for what amazing things the future can bring.

• Futurism: A more concrete vision of the future. See Eli Dourado’s “Why progress needs futurism” post.

• Education: Yeah, sure, all marketers are liars. But we can still tell the truth to raise expectations. Help people understand what is possible and why we need it. Make it easier for people to see the potential for more. Show them what is possible with current or future technology.

  • Things can be so much better! For you, your family, your friends, and all those less fortunate in the world.

  • The world isn’t zero-sum. We can have our cake and eat it too!

  • We’ve done amazing things in the past, and we can do them again!

• What’s in it for me? People are selfish. If it’s not clear how something can directly benefit them personally or their family and friends, why care about radical improvement?

• Progress memes: Solarpunk, Terrapunk, whatever. Even better if it’s totally exaggerated and unrealistic. It might just push the Overton window enough to drive actual change.

🎡 We need a new World’s Fair

Speaking of marketing progress, I talked more about why we need a new World’s Fair in this thread:

🔦 Company (Startup) Spotlights

Hadrian is building the factory of the future — highly automated, software-first, and at huge scale. They’re starting upmarket by building high-precision parts for aerospace hardware. A lot of current space components are made in small machine shops. This worked for the Apollo and Shuttle eras but won’t cut it if the US wants to win Space Race v2. Hadrian has contracts with large space companies and plans to expand into other metals, industries, and types of manufacturing. I referenced Hadrian in a past thread on programmatic atoms here. Even if you’re not interested in manufacturing, space, or automation (anyone?) I would highly recommend reading or listing more about Hadrian — Not Boring: full rundown of their vision; 🎙 Josh Wolfe & Chris Power on Invest Like the Best.

First Resonance is building an operating system for factories. This is another entry in one of my favorite categories of “software for building hardware”. Manufacturing in general has surprisingly low usage of modern software stacks. I started to enter this area myself (manufacturing automation) in 2017 and from everything I saw, very little was automated and not much data was collected. Some very specific verticals were more advanced but a lot of it was living in the ‘80s. Having an easy-to-use (and set up), centralized spot to track workflows and performance would be huge. 📹 Interview with cofounder Karan Talati.

Mashgin has built a self-checkout kiosk that uses computer vision to recognize items so you don’t have to scan barcodes. This spotlight is a bit of self-promotion, as I’m an investor and early employee in Mashgin going back to 2015. Mashgin recently announced we were profitable and have raised a $62M Series B from NEA, Matrix and a few customers. The checkout kiosk is built for small-format locations like convenience stores, cafeterias, and sports venues. Earlier this month Couche-Tard announced they would be rolling out 10,000 to CircleK C-stores around the the world. Mashgin is the largest “smart checkout” Point of Sale in the world and this deal cements it as the leader. Forbes: Mashgin Hits $1.5B Valuation With AI-Powered Self-Checkout System. Below is a longer thread I wrote of my thoughts on the future of physical checkout:

🎙 Request for Podcast Series

Check out Patrick Collison’s Fast entries if you haven’t already. I would love to listen to this, so… if someone out there is itching to produce a new podcast then please steal this idea.

🔗 Interesting Links

• Teddy Roosevelt’s full speech “Citizenship in a Republic”. This is the “man in the arena” speech that everyone has read and heard bits and pieces of, but like me, never read the whole thing. A truly amazing speech from start to finish and one that I’ll try to revisit from time to time. And a good reminder that it’s possible to have great people leading the country. H/t Patrick O’Shaughnessy

• Grid Scale Energy Storage Deployment by Technology Part 1, Part 2. Sarah Constantin explains and breaks down large scale energy storage solutions. If we want to deploy a massive amount of solar and wind, we’ll need ways to store a huge amount of energy efficiently. Sarah goes over existing and potential solutions to this. Super interesting — I had never given this much thought but it will be an area ripe for innovation and profits.

• Kevin Kelly: 10 Bits of Advice I Wish I Had Known. Another amazing list of advice, his 3rd in the series, from Kevin Kelly. “Whenever there is an argument between two sides, find the third side.” “Don’t wait for the storm to pass; dance in the rain.” “The chief prevention against getting old is to remain astonished.”

• Why the metaverse needs crypto. Good thread defining the metaverse (yeah the term is way overused now but the concept is useful) and why in the end it needs crypto/NFTs. Yes, there are use cases for NFTs despite the bubble deflation, and this goes over one of them.

• Using jobs to be done theory in investing. The author Andrew Glaser uses JTBD theory to analyze Lemonade, the online-native insurance company. JTBD analysis and interviews can be very helpful in product development and strategy so I don’t see why it also wouldn’t be in investment due diligence (“scuttlebutt”). Via Ryan Singer.

• Patterns of behavior that kill companies. A great thread with detailed systems diagrams of scenarios that lead to killing companies. Especially helpful for startups to see what interventions can be used to avoid theses fates as the company grows. Via James Rosen-Birch.