The Summit

We didn't stagnate after Apollo. We reached the top of one mountain and spent fifty years walking to the next.

There’s been a thought that’s been rattling around in my head for a decade or so since my colleagues in the tech world started calling anything orbiting around the decades old machine learning domain AI. The thought’s genesis was there long before that though; a few weeks out from graduation while cramming for my Weapons Systems final at the Naval Academy.

Weapons as we called it was the last gauntlet to run for humanities majors; the Academy’s excuse to cram differential equations into the innocent minds of history or English or oceanography majors. Calling it Weapons Systems was a stretch. The only thing Weapons related about it was that the calculations we had to solve could double as the ones used for rockets or other projectiles. So for the last semester of your last year at the Naval Academy, you learn the math that put us on the moon. At least you did in the 1900s when I went there.

If you know anything about that math, it’s hard to not see the connection to the math that ultimately brings us to concepts behind machine learning and the gradient descent of LLMs. Which brings me full circle. The thought that’s hard for me to shake is that it’s all just one long arc. And in that arc I see a crack in the great stagnation narrative.

There’s a story we tell ourselves about the decades after Apollo: that we lost our nerve, that bureaucracy calcified our ambition, that we chose consumer gadgets over cosmic exploration. We look at the 1970s through the 2010s and see stagnation, a civilization that went to the moon and then inexplicably stopped climbing.

But what if we didn’t stagnate at all? What if we simply reached the summit?

The physical world, it turns out, has fewer variables than we thought. When you’re designing a guidance system to land on the moon, you’re working within a remarkably constrained problem space. Newton’s laws don’t change. Gravity is predictable. Orbital mechanics can be calculated on paper. The entire challenge of Apollo, as technically magnificent as it was, operated within a universe of perhaps a few dozen critical parameters that mattered for navigation and control.

Compare that to a game of chess. A chessboard has 64 squares and 32 pieces, yet the number of possible board positions exceeds 10^40. The number of possible games is functionally infinite. And chess is simple. It’s a toy universe with rigid rules and perfect information.

Now consider language. Or images. Or the full space of human knowledge and expression. The dimensional space of possible configurations isn’t just larger than physics. It’s larger by a factor that makes the comparison absurd. When you’re training a neural network to generate coherent text, you’re navigating a probability landscape with billions of parameters, each one a dial that can be turned continuously. The position space of a large language model makes chess look like tic tac toe.

This is the thing we got backwards. We thought the physical world was infinite in its complexity, that we’d spend centuries mastering spaceflight and energy and materials. But the physical world turned out to be the easy problem. Not because physics is simple, but because it’s constrained. The laws are few. The variables are knowable. You can write them down.

The abstract world, the world of information, pattern, meaning, and representation, is where the real exponential lives. And we couldn’t touch it until the hardware caught up.

People frame the post Apollo era as a failure of will, but that’s not quite right. We stopped climbing Everest because we’d reached the top. What else were you going to do? Stand on the summit and jump?

After Apollo, we built Space Shuttles and space stations. We sent robots to every planet. We put telescopes in orbit that could see to the edge of the observable universe. All of this was impressive, but none of it was fundamentally more difficult than what we’d already done. It was lateral motion, not vertical. The engineering challenges were real, but they weren’t deeper. They were variations on a theme we’d already mastered: trajectory, propulsion, life support, navigation.

We could have gone to Mars. We still can. But Mars is just the moon with more fuel required. It’s the same equations, the same closed form solutions, the same explicit control systems. It’s not a new mountain. It’s the same mountain with a longer trail.

So we did what any reasonable civilization would do: we stopped pouring infinite resources into climbing the same mountain and started looking for the next one. The fact that the next mountain wasn’t visible yet, that it required fifty years of semiconductor advancement before we could even see its base, doesn’t mean we were standing still. We were building the ladder.

The stagnation narrative only makes sense if you think the only meaningful progress is the kind that involves rockets. But rockets, it turns out, were just the first act. They were what we could build with the physics we understood and the compute we had. When both of those constraints were maxed out, the frontier moved somewhere else.

Here’s the thing about guidance systems: they work because the universe is compressible. You can describe the motion of a spacecraft with a handful of differential equations. You can model aerodynamics, gravity, thrust, and drag with maybe a few hundred parameters if you’re being really thorough. Everything else is error correction and noise management.

This is what made the Apollo Guidance Computer possible. It had 2K of RAM and ran at one megahertz, but that was enough because the problem was small enough to fit. The engineers could write down every equation, every failure mode, every edge case. They could prove the system would converge. They could test it exhaustively.

And because they could do all of that, they did. Apollo wasn’t just a success. It was a complete success. We didn’t leave anything on the table. There was no secret remaining insight into Newtonian mechanics that would have made the guidance computer ten times better. We’d hit the ceiling of what that problem space allowed.

This is the wall that people mistake for stagnation. We didn’t stop progressing in guidance and control systems because we got lazy. We stopped because the problem was solved. You can’t improve a solution that’s already optimal within its constraints. You can make it cheaper, more reliable, more efficient, and we did, but you can’t make it fundamentally more.

The physical world has an edge. We found it.

But while the exterior universe is bounded by physics, the interior universe, the space of thought, language, representation, and meaning, is functionally infinite. This isn’t a poetic metaphor. It’s a mathematical fact. The configuration space of abstract symbolic systems is vastly larger than the configuration space of physical systems.

A neural network with a billion parameters isn’t modeling a billion independent things. It’s modeling the relationships between those parameters, which means the expressive capacity grows combinatorially. The space it navigates is so large that we don’t have the vocabulary to describe it. We just call it high dimensional and move on, because the alternative is to say unimaginably vast every time, and that gets old.

This is why machine learning took fifty years to catch up to control theory. Not because the ideas were missing. Cybernetics and early neural networks were contemporaries of Apollo. But because the compute wasn’t there. You can’t explore an infinite space with finite resources. You need the resources to scale with the space, and for symbolic systems, that scaling is exponential.

So we waited. We built faster chips, denser memory, cheaper storage. We laid fiber and built data centers. We did fifty years of grinding, incremental improvement on hardware, and the entire time, people said we were stagnating. But we weren’t stagnating. We were loading. We were building the infrastructure to climb the next mountain, which happened to be invisible from the base.

And then, around 2015 or so, we had enough compute to start training networks at scale. And the moment we did, everything exploded. Not because we invented a new kind of intelligence. Backpropagation is older than the moon landing. But because we finally had enough compute to let it run.

When historians look back a thousand years from now, the period between 1969 and 2020 is going to look like a rounding error. Apollo and GPT will appear almost simultaneous. The gap between them will seem like the gap between Hobbes and Locke. Different thinkers, same intellectual moment, same chapter in the story.

Hobbes published Leviathan in 1651. Locke published Two Treatises of Government in 1689. That’s 38 years. If you’re reading about the Enlightenment in a history book, those two names appear on the same page, sometimes in the same sentence. The fact that Locke was six years old when Hobbes wrote his masterwork doesn’t matter. They’re part of the same movement, the same turn in human thought.

Now zoom out further. The moon landing was 1969. GPT 3 was 2020. That’s 51 years. From the perspective of someone in 2500 or 3000, those events are going to be part of the same transition: the moment when humanity figured out how to build machines that could navigate complexity. First physical, then abstract. One chapter. Same revolution.

The fifty years in between won’t disappear, but they’ll be compressed. The Space Shuttle, the personal computer, the internet, the smartphone, all of that will be context, background, the industrial base that made the second leap possible. But the story will be about two peaks: the peak of explicit human engineering, and the peak of emergent machine intelligence. Everything else is the bridge.

So maybe the right way to think about the last century isn’t as a story of decline or stagnation, but as a story of completion.

We climbed the mountain of physical mastery as high as it could go. We built machines that could fly themselves to the moon using equations we could write by hand and computers we could fully understand. That was the peak. We planted the flag. And then we looked around and realized: there’s nothing higher here.

But there was another mountain visible in the distance. Taller, stranger, harder to climb. A mountain made not of rock but of information. And to climb it, we needed different tools. Not slide rules and differential equations, but matrix multiplications and gradient descent. Not systems we could interpret, but systems we could grow.

The gap between those two peaks wasn’t stagnation. It was the descent from one summit and the long approach to the next. It was fifty years of building the roads, the base camps, the supply chains. It was the preparation required to attempt something harder than anything we’d done before.

And now we’re climbing again. Not with rockets this time, but with models. Not by imposing order on the physical world, but by discovering order in the abstract. The mathematics are the same. Feedback, correction, convergence. But the territory is different. Bigger. Deeper. Stranger.

We didn’t stagnate. We finished one thing and started another. And when the far future looks back, they’ll see it for what it was: a single motion, from guidance to generation, from the world of things to the world of thought.

The same journey. Just harder terrain.