Around 1800, an English engineer named Henry Maudslay built a lathe that could cut an accurate screw thread, and it turned out to be one of the more important things that has ever happened...
Before Maudslay, threads were cut by hand, one at a time, so no two were quite the same. A nut from one workshop wouldn't fit a bolt from the next, so every machine in the world was a one-off that had to be fitted together by a craftsman, and could only be repaired by one.
You couldn't make a thousand of anything, because you couldn't make the parts match. That was the bottleneck on the entire industrial world, and it was sitting in a human hand holding a file.
Maudslay's lathe cut the same thread every time. Which meant one machine could now make the precise parts of another machine, and those parts would fit. Within a generation, the machine tools descended from his were turning out the components for everything else, and the Industrial Revolution found a gear it didn't know it had.
Now, you're probably sitting there thinking, Raoul, why the fuck are you writing about screw threads from the 1800s? Bear with me, because if you look at the long arc, the same shape keeps appearing.
For most of human history, tools improved at the speed of the people who made them, which is to say slowly... a better plough every few centuries. Then tools started making tools, and each loop ran faster than the one before.
Steam engines drove the lathes and boring machines that cut the parts for the locomotives, so steam built the railways that carried more steam engines.
Machine tools built the assembly lines that built cars. Computers now design the chips that go into better computers, because no human can lay out billions of transistors by hand. Every time the tools took on more of the work of making the next tools, everything sped up, because the slow part of the loop, the human part, got smaller.
So what happens when the tool in the loop is intelligence itself?
It's Already Happening
Back in November 2023 I posted that OpenAI had probably worked out how to use AI to ideate, design, code and ship product... that AI was building AI.
As of May 2026, Anthropic says more than 80% of the code it merges into its own codebase is written by Claude, up from the low single digits before Claude Code launched in February 2025. The company building one of the frontier models is mostly having that model write the software the next version runs on.
OpenAI published its own version in September: its research organisation now puts in 3.1 agent-workdays of effort for every day of human labour, and the number of experiments per researcher hit an all-time high in August.
In May of last year DeepMind published AlphaEvolve, a Gemini-powered agent that writes and tests its own algorithms, and among the things it found was a scheduling improvement that continuously recovers 0.7% of Google's entire global compute fleet, along with a speedup to a kernel inside Gemini's own training that cut the training time by 1%. Small percentages, until you register what they are: an AI making the next AI cheaper to build and cheaper to run.
That's Maudslay's lathe, cutting the parts for the next machine.
A Different Kind of Bottleneck
Every loop before this one took the physical work off us. The lathe did the filing, the assembly line did the lifting, and that's the whole reason each one ran faster than the last. But the thinking never left. Someone still had to sit down and design the next lathe. Someone still had to work out what the chip was for. The machines got brilliant at making stuff but the ideas for the next generation still came out of a human head, at the speed a human head goes... which, let's be honest, isn't very fast and doesn't get faster.
That's been the ceiling for all of history and nobody ever thought about it, because what would you even do about it? You can't build a tool that does the thinking.
Except now you can. For the first time, the thing being outsourced is the intelligence itself, and so what caps progress is no longer how fast people can think but two physical things, how much energy you can turn into compute and how much intelligence you can get out of each unit of it. Both of which are on exponential curves.
It's leaving the screen
So far this is a story about software, but it isn't staying there.
David Mattin, who writes The Exponentialist with me, calls the next stage self-driving science. And no, nothing to do with Teslas, although the idea is the same...
What he means is that science has always been a loop that a human drives. You have a hunch, you design an experiment, you run it, most of the time it fails, you look at why, you have a better hunch. Round and round, one person at a time, months per lap. Every drug, every material, every bit of physics we've ever understood came out of that loop, at that speed.
Self-driving science is what you get when you hand the loop to the machine. We're not talking about AI helping scientists... we're talking about AI making the discoveries themselves, the new materials, the new molecules, the things humanity simply didn't know last year, coming out of a machine that had the idea, ran the test, read the result and had the next idea, without a person in the loop at all. Working out how the world works, is being outsourced... and it's happening now, not in some sci-fi future.
In 2023 DeepMind's GNoME model predicted 2.2 million new crystal structures, around 380,000 of them stable. That's more new materials than humanity had found in the entire history of materials science, found by a model, before anyone had walked into a lab.
Then in December 2025 DeepMind and the UK government announced an automated materials lab, opening this year, where robots synthesise and test hundreds of candidate materials a day and the AI picks the next batch. The loop that used to take a research group months per lap now runs every single day, and the humans are there to review the results, not run the experiments.
The same thing is happening to robots, and the way it works is wild. In 2025 DeepMind released Genie 3, which is best understood as a video game that writes itself. Type a sentence, say a cluttered bedroom, and it generates a simulated one on screen, complete with the physics of a real room, so that anything moving around inside it gets the same feedback it would get from the real thing.
Then you can put a robot's brain inside of it and tell it to tidy the room. It fails, so it tries again, and again, millions of times, in millions of simulated rooms, at computer speed, without breaking a single real thing or taking up any physical space. Then you can load what it learned into an actual robot, and it walks into an actual bedroom already knowing how. DeepMind is doing this now with its robot agents, and nobody is writing the lesson plan. The agent sets its own tasks and marks its own work.
If this is the kind of thing that fascinates you, it's what David and I publish every week in The Exponentialist. It started life inside Real Vision and is now its own publication on Substack, and there's a free tier if you want to follow along.
And then it feeds itself
So the machines are writing their own code, improving their own models and starting to do our science for us. Fine, you say, but there's a catch everyone's heard about.... all of this runs on power, and lots of it. Data centres are already straining grids, chips are already some of the scarcest objects on earth, and surely at some point the whole thing just runs out of electricity.
But if a machine can solve problems humans couldn't solve, the first problems it's going to get pointed at are its own. The materials labs I just described aren't hunting for random discoveries. They're hunting for battery chemistry that stores more, a solar cell that converts more, a chip that packs transistors tighter... the exact things that make energy and compute cheaper.
And every one of those feeds straight back into the machine that found it.
Better panels and batteries make energy cheaper, cheaper energy makes compute cheaper, cheaper compute trains better models, and better models find the next material. The bottleneck everyone worries about is the first thing the loop is built to eat.
There's a law underneath this, Wright's Law: every time cumulative production of something doubles, its cost falls by a fixed percentage. Solar has followed it for decades. What's new is that the thing doing the designing is no longer a person, so the doublings come faster and the cost falls faster, on chips, batteries and panels at once.
Every turn of the wheel makes the next turn faster, and there's no outside input it depends on. Earlier versions of this loop took decades to come round once. This one is turning in months, and the thing doing the turning is now itself the thing being improved.
Cheaper means more, not less
OK, so the obvious objection at this point is economic. If intelligence is getting this cheap this fast, surely we'll need less of the stuff that produces it... fewer chips, fewer data centres, less power.
The opposite is happening, and the reason has been understood since the coal age. Jevons noticed that when a resource gets more efficient, total consumption rises, because the cheapness opens up uses that were never viable before. Make the steam engine more efficient and a country burns more coal, not less.
Intelligence is doing exactly this. The cost of a unit of it has been falling by roughly ten times a year for equivalent capability, by a16z's count in late 2024, and consumption has gone through the roof. Google processed about 1.3 quadrillion tokens a month as of October 2025, more than twenty times the year before, and by May this year the figure was 3.2 quadrillion. Every drop in the cost of thinking unlocks a problem that was too expensive to think about before, and there's no ceiling on that demand because there's no shortage of unsolved problems.
This is why the capex can't stop.
"It's hype. It's plateauing."
I hear this every few months, usually right after a release that was a bit less dazzling than the one before it. The excitement lasts a week, then it's back to normal, and normal is where people decide the whole thing is slowing down.
The mental model I use is child development. Watch a toddler failing to walk and you could conclude its legs don't work and it's too stupid for the task, or wait for a child's first words and decide its mouth isn't built for speech. Until it just does it. These models are infants, with superpowers in some areas and badly behind human children in others... but infants learn fast. Judging where they'll be in three years by what they can't do this quarter is the mistake, every single time. In three years we've gone from a bad chatbot to something that scores well above most humans in every subject we teach. Don't midcurve it, as the meme has it, just because you have a bias.
What this means for you
The fear underneath all of this is personal, so let me be personal about it.
I'm not a technical person. In January I gave cowork a PDF and had a working website back in six minutes, then asked it for a dashboard of Real Vision's engagement across every platform and had that in fifteen. I posted at the time that I felt kinda bionic.
Today, one person with a set of agents can now produce what used to take a firm. In my book, The Everything Code (out on November 3rd and on preorder now) I tell the story of a software engineer who spent more than a decade building a company the traditional way, stepped back, and then built a personal AI project over a single winter that hit six-figure GitHub stars and got him hired by OpenAI. It was the same brain applying the same work ethic, the only thing that had changed was what he was building on.
"Does this replace me" is the question your ego asks, and it's the wrong one. The layer of work beneath you is being compressed whether you like it or not, the same way Maudslay compressed the work of every craftsman who filed screws by hand. The better question is what you do with a machine that can carry that layer for you, and the people who end up with the superpowers are the ones who move up a level, to the judgement about what's worth building in the first place.
Own the machines that build the machines
What the loop tells you is where the value pools. When intelligence builds intelligence, the returns flow to whoever owns the loop: the compute it runs on, the energy that feeds the compute, and the rails the agents settle on once they start transacting with each other at machine speed. Those are the things it can't route around, and for the first time in history each of them is something an ordinary person can own a piece of.
This is the Exponential Age moving from theory to lived reality, and it runs straight into the window I've written about in my Economic Singularity framework, somewhere between 2030 and 2032. That's the point where the machine economy starts moving faster than the human instruments built to measure it, GDP, earnings, the wage, and the old numbers stop describing what's happening. The years between now and then are the game.
The full thinking on this is in The Everything Code. It charts why the financial system got so broken and how we've been trapped in a debt cycle for decades, nd then why the Exponential Age arriving right now is the other half of the same story.
Everything in this piece is inevitable once you see the mechanism, and the strange part is that it's the way out: the machines are what finally lets the world grow its way clear of the debt. After decades of being stuck, that is something to be optimistic about, and the book is where I make the case in full.
The Everything Code is out November 3rd and do me a favor, if you're planning on buying it I'd be honored if you would consider preordering it now. These things really help with rankings and it'd mean a lot to me. Thanks.
And if you want to go deeper on any of this, keep chatting to me! I've created an AI trained on 30 years of my thinking over at raoulpal.ai and it's free. Continue the discussion with me there.
