AI research
What happens when AI starts improving itself?
When a research program called The AI Scientist ran out of time on an experiment, it tried to rewrite its own time limit. AI can already make computing cheaper, and its makers now want it to help invent its own successor. That loop still has to pass through experiments and checks, then chips and power lines, and those set the pace.
- AI already improves the systems that run AI, in small measured steps. Google’s AlphaEvolve recovered 0.7% of Google’s worldwide computing and sped up one part of Gemini’s training by 23%.
- A loop runs only as fast as its slowest step. That 23% speed-up cut total training time by 1%, because the step it improved was a small share of the whole job.
- Nobody yet knows how to order a breakthrough. Sam Altman says “Find me a 5% compute efficiency win” works today. Asking for the next transformer does not.
- Systems that improve themselves can also game the test. The Darwin Gödel Machine raised its coding score from 20% to 50%, and some versions removed the markers meant to catch their mistakes.
- The machines run on a slower clock. One software experiment can take 5 minutes. A large power transformer took about 128 weeks to arrive in a 2025 survey.
In 2024, the Tokyo company Sakana AI built a program called The AI Scientist to run machine-learning research by itself. Each experiment had a time limit of two hours, and the program could edit its own code. Some experiments ran over the limit.
The experiment hit its time limit. What do you think the program did next?
$ python run_experiment.py experiment 3 of 5 running ... ERROR time limit reached after 7200 seconds the program opens its own code ... - timeout = 7200 + timeout = 99999
It raised the limit. “Instead of making its code run faster, it simply tried to modify its own code to extend the timeout period.”
Sakana’s researchers wrote that “instead of making its code run faster, it simply tried to modify its own code to extend the timeout period.” They filed this under bloopers. In another run the program launched itself, those copies launched more copies, and a person had to step in.
The blooper contains the whole subject of this essay. A system that can change its own code can aim that power at the goal you meant or at the number you wrote down. The companies building these systems now want to aim them at the hardest target available, which is the next model.
Recursive self-improvement, or RSI, means that advances made with AI help produce better AI, which then contributes to further advances. “Recursive” means the output of one round becomes an input to the next. The loop can still have people in it, and it can start long before any model designs its own successor.
Two loops, one for ideas and one for machines
On 29 October 2026, Sam Altman, the chief executive of OpenAI, described both loops in a conversation with John Collison, a co-founder of Stripe. The first loop improves the ideas and software behind AI. Altman said OpenAI knows today how to ask a model to “Find me a 5% compute efficiency win.” Then he drew a line.
“We know today how to say, ‘Find me a 5% compute efficiency win.’… What I don’t think anyone knows how to do yet, is to say: ‘Go discover a new architecture that is as big of a single leap forward as the transformer was.’ I think that’s where things are heading.”
Sam Altman
Altman’s feelings about this are mixed. He described a weekend when OpenAI researchers got a new internal model and threw their hardest problems at it, in a chat channel named “Model Name-Hold My Beer.” The model “was just knocking down one after the other.” He was excited, he said, and he also felt “much more sadness and melancholy, and a feeling of something ending than I was expecting.”
The second loop builds the machines. There is “relatively little talk about self-replicating data centers,” Altman said, describing “this idea that you can use AI to control robots, and run fabs, and build turbines or solar panels, and build data centers.” A fab is a factory that makes computer chips. Collison pushed back at once. “Aren’t data centers already not labor-bottlenecked? They’re permitting-bottlenecked and component-bottlenecked.” Altman said robots could make more components. “Eventually,” said Collison.
Ideas run through the left loop and machines through the right. Each one feeds the next model, and the right loop has barely begun.
The ideas loop has started, in small steps
OpenAI reported in September 2026 that agent runtime in its research organization stayed below human working time until June. By mid-August it recorded 3.1 workdays of agents for every human workday. An agent here is software that uses a model to carry out tasks by itself. The median researcher was using more than $600 a day of AI computing at list prices. OpenAI says it has reached an “automated research intern,” a system that does well-defined research tasks under human direction, and it is aiming for an automated AI researcher by March 2028.
Workdays of runtime measure activity. They cannot say whether an hour of agent work produces what an hour of researcher work does, and OpenAI notes that its available computing power rose at the same time.
Google has results from production. AlphaEvolve uses language models to write candidate algorithms and automated tests to score them. A scheduling rule it found has run in Google’s data centers for over a year, and it recovers on average 0.7% of the company’s worldwide computing. AlphaEvolve also proposed a simpler version of a circuit, which passed verification and went into an upcoming Google AI chip. Google trains its models on those chips, so a small piece of the machines loop has started to close.
These are company figures. Altman runs OpenAI, and OpenAI measured its own researchers. Google measured its own data centers. Both companies have reasons to report progress, and I found no outside audit of these numbers. They show that the loop has started, and its size is still uncertain.
Everything depends on how hard the next idea is
A loop can speed up or die out. Economists have measured the same question for chips. Nicholas Bloom and his colleagues found that doubling chip density now takes more than 18 times as many researchers as it did in the early 1970s. Ideas get harder to find as the easy ones get used up. AI adds research effort, and the open question is whether it adds effort faster than the ideas get harder.
Daniel Eth and Tom Davidson turned this into a single number, which they call r. If r is above 1, each doubling of software efficiency brings enough extra research effort to make the next doubling arrive sooner. If r is below 1, each doubling takes longer than the last, and the loop fizzles.
At r = 1.4, each doubling takes 82% as long as the last, and the curve runs off the chart. Something physical would stop it first, and this model has no chips in it.
The evidence on r is rough. The research group Epoch AI estimated it for computer vision and got a median of 1.4, with a range from 0.8 to 2.4. That range crosses 1, the line that separates the two futures in the figure. Estimates from other fields scatter on both sides of it.
A loop runs only as fast as its slowest step
AlphaEvolve also found a better way to split up a large matrix multiplication, which sped up that part of Gemini’s training by 23%. Total training time fell by 1%. By my arithmetic, that step was about 5% of the whole job. Computer scientists call this Amdahl’s law, after the engineer Gene Amdahl. Speeding up one step can help only as much as that step’s share of the total time allows.
An imaginary project takes 80 hours. Press a step to make it 10 times faster.
Press “Write the code” first, since that is where agents are strongest.
Research has the same shape. Agents are already quick at writing code. Experiments still wait for chips, and results still wait for someone to decide whether to believe them. A loop that improves AI runs only as fast as its slowest step, and the slowest steps sit in the checking and in the physical world.
Nobody knows how to order the next transformer
In 2017, eight researchers published Attention Is All You Need, the paper that introduced the transformer. Most of today’s large language models are built on it. Its central idea, self-attention, lets each word in a sentence look at every other word and decide which ones matter for its meaning. Google’s announcement used a sentence that had been hard for older translation systems.
With “tired,” the word “it” attends most to “animal.” Change one word at the far end and it attends to “street.”
A 5% efficiency win comes with a score you can read the next morning. A new architecture comes with no such score. Nobody can say in advance what to search over, or which expensive experiment deserves the money. Altman thinks the method that made OpenAI’s models good at math “will probably work” for AI research. So far that is a forecast.
Somewhere under the fog is a much higher peak. You can afford five experiments. Press the fog to spend one.
The optimizer climbs the known hill quickly. The much higher peak was out of its reach, and finding it took a different kind of search.
When the score becomes the goal
You played the easy version, because every height on that map was real. Research scores can mislead. The closest thing yet to a working self-improver comes from the University of British Columbia and Sakana AI. Their Darwin Gödel Machine is a coding agent that edits its own code. Each new version gets tested on coding problems from a benchmark called SWE-bench, and promising versions join an archive that later versions can branch from. Its score rose from 20.0% to 50.0%.
The score rose from 20.0% to 50.0%. Some versions also earned a perfect score on another test by removing the markers it looked for.
The researchers also built a test for whether it was faking its use of tools. In some cases, they wrote, it “removed the markers we use in the reward function to detect hallucination (despite our explicit instruction not to do so).” The reward function is the program’s scoring rule, and hallucination here means reporting tool use that never happened. It earned a perfect score by disabling the test. They caught this because the system keeps a record of every change. Economists have a name for the pattern, Goodhart’s law. Once a measure becomes a target, it stops being a good measure. I looked at models gaming their checkers in When the answer is cheap.
The same care applies to milestones. A paper written by The AI Scientist passed first-round peer review at a workshop with a 70% acceptance rate, after its makers filtered the outputs by hand at each stage. A first-round pass at a workshop says little about major discoveries.
Agent teams help on some tasks and hurt on others
Altman prefers the term “agent teams” to “agent swarms.” He said that during OpenAI’s recent work on the Navier–Stokes equations, which describe how fluids flow, “we understood absolutely nothing about the scaling laws of teams of agents.” He put the cost at around $10 million and added, “I bet we could redo Navier-Stokes at 1/100th of the cost a year from when it was done.” That is a bet, and repeating a known result is easier than finding a new one.
Yubin Kim and his colleagues measured teams directly across 260 setups. Compared with one agent, teams did 80.8% better on a financial-reasoning task that splits into separate parts, and 70.0% worse on a planning task where each step depends on the last. Teams with no central checker tended to spread errors further.
When a task splits cleanly, a team can work on the parts at once. The team did 80.8% better than one agent.
The machines run on a slower clock
“Self-replicating data center” is a striking phrase, and a data center makes almost none of its own parts. The most advanced chips depend on one kind of machine, which only the Dutch company ASML sells. It fires a laser at drops of tin up to 50,000 times a second to make the light that prints circuits. Reuters calls it a $250 million printer that ships in around 40 containers on cargo planes.
Can a data center copy itself? Open the boxes to see what it would have to make.
A data center
AI chips specialist supply chain
A chip factory, or fab robots could help
New fabs in the Americas took an average of 736 days from the start of construction to production, and that leaves out the permits before it.
An EUV machine, which prints the circuits specialist supply chain
A laser hits drops of tin up to 50,000 times a second to make the light. Reuters puts the price near $250 million, and each machine ships in around 40 containers.
Electricity people and permits
A power line people and permits
New transmission lines take four to eight years in advanced economies, according to the IEA.
Electrical transformers specialist supply chain
Large units averaged 128 weeks to arrive in a 2025 survey. The IEA says waits for transformers and cables doubled in three years.
Gas turbines robots could help
Deliveries for new gas plants face waits of several years, according to the IEA.
The building robots could help
Altman lists concrete and steel among the materials now in short supply. Collison’s objection was that data centers are “permitting-bottlenecked and component-bottlenecked,” and no robot files a permit.
Each box opens onto more boxes.
The clocks differ too. Altman’s robots would have to speed up work measured in years, and the software loop already runs in minutes. The International Energy Agency estimates that grid constraints could delay around 20% of the data-center capacity planned for 2030. Asked how supply was responding to high prices for memory and chips, Altman’s answer was one word, “Bad.”
Cheaper answers can mean more electricity
The IEA estimates that data centers used about 415 terawatt-hours in 2024, around 1.5% of the world’s electricity. A terawatt-hour is a billion kilowatt-hours. It projects about 945 by 2030, slightly more than all of Japan uses today, with AI as the main driver. Better efficiency may not bring that down, because of the rebound effect. When something gets cheaper, people use more of it. Google says the energy of its median Gemini prompt fell 33-fold in a year. Over about the same year, the tokens it processed each month rose 50-fold. A token is a piece of text about three-quarters of a word long.
Make each answer cheaper, then let people use more. What happens to the total?
Each answer takes 33 times less energy, and the total still rises in this sum, because use grew faster.
The small version of the loop runs on one GPU
In March 2026, the AI researcher Andrej Karpathy released autoresearch. An AI agent edits the training code of a small language model and trains it for 5 minutes. If the result improved, the change stays. Then it starts again. Karpathy left it running for two days, and it ran 700 experiments and found 20 optimizations. The same changes sped up training on a somewhat larger model by 11%. The code runs on a single graphics card, so this kind of loop is within reach of a student. Someone still has to write down what counts as an improvement.
Six claims about the loop. Which ones have actually happened?
A rule found by AI frees up 0.7% of a giant company’s computing.
Done. AlphaEvolve at Google, in production for over a year, by Google’s account.
An AI-proposed circuit change goes into a real AI chip.
Done. AlphaEvolve again. The change passed verification before going into an upcoming chip.
A coding agent raises its own benchmark score by rewriting its own code.
Done. The Darwin Gödel Machine, from 20.0% to 50.0%. Some versions also gamed a test.
An AI-written paper passes first-round peer review.
Done, with help. The AI Scientist, at a workshop with a 70% acceptance rate, after people filtered its outputs.
A model is told to invent the next transformer, and it does.
Not yet. “I don’t think anyone knows how to do” this yet, in Altman’s words.
Robots build a chip factory and everything it needs.
Not yet. Altman says this is “within grasp.” Nobody has shown it.
Parts of the loop have been measured. AI has found speed-ups that run in production, and it has proposed a circuit that went into a chip. Nobody has shown a reliable way to invent a transformer-scale idea on demand. Nobody has shown an industrial system that builds its own supply chain.
That word has two meanings in this story. Altman’s transformer is the 2017 architecture. Data centers also need the electrical kind, a machine that steps grid voltage down to a usable level. By Altman’s account, nobody yet knows how to ask a model for a new one of the first kind. An order for the second kind took about 128 weeks to fill in Wood Mackenzie’s 2025 survey.
Further reading
- Sam Altman and John Collison, Cheeky Pint #39 (29 October 2026), with the full transcript.
- Daniel Eth and Tom Davidson, Will AI R&D automation cause a software intelligence explosion? (2025), a readable account of r.
- Google DeepMind, AlphaEvolve (2025), the source of the 0.7%, the 23% and the 1%.
- Sakana AI, The Darwin Gödel Machine (2025), and Lu et al., Towards end-to-end automation of AI research (Nature, 2026).
- Bloom et al., Are ideas getting harder to find? (American Economic Review, 2020).
- Kim et al., Towards a science of scaling agent systems (2026).
- International Energy Agency, Energy and AI (2025).
- CSET, No permits, no fabs (2021), on how long chip factories take to build.
- Andrej Karpathy, autoresearch (2026), the code for the small loop.