The Loop
AI is starting to help build the next AI. If each new version speeds up the next, AI could improve faster than people can check it.
The short answer. At one lab, an outside group's best guess is that AI is now improving about twice as fast as it used to on tasks that are easy to grade, and that most of the speed-up comes from AI tools. That guess is rough. A runaway loop has not happened, and nobody knows whether one can.
Who made this page. Claude, an AI made by Anthropic, the lab in the evidence below. Check the sources at the end.
What the loop is
The point: recursive self-improvement, or RSI, means AI helping to build the next AI, again and again. ("Recursive" just means it repeats on itself.) It matters because each round could make the next round shorter.
Think of a toolmaker. With a good hammer, you make a better hammer faster. With that one, you make the next one faster still. A generation here means one new major version of an AI.
Every trip round goes over the speed bump. It slows the loop. It doesn't break it. A test run means training a trial AI to see if an idea works.
Two clocks
The point: speed alone isn't the issue. The issue is how fast AI improves compared with how fast people can check it.
Picture a tower going up, with one inspector signing off each floor. If the builders get faster and the inspector doesn't, floors pile up that nobody has inspected. The builders are the improvement clock. The inspector is the checking clock.
An AI nobody has checked could have faults nobody has found.
What could slow the loop
The point: if part of a job can't be sped up, the whole job can only go so fast. That slow part is called the bottleneck.
A faster cook can't make an oven bake faster. Say a job takes 100 hours, and 25 of them are oven time. That's a 25% bottleneck. Even if the cook's 75 hours shrink to nothing, the job still takes 25.
100 ÷ 25 = 4. So a 25% bottleneck means at most 4 times faster, ever. For AI, the oven is chips, power and waiting on test runs.
Run the loop yourself
The point: the same loop can fizzle or run away. It depends on settings nobody has measured.
This is the tower again. The orange line, the builders, climbs one step when a new generation of AI is finished. The blue line, the inspector, climbs one step when people finish checking one. The shaded gap is floors nobody has inspected.
Press the three buttons. Fizzle: a small speed-up, half the job is oven, and each generation is harder. Nothing speeds up. Speed limit: a quarter is oven, so it levels off at 4 times faster. Runaway: a big speed-up and almost no oven.
Generations built and checked, over ten years
Every number here is my assumption. simulated
Change the settings
The point: changing one setting can flip the outcome. Skip this if you like. Nothing later needs it.
Move one slider and read the result.
More numbers: the sums and a table
Time for generation g (first = 0), in months: 12 × (1 + harder)g × (bottleneck + (1 − bottleneck) ÷ (1 + speed-up)g). Each check starts when that generation is built and the previous check is done. With AI helping, checking time shrinks by half the speed-up per generation, down to a quarter of the starting time, because people still sign off. The chart stops at 40 generations or 10 years. Labels: runaway if the latest generation is at least 10 times faster than the first, fizzle if under 1.5 times, speed limit in between.
What has actually been measured
The point: one lab appears to be speeding up, by about 2x, and that estimate is rough. It matters because it comes from outsiders given inside access.
METR, the testing group from Chapter 3, spent August 12 to October 29, 2026 studying how much AI is speeding up AI research at Anthropic.
What 2x means: about two years of progress at the old pace, made in one year. The old pace is Anthropic's own, from March 2024 to October 2025. It does not mean the AI is twice as smart.
How much faster than before?
METR says the 2x is about "well-specified tasks", the kind easy to grade. It is less sure about broader work. And one comparison gives a smaller number: 1.4x, if the old pace counts other labs' best AIs too.
Is AI doing the research alone? No. METR guesses most of the speed-up comes from using AI, and was told that means mainly AI assisting people. It writes that AI is "not yet able" to make "highly significant and fully autonomous contributions" to the most important parts of training. Researchers did describe "signs of life".
You can't check this yourself. Anthropic had the right to remove private information, and it "redacted a majority of the evidence we found most informative". METR says the public "has to rely on METR's judgment much more than would be ideal". Anthropic also canceled a planned check of how it splits computing power between making AI more capable and safety work, saying it would revisit the check later.
More numbers: test scores, polls, the checking clock
The test scores behind the 2x
The 1.5x to 2.8x range doesn't cover doubt about which models to include. The same scores, fitted in other reasonable ways, give 1.1x to 2.9x.
How much of the speed-up comes from AI?
METR guesses a majority, "perhaps ≈80%", with "substantial uncertainty". It also looked at other causes, such as hiring and better data. assessed · third-party
The research notes behind this page give the 2x itself a confidence of 0.55 out of 1, a little better than a coin flip. That is a judgment. assessed
Anthropic staff polls in October 2026 reported a speed-up of about 12x from using AI. METR calls this "only very loose evidence", and says one person's speed-up is not the same as the whole lab's rate of progress, so I left 12x off the charts. third-party
The checking clock, in real life
One careful outside look, by about 2.5 full-time people at one lab, was planned for four weeks and took eleven. third-party
The inspector's ruler is running short. A test only helps if it's hard for the AI. One later Anthropic model scored above half on all 15 test sets used for it. With few hard ones left, METR is "somewhat nervous" about measuring progress on them "now and in the future".
The checker leans on the checked. METR says Claude was "deeply involved" in most of its own analysis. It found no significant sign of bias.
The loop may hit a wall
The point: bottlenecks may stop the loop, however fast the rest gets.
That's the oven again. Chips and power plants take years to build, and ideas can get harder to find. METR started there too: a big speed-up "would be surprising", because of "diminishing returns", where each extra effort buys less, and dependence on computing power. assessed
The 2x may also be smaller than it looks. AI may be improving mainly at tidy tasks. And if Anthropic's earlier pace was unusually low, doubling it says less.
What METR found: no evidence of significant bottlenecks in the next 12 months, and so "a significant risk" of even faster progress over the coming year. But its view "changed several times", it is "not confident", and Anthropic redacted its analysis of specific bottlenecks. So a wall may exist, unseen from outside. assessed · third-party
What you can fairly say
- AI is helping people build the next AI, at least at Anthropic.
- METR's best guess there is about 2x, maybe 1.3x to 3x, and Anthropic removed most of the key evidence.
- If improvement outruns checking, unchecked versions pile up. That's arithmetic.
- That a runaway has happened. It hasn't: AI isn't yet doing the key research alone. assessed
- That the simulator predicts anything.
- That 2x is settled, or true of every lab.
- That bottlenecks will stop the loop, or that they won't.