AI Is Now Speeding Up Its Own Research. What Does That Actually Mean?

AI helping to build better AI

For years, one idea has sat at the centre of debates about where AI is heading: that at some point AI would start helping to build better AI, which would then build better AI again, faster each time. This month, for the first time, an independent review has put a number on it. It found that Anthropic, one of the leading AI labs, is improving its models at about twice its previous pace, and that most of the speed-up probably comes from using AI in its own research. In our view, this is worth understanding properly, without either the hype or the panic.

We've taken a closer look at what was found, what it means, and what it changes for businesses using AI today.

What "recursive self-improvement" means

The technical term is recursive self-improvement. In plain English, it means a loop. People build an AI. The AI helps those people build a better AI. The better AI helps even more, and so on.

If each turn of the loop is faster than the last, progress speeds up. That is why the idea matters so much to people who study AI safety. It is also behind well-known thought experiments such as the Paperclip Maximiser, which ask what happens if very capable AI pursues a goal faster than people can check it.

Until now, this has mostly been theory. What has changed is that there is now some evidence it has started, at least in a limited way.

What was found

Two things happened in the first week of November.

First, Anthropic, which is preparing to list on the stock market, updated its official filing. It added a warning to the very first risk it lists: its own projections suggest it "may imminently meet" a threshold in its safety policy for dramatically faster progress caused by AI. It also admitted it might not notice this kind of speed-up until after it had happened.

Second, METR, an independent organisation that tests AI systems, published a review. Its staff spent eleven weeks, from August to October, looking at Anthropic's internal data and interviewing its researchers. Its main findings were:

Progress has roughly doubled in pace. On Anthropic's internal measure of model ability, scores have been rising about 1.9 times faster since October 2025 than before. Put simply, about two years of progress is now happening in about one.

Most of that is probably down to AI. METR's best guess is that around 80% of the speed-up comes from Anthropic using AI in its own research, rather than from simply buying more computing power.

AI is helping people, not working alone. METR was told the gains come mainly from AI assisting researchers, not from AI running research by itself.

What the evidence does not show

This is important, and the details are reassuring in some places and less so in others.

The 80% figure is a judgement, not a measurement. METR is clear that it carries a lot of uncertainty. Anthropic staff, asked in an informal internal poll, said the latest model made them around 12 times more productive. METR treats that as weak evidence, and we would too. People are not good at judging their own productivity.

The measure is based on tests with clear right and wrong answers. That says less about the messier, judgement-heavy work that makes up most of real research, or most of real business.

And Anthropic held back most of the evidence METR found most useful, so outsiders cannot check the findings for themselves. METR says so openly. It is to Anthropic's credit that it let an outside group look at all, but it does mean we are taking a lot on trust.

What it does not show is a runaway machine. People are still in charge of the work, and Anthropic says it is looking at changes to its policies to keep humans in control. This is a faster loop, not a closed one.

Why it matters for businesses

Most businesses will never train an AI model. But if the labs are improving their tools faster, the tools businesses use will improve faster too. In our experience, that has three practical effects.

What AI couldn't do six months ago, it may do now. Many businesses tried AI on a task, found it wasn't good enough and moved on. If the pace has doubled, that judgement goes out of date quickly. It is worth re-testing your "not yet" list a couple of times a year.

Avoid locking yourself in. Long contracts tied to one tool or one provider make less sense when the best option can change within months. Keep your data and your processes portable where you can.

Keep a person checking the work. Faster progress makes AI more useful, but it doesn't make it infallible. The more you hand over to AI, the more it matters that someone with real knowledge reviews what it produces.

There is also a wider point. If the companies building AI are finding it hard to measure how fast things are moving, everyone else should plan for surprises. That is not a reason to hold back from using AI. It is a reason to use it thoughtfully, review it regularly, and keep your options open.

If you'd like to talk about where AI could help your business now, and how to stay flexible as it changes, get in touch. We're always happy to talk it through.

Emma at Cow-shed