AI is helping build better AI

In August I ended a post about AI improving itself by saying we'd almost certainly know by 2040 whether it could do the hard parts of AI research. I may have given us too long.

METR, a non-profit that tests AI systems, has just spent 11 weeks inside Anthropic working out whether Claude is speeding up the building of the next Claude. Its answer is yes, probably, by a lot. On well-specified tasks, Anthropic's models have been improving at about twice the pace they managed before October 2025, and METR puts perhaps 80% of the current speed-up down to AI.

Anthropic sells a product that's now doing part of the job of improving the product.

Ryan Greenblatt, who I quoted in August expecting labs to automate AI research around 2030 or 2031, read the report and called it "a big update towards ASI within 1-2 years". ASI is artificial superintelligence. He also wrote "I was seriously wrong", which isn't a sentence you see much on LinkedIn.

How much faster are the researchers?

If the whole lab is going twice as fast, the researchers are going a lot faster than that. Better ideas are only one of the things that make a model improve, and every idea still needs computing power to test it. Ajeya Cotra, who works at METR but wasn't involved in this report, works backwards from the doubling and lands on roughly 10x.

She's careful to say that doesn't mean every researcher has nine assistants, as anyone who's managed a team of ten will understand. She says it's as if each researcher is in a "time warp bubble", getting ten days for every day that passes outside, which she doesn't think any amount of hiring could buy.

Monday now has a fortnight in it. Presumably some of that's still meetings.

I'd treat those numbers with care. The 12x comes from staff voting with emoji reactions, which METR calls "only very loose evidence". And when it looked at seven people claiming 10x or more in its May survey, it was confident two were overstating it and suspicious of two more.

How far could the loop go?

Imagine hiring someone who takes a year to train a successor who's twice as good. The successor trains the next one in six months. That one takes three.

Cotra's maths works the same way. AI took about a year to get Anthropic to double speed, so at that speed the next doubling takes six months, the one after takes three, then six weeks.

Keep halving and the whole staircase fits inside a year, however many steps you add. Cotra says this "naively implies" extreme superintelligence within a year. She expects a ceiling somewhere, though not one close to where we are now. METR goes no further than "a significant risk" of even faster acceleration over the coming year. Cotra thinks systems that capable could overpower us if they don't share our goals, and that the industry is nowhere near ready.

What could slow it down

From the pavement, a car that's gone up a gear and a car that's still accelerating look much the same for the first few seconds.

That's the argument about METR's data. One commenter, Thane Ruthenis, reads it as "a one-off slope change somewhere around the start of 2026", meaning a faster pace that then holds. METR thinks the faster pace is lasting, and says it "could be undergoing continuous acceleration" on top of that. Cotra's staircase needs that second part.

The loop may also be weaker than it looks. Ramez Naam argued in September that it would need to be "roughly 5–10× stronger to sustain itself", because each extra bit of progress costs more than the last. And METR says much of its evidence "measures performance on well-specified tasks", the kind with a score at the end. AI is "not yet able to make highly significant and fully autonomous contributions" to the most important parts of training. That's the question I ended on in August, and I still can't answer it.

We can't check much of this ourselves.

METR wasn't paid, thinks it "extremely unlikely" Anthropic was trying to mislead it, and admits the public now has to rely on its judgment "much more than would be ideal". It also says Claude "was involved in many stages of this project", so AI helped with the report saying AI is helping.

Meanwhile Anthropic's amended IPO filing says it "may imminently meet" its own threshold for dramatically accelerated progress, among the warnings to investors about what could go wrong. It's an unusual risk factor, in that some investors will read it as the pitch.

Where the work piles up

METR didn't see bottlenecks that would stop progress over the next 12 months, though its understanding "changed several times". Anthropic's engineers did run into one. The company says the new constraint is "all of the developer infrastructure built for human scale".

If your chefs get ten times faster, you find out very quickly how many waiters you've got.

A team of seven replaced Anthropic's source control and moved everyone onto it in five months, with Claude writing and tuning parts of it, so the waiters are getting faster too. Anthropic's own summary is that "every bottleneck we remove immediately leads to a further increase in the amount of code we produce".

Not everyone has that many waiters. Postgres 19 slipped to around 19 November and lost eight to ten features after AI reviewers found more problems than its developers had time to fix. Peter Eisentraut, one of its committers, described "staring at a list of forty defect reports". Each might only need three lines, "and then you run out of time".

I'd expect most businesses to meet that problem well before they meet superintelligence.

In August I said 2040. I'd like to change my answer to: ask me again in six months. Then three.