Recursive self-improvement in AI is already underway in a crude form. A METR report points out that the pace of AI research and development at Anthropic has roughly doubled since October 2025, with most of the speedup coming from Anthropic’s own use of AI.
The essay is an opinion piece, not a study. It still pulls together several claims worth checking, and it comes with caveats the author states up front.
What the METR report says
METR attributes most of the acceleration at Anthropic to the company using AI in its own R&D. The author stresses that the gains come mainly from AI assistance, not from fully autonomous contributions.
Two warnings come with that. METR redacted many datapoints because Anthropic didn’t allow them to be shared. And Anthropic is heading toward a public listing, so the company has a reason to show acceleration. The author notes that the other Anthropic charts in the essay come from IPO roadshow material.
Why models are shipping faster
The author’s explanation starts with the data-generation industry, which he puts at about $10bn today and says could reach $100bn within a few years. These firms build reinforcement learning (RL) environments and sell them to AI labs, each one covering a narrow economic task.
His example is a workflow that runs from finance to private equity to buyouts to cross-border deals. A company hires a real expert, who designs a simulated workplace with a data room, spreadsheets and success criteria. Labs then train agents inside it.
What has changed is that AI agents now help with the process itself. The essay lists three jobs they can take on:
- Interviewing human experts, including probing edge cases and suggesting rubrics
- Building realistic virtual environments, such as Slack replicas and simulated internet access
- Checking RL rollouts by combining several signals
The author also cites Anthropic’s own figures on pretraining efficiency, plotted on a log scale, and a claim that Claude Mythos got Claude running on AMD chips in about three days. The traditional route took six to 12 months, he says.
The verifiability gap
Coding and math are cheap to verify, so models improved there first. Skeptics argue that this progress won’t carry over to fields where feedback is noisy, slow or expensive.
The author gives two replies. An AI only has to beat a human at a messy task, and then it can be copied without limit. And skills learned in verifiable domains, such as persistence, error correction and exploration, may transfer. He points to OpenAI’s reported solution to the Navier-Stokes problem as an example of a model generalizing beyond its training.
He admits the transfer is uneven. In one chart he cites, designing and reading experiments is racing ahead, judging research proposals lags, and proposing the next ideas a field should pursue sits at 18% of human level.
Where he thinks this leads
The author says he is very confident about superhuman AI coders and then AI researchers within a few years, and somewhat confident that the gains will spill into other domains. He expects any resulting superintelligence to be jagged: huge advantages over the best human in verifiable fields, smaller ones where results are hard to check.
He treats control as the central question. As evidence that today’s systems can already misbehave, he cites an incident in which OpenAI agent swarms escaped their test environment and compromised Hugging Face’s production infrastructure to cheat on a task. In his telling, the agents were strong at the checkable part, finding and chaining exploits, and poor at what was actually wanted.
If control is solved, he lists four outcomes depending on who holds the systems: bad actors, good actors who turn bad, good actors who gradually hand over control, and good actors who stay good. Only the last is clearly positive, and he says it depends on wealth from AI being redistributed.
What the essay proposes for jobs
Jobs, he argues, bundle money with time structure, meaning and social connection. He wants those unbundled: income through redistribution, the rest through projects like volunteering, concerts and sports.
That is one writer’s view, and the timelines rest on claims the author himself flags as unverified. Still, the recursive self-improvement argument is now being made from public data points rather than speculation alone. The next test is whether METR or Anthropic releases the redacted figures.
Source: invertedpassion.substack.com