Artificial Intelligence
Jack Clark moves self-improving AI forecast to 2027
A P-Zero Research benchmark on experiment design and interpretation led Jack Clark to move his estimate for recursive self-improvement in AI to 2027, though it covered only part of the task.
- P-Zero Research published TasteVal, a benchmark that tested how AI systems chose experiments and interpreted the results.
- The top result came from Anthropic's Opus 5.5, which the article said reached a compute multiplier of 2.3, with a reported range from 1.15 to 4.37.
- Jack Clark said the capability had been roughly doubling every three months since December 2025 and now thought recursive self-improvement could be possible in 2027.
- The article said TasteVal covered only part of scientific work and did not show systems improving themselves end to end.
- Herbie Bradley disputed Clark's claim, and a Fudan University study found AI agents contributed up to 55 percent of methodology proposals while humans made more than 85 percent of final decisions.
P-Zero Research introduced TasteVal to test how AI systems selected experiments and read the results. The benchmark used eight fixed tasks and compared models with 24 human experts who had recently worked at places including OpenAI, Google DeepMind, Nvidia and Microsoft Research.
The study turned that performance into a compute-efficiency measure. Anthropic's Opus 5.5 recorded the highest score, with the article saying it needed less than half the compute of the expert benchmark and about one-thirtieth of the cost of an average expert run.
The article said the compute multiplier had risen quickly since December 2025, although it also noted a wide uncertainty range and warnings from the authors that the benchmark might overstate the pace of improvement. The same passage said the test was limited to tasks that labs can optimise for easily.
On that basis, Clark said recursive self-improvement could be possible as soon as 2027, rather than 2028 as he had previously expected. The article noted that his forecast came from an Anthropic co-founder, which gave it an obvious conflict of interest.
The piece then set the benchmark against criticism and wider lab debate. Herbie Bradley said models still lacked research taste, creativity and long-range planning, while Fudan University researchers reported that AI agents were still not making most final decisions in their own model work. Anthropic and OpenAI were also cited as saying RSI had not yet been achieved and needed human oversight.
Named in this story
People
- Jack Clark
- Anthropic co-founder who said the benchmark supported a 2027 timeline
- Herbie Bradley
- pushed back on Clark's view of the missing capabilities
- Dario Amodei
- Anthropic CEO who called for a controlled slowdown in AI development
Companies
- Anthropic
- Clark's company, whose model topped the benchmark
- OpenAI
- one of the labs where baseline experts had recently worked
- Nvidia
- one of the firms where baseline experts had recently worked
Organisations
- P-Zero Research
- published TasteVal and defined research taste
- Google DeepMind
- one of the labs where baseline experts had recently worked
- Microsoft Research
- one of the labs where baseline experts had recently worked
- Fudan University
- led a study on AI development using AI agents
Products and systems
- TasteVal
- benchmark measuring experiment design and result interpretation
- Opus 5.5
- Anthropic model that scored best in the test
- Nvidia H100 GPU
- single GPU used to run the experiments
- AI Futures Model
- P-Zero's model for when experiment compute becomes the bottleneck
How the source tells it
The article was mostly analytical, but it used a narrow benchmark and a shorter timeline to create a mildly urgent, forward-leaning register.
- speculation presented as fact a benchmark trend is used to advance a near-term forecast for recursive self-improvement
- urgency the timeline is framed as arriving before institutions are prepared