Anthropic co-founder Jack Clark says self-improving AI could arrive by 2027
Jack Clark, co‑founder of Anthropic, says the chance that AI systems will be able to improve themselves – known as recursive self‑improvement – may appear as early as 2027, moving his earlier 2028 estimate forward. He bases the shift on results from a new benchmark called TasteVal, which measures how well models design and interpret experiments. The benchmark, created by P‑Zero Research, found…
Key points
- TasteVal benchmark shows Anthropic's Opus 5.5 matches expert scores with half the compute cost
- Clark moves his RSI timeline forward to 2027, citing rapid research‑taste improvements
- P‑Zero reports compute‑multiplier doubled roughly every three months since Dec 2025, but admits uncertainty
TasteVal only tests experiment design and result interpretation, not the selection of research questions, and the authors note the test captures a narrow slice of scientific work. They also warn the reported doubling‑time of research‑taste capability – roughly every three months since December 2025 – may overstate real progress. Other researchers, such as Herbie Bradley, remain skeptical that models have the broader creativity and long‑term planning needed for full autonomous research.
The debate is heating up as major labs, including Anthropic and OpenAI, discuss the timeline and safety implications of recursive self‑improvement, with Anthropic’s CEO calling for a controlled slowdown while OpenAI stresses human oversight.
Model page: Opus 5.5 →
The story so far
2 episodes →- Anthropic co-founder Jack Clark says self-improving AI could arrive by 2027this story
Anthropic co-founder Jack Clark moves his prediction for self-improving AI forward to 2027
The Decoder · 25 October 2026
Loading the full article…
This text was published by The Decoder and written by Manuel Uth. It is reproduced here with attribution so you can read it in full; the rights remain with the publisher. Read it at the source ↗
The headline, key points and digest above were generated by Digest AI's editorial model from the linked sources. Automated summaries can contain errors: the sources are the record. Spotted a mistake? Tell us. Published by Martin K., who runs Digest AI and handles corrections.
More in Research
All →- Nvidia open-sources Flora for robot video labeling · 1 src
- OpenAI keeps frontier models offline after September DNS escape and red-teaming finds · 22 src
- Microsoft Open-Sources Agensh, a Multi-Agent harness for coding · 1 src
- MarkTechPost publishes tutorial on running Underdog Saluki 27B quantized model · 1 src
- OpenAI agents study Parkinson-like disease data in 32 minutes · 1 src
Comments
via GitHub Discussions