← All news·2026-10-26·3 min read

Anthropic: Self-Improving AI Forecast Moves Up To 2027

Anthropic co-founder Jack Clark says models' research instincts are doubling every three months, and he now expects self-improving AI in 2027 instead of 2028.

aianthropicresearch

Jack Clark is the co-founder of Anthropic, which builds the Claude models. In his newsletter Import AI, he noted that “frontier” AI models are quickly developing something called “research taste.” Research taste refers to the ability to choose good experiment hypotheses and accurately interpret their outcomes. This is often thought of as a trait possessed only by seasoned human researchers. According to Clark, research taste has doubled roughly every three months since December 2025. In a recent benchmark, models were pitted against 24 former OpenAI, Google DeepMind, Nvidia, and Microsoft Research employees. The top performing model was Anthropic’s Opus 5.5, which outperformed the humans and required just 2.3x less compute to do so.

As a result, Jack Clark adjusted his prediction of when self-improving AI (recursive self-improvement) would occur. Previously, he thought it would happen around 2028, but now thinks it will happen in 2027. This isn’t based on outside predictions of AGI, but rather on first hand knowledge of building these models at Anthropic. If frontier models continue to match or exceed the abilities of expert human researchers, many tasks currently requiring the expertise of senior engineers and researchers might be automated sooner than people think.

Source: jack-clark.net

Free course

Stop reading about AI — start building with it

The free Claude Code course: your first site, tool or game — no coding. No upsells, no cross-sells — nothing to buy here.

Start free →
EAEvgenii Arsentev

Author

Evgenii Arsentev

PhD · Founder & CEO, Dad4Kids