Anthropic: Self-Improving AI Forecast Moves Up To 2027

Jack Clark is a co-founder of Anthropic, the company behind the Claude models. In his newsletter Import AI, he wrote that the best AI models are rapidly gaining what he calls "research taste." Research taste is the ability to pick promising experiments and read their results correctly. People usually see this as a skill that belongs to experienced human researchers. Clark says this skill has been doubling every three months since December 2025. In one benchmark, models were compared against 24 experts who had previously worked at OpenAI, Google DeepMind, Nvidia, and Microsoft Research. Anthropic's Opus 5.5 scored best in the test. It beat the human experts' results while using 2.3 times less compute than they needed.

Because of this, Clark moved up his own forecast for when AI could start improving itself without human help. This process is called recursive self-improvement. He used to expect it around 2028. Now he expects it in 2027. This forecast comes from someone who actually helps build these models at Anthropic, not from outside speculation about AGI. If frontier models keep closing the gap with expert researchers at this pace, research and engineering work that today requires experienced specialists may start shifting to automated systems sooner than most forecasts assumed.