Anthropic Co-Founder Jack Clark Now Says Recursive Self-Improvement Could Begin as Early as 2027

Jack Clark, who co-founded Anthropic and writes the Import AI newsletter, has revised one of the most-watched dates in AI forecasting. Based on a new benchmark called TasteVal, he now believes recursive self-improvement, where AI systems design and train their own successors, could become possible as early as 2027. In a long essay in May, he put the odds of a system independently training a stronger successor at roughly 30 percent by 2027 and 60 percent by the end of 2028. The new position is a meaningful pull forward.

The benchmark behind the change comes from research lab P-Zero Research, which defines research taste as the ability to design the right experiments and interpret them correctly. TasteVal runs eight fixed tasks drawn from real AI model development, including training data selection, pretraining and fine-tuning, preference training, and defense against adversarial prompts. A fixed coding agent executes every experiment on a single Nvidia H100, with each run capped at 40 GPU-hours or 120 hours of wall-clock time.

The human baseline is unusually strong: 24 experts who recently worked at OpenAI, Google DeepMind, Nvidia and Microsoft Research, with at least two per task and the best human result counting. The top model, Anthropic Claude Opus 5.5, reached expert-level scores with a compute multiplier of 2.3, meaning it needed less than half the compute. The statistical range is wide, from 1.15 to 4.37, and a single model run cost roughly one-thirtieth of an average expert attempt.

The trend line is what moved Clark. Across model generations, the compute multiplier has roughly doubled every three months since December 2025, though the reported doubling time ranges from 1.7 to 5.0 months. Before that inflection, research taste grew far more slowly. All of these numbers come from P-Zero Research own evaluation and have not been independently verified.

The paper is candid about its limits. TasteVal measures experiment design and interpretation but not the choice of which research question to pursue, which is often considered the core of scientific intuition. The authors themselves warn the results may overstate how quickly research taste is improving, because the tasks are quick and cheap to check, exactly the kind labs find easiest to optimize, and the expert comparison group included no top-tier researchers.

Critics had drawn the line at precisely this capability. AI researcher Herbie Bradley pushed back on Clark May essay by arguing that models could plausibly take over junior-researcher work, but that research taste, creativity and coherent long-term planning were nowhere in sight. TasteVal pushes against that assessment for planning and evaluating experiments, while leaving the agenda-setting question untouched.

Real-world data still shows humans firmly in charge. A study by a Fudan University team analyzing the development of their own model found AI agents contributed up to 55 percent of methodology proposals, but humans made more than 85 percent of the final decisions. The authors noted it remains unclear how AI can generate diverse research directions and judge their value before results arrive.

The broader debate is shifting from raw benchmark scores to a different question: how much independent research work can language models actually take on, and who verifies it. If the P-Zero trend holds, AI development gets faster and cheaper, but a growing share of research decisions becomes harder for humans to audit. Clark own bias, as co-founder of the company whose model tops the benchmark, is worth keeping in view.