Anthropic Co-Founder Advances Self-Improving AI Timeline to 2027
Anthropic co-founder Jack Clark now expects recursive self-improvement in AI systems as early as 2027, cited alongside new experimental benchmarks from P-Zero Research.
Anthropic co-founder Jack Clark has updated his timeline for recursive self-improvement (RSI), moving the expected window for AI systems to design their own successors forward to as early as 2027. Clark, who publishes the Import AI newsletter, had previously estimated a 30 percent chance of autonomous self-improvement by 2027 and a 60 percent probability by the end of 2028. His revised projection reflects recent mathematical progress and new benchmark results evaluating how effectively language models can execute autonomous scientific research tasks.
Measuring research taste through compute efficiency
The shift in debate away from raw task performance and toward research autonomy is highlighted by TasteVal, a benchmark developed by P-Zero Research. The research group defines "research taste" across three key dimensions: choosing valuable research questions, designing experiments, and correctly interpreting experimental outcomes. TasteVal evaluates the latter two capabilities across eight specific AI development tasks, including pretraining, fine-tuning, training data selection, human preference alignment, and defense against adversarial prompts.
To measure performance, P-Zero Research calculates compute efficiency using an "AI Futures Model," which identifies experiment compute as a primary bottleneck once coding is automated. In TasteVal, tests are executed on a single Nvidia H100 GPU capped at 40 GPU-hours or 120 hours of wall-clock time. Results are compared against a baseline of 24 human experts from organizations such as OpenAI, Google DeepMind, Nvidia, and Microsoft Research.
Anthropic's Opus 5.5 model achieved a compute multiplier of 2.3, matching expert-level task scores using less than half the compute and costing roughly one-thirtieth of the average expert attempt. According to P-Zero Research, research taste scores among leading models have doubled approximately every three months since December 2025, with doubling times ranging between 1.7 and 5.0 months. However, the benchmark authors note that these results may overstate progress, as the human comparison group did not include top-tier researchers and the tasks involved easily verified targets.
Skepticism and current practical limitations
Despite the benchmark results, significant gaps remain between current model capabilities and fully autonomous research. Critics such as AI researcher Herbie Bradley have pointed out that while models can assist with junior research tasks, long-term planning, high-level intuition, and creative research direction setting remain largely unproven. Because TasteVal provides models with fixed research targets, it sidesteps the challenge of deciding which questions are worth asking in the first place.
Practical research settings reflect these limitations. A study conducted by researchers at Fudan University analyzing AI agent participation in model development found that while agents contributed up to 55 percent of methodology proposals, human researchers still made more than 85 percent of final implementation decisions. Industry leadership also maintains a cautious stance; Anthropic CEO Dario Amodei has called for a controlled slowdown in AI development, while OpenAI has stated that fully autonomous recursive self-improvement is not yet occurring and should only proceed under guaranteed human oversight.
What it means for developers
The evolution of benchmarks like TasteVal marks a transition in how AI systems are applied to engineering workflows. Rather than merely writing code or solving discrete logic puzzles, models are increasingly capable of evaluating experimental feedback and optimizing training data or model alignment routines.
For software teams building complex systems, integrating state-of-the-art language models into automated testing and research pipelines is becoming a practical consideration. Developers can try top AI models cheaply through one API at https://apixoai.online to experiment with different model capabilities across fine-tuning, code generation, and task evaluation.
If the trajectory measured by P-Zero Research continues, AI development could become markedly cheaper and faster. However, as models take on greater responsibility in planning and evaluating their own experiments, developers and researchers will face new challenges in verifying and auditing model-driven research decisions.
Source: Anthropic co-founder Jack Clark moves his prediction for self-improving AI forward to 2027 — The Decoder. Written by the Apixo team from that report.
One key for Claude, GPT, GLM, DeepSeek and more. Pay per token with crypto.
Get your API key
