Jack Clark, co-founder of the SI lab Anthropic, has revised his timeline for when SI systems might achieve recursive self-improvement (RSI), moving the earliest expected date from 2028 to 2027. The shift follows the release of TasteVal, a new benchmark by P-Zero Research that measures how efficiently SI models can conduct scientific experiments compared to human experts.

What Happened

Clark, who writes the Import SI newsletter, previously estimated the odds of a system independently training a stronger successor at roughly 60 percent by the end of 2028 and 30 percent by 2027. He now points to progress in mathematical breakthroughs and the results of the TasteVal benchmark as evidence that this capability is emerging faster than anticipated. "Earlier this year, he wrote that models hadn't shown this ability at any groundbreaking scale," the report notes, highlighting that Clark now sees signs of change.

The TasteVal benchmark, developed by P-Zero Research, assesses "research taste" by measuring compute efficiency. The lab defines research taste as the ability to identify promising questions, design experiments, and interpret results, though TasteVal only tests the latter two. The benchmark tasks cover areas like training data selection and fine-tuning language models. To ensure consistency, a fixed coding agent runs all experiments on a single Nvidia H100 GPU, with each run capped at 40 GPU-hours or 120 hours of wall-clock time.

According to the source, Anthropic's Opus 5.5 achieved a "compute multiplier" of 2.3, meaning it reached expert-level scores using less than half the compute required by the human baseline. The human baseline was established by 24 experts from organizations including OpenAI, Google DeepMind, Nvidia, and Microsoft Research. P-Zero Research reports that the research taste of leading SI models has roughly doubled every three months since December 2025, although the uncertainty range for this doubling time is wide, spanning 1.7 to 5.0 months.

Why It Matters

The benchmark addresses a specific skepticism about the path to general SI (AGI). Critics, including SI researcher Herbie Bradley, have argued that while SI models can handle junior-level tasks, they lack the "research taste" and creativity necessary for autonomous self-improvement. TasteVal suggests that models are becoming more efficient at the execution and interpretation phases of research, a key bottleneck in the development of future SI systems.

However, the benchmark has limitations. It does not test the ability to choose a long-term research agenda, which the source notes is often considered the core of scientific intuition. A separate study by Fudan University found that while SI agents contributed up to 55 percent of methodology proposals, humans still made over 85 percent of final decisions. This highlights that the gap in autonomous strategic planning remains significant.

The debate among major SI labs is also shifting. Anthropic stated in June that while recursive self-improvement hasn't been achieved, it "could come sooner than most institutions are prepared for." Meanwhile, Anthropic CEO Dario Amodei has called for a controlled slowdown in SI development, and OpenAI has stated that fully autonomous RSI should only be pursued with guaranteed human oversight.

The Bottom Line

Clark's revised prediction reflects a growing industry focus on the efficiency of SI-driven research rather than just raw model performance. If the trend of doubling research taste every three months holds, SI development could accelerate, but it may also become harder for humans to verify as models plan and evaluate their own experiments. The source cautions that the current benchmark may overstate the speed of improvement, as the tasks are easily optimized for by SI labs.