Jack Clark, who helped found Anthropic and writes Import AI, has revised his call on when machines might start building better machines. The 2028 target he had settled on has moved up to 2027.
Behind the shift sits TasteVal, a benchmark from P-Zero Research. The lab describes its subject as research taste: selecting which experiment deserves a run and then reading what comes back correctly. TasteVal scores show that skill compounding at a three-month cadence beginning in December 2025.
The suite holds eight tasks drawn from model development, covering pretraining, selection of training data, preference tuning and hardening against adversarial prompts. Every task carries a fixed objective and a score. The model picks its next experiment and lets the result steer the one after.
All runs go through one fixed coding agent on one Nvidia H100. Each attempt stops at 40 GPU-hours or 120 hours of wall clock, which keeps programming ability out of the comparison and isolates experiment choice.
Twenty-four specialists, freshly arrived from OpenAI, Google DeepMind, Nvidia and Microsoft Research, supplied the yardstick, at least two on each task. The strongest human attempt set each bar. Anthropic’s Opus 5.5 topped the field: a multiplier of 2.3, human-level scores, under 50% of the compute.
P-Zero treats research taste as compute efficiency, reasoning that once coding is automated the experiment budget becomes the binding limit.
The scope is narrow, and nothing here shows systems spawning successors today. Clark’s earlier date was pulled in by evidence, not by an event.
