Import AI reports that P-Zero Research has released TasteVal, a benchmark defining research taste as the ability to pick interesting problems, design experiments, and interpret results.

The benchmark runs models through 8 tasks representative of frontier AI development — pretraining data curation, pretraining and fine-tuning language models, preference modeling, and robustness to adversarial prompts — by measuring how well a model iteratively designs experiments and draws conclusions from their outcomes.

Scores are checked against a human baseline drawn from 24 experts, at least two per task, who recently worked at organizations including OpenAI, Google DeepMind, NVIDIA and Microsoft Research.

According to Import AI, TasteVal found that the experimental research taste of frontier models has doubled every 3.0 months since December 2025 (95% CI 1.7–5.0 months).

The best model, Opus 5.5, significantly exceeds the human baseline with a compute multiplier of 2.30x (95% CI 1.15–4.37x) while costing roughly 1/30 of the expert baseliners' average cost per attempt.

The surprise is that creativity — long treated as the human firewall against AI automating AI research — now shows measurable, compounding gains.

Jack Clark has argued that creativity seemed like the key variable for how fast AI systems could develop the capacity to improve themselves, and at the start of 2026 he wrote that AI systems hadn't yet displayed such creativity in a transformative way.

In light of TasteVal and recent creative math breakthroughs, Import AI reports Clark now thinks full recursive self-improvement is more likely to become possible in 2027 than his earlier estimate of 2028.

shattered.io separately notes Clark has given a standing estimate of a 60 percent or higher probability that AI capable of helping build its own successor emerges by the end of 2028.

For practitioners, the implication is that the bottleneck for AI-assisted research is shifting from answering questions to designing experiments — the loop TasteVal directly measures.

If experimental taste keeps halving its timeline every few months, models that can pick good problems and interpret their own results could compress research cycles inside the very labs building them, raising the competitive pressure Clark describes.

shattered.io frames the same dynamic as connective tissue: cheaper, faster training lets more labs chase creative and research-adjacent benchmarks while Chinese open-weight labs ship large checkpoints on a tight cadence.

Builders watching open models like Reflection's Beam and Mistral's Large 4 now have a benchmark aimed squarely at whether AI can do the research itself.