Import AI #457: Prime Intellect Tests GPT-5.5 and Opus 4.7 as Autonomous Research Agents — Beat Human Baselines but Lack Creative Insight

Import AI·medium signal

Prime Intellect ran GPT-5.5 and Claude Opus 4.7 on the nanoGPT speedrun benchmark, completing approximately 10,000 runs using 14,000 H200 GPU-hours. The AI agents beat human baselines in every session at engineering optimization and hyperparameter search, but Jack Clark notes they 'lack creative insight' — they refine existing approaches without generating fundamentally novel research directions. This is the clearest empirical evidence yet of the gap between AI engineering competence and scientific creativity.