闭源RSI的严父：18个Agent自主科研，Kimi K3靠Harness逼近Opus 5

TL;DR - Prime Intellect reports that its multi-agent research harness lets cheaper open-weight models autonomously optimize nanoGPT training, with Kimi K3 reaching 2,930 steps—close to Opus 5’s 2,920 and ahead of GPT-5.6 Sol’s 3,042. The results suggest AI-for-AI progress may depend as much on experimental throughput and infrastructure as on the underlying model.

• Across 153 autonomous runs involving 18 models, agents modified code, launched training, analyzed noisy results, and selected follow-up experiments without internet access.
• The benchmark measured how quickly agents could reduce a 124M-parameter GPT’s validation loss below 3.28, starting from a 3,290-step baseline; the human record is 2,600 steps.
• Fable 5 achieved the best agent result at 2,726 steps, capturing about 82% of the available improvement between the baseline and human record.
• Successful agents did not invent fundamentally new methods; their advantage came from repeated validation, noise handling, revisiting discarded ideas, tool creation, and higher experimentation throughput.