18 AI Models Ran Autonomous Research for 8 Days; Best Closed 82% of Gap to Human Record

新智元 · wechat · 2026-08-17

PrimeIntellect published the largest autonomous-research experiment yet: 18 frontier models ran fully autonomously for 8 days in offline sandboxes (one dedicated 8xH200 node each, bwrap + network namespaces), tackling the nanoGPT optimizer speedrun — no architecture or data changes, just reach validation loss 3.28 in as few training steps as possible. The human record, set by dozens of engineers over months, stands at 2,600 steps; models started at 3,290. Records had to reproduce across 8 fixed seeds (1-in-1,000 odds of luck), and an independent LLM monitor auditing 100+ reports found no cheating. Fable5 hit 2,726 steps, closing 82% of the gap; Opus5 closed 54%; GPT-5.5 managed just 8%. A unified-budget comparison shows the ranking holds whether measured by hours, experiment count, or output tokens — the edge came from method, not brute force.

The most counterintuitive finding: no model invented a fundamentally new method. Weak models killed whole method families off single-seed negatives and mistook their own code crashes for bad ideas; strong models gated spending with 3-seed probes before committing 8 seeds, and re-ablated the entire stack after every merge — Opus5 set a record by reviving a discarded β2 setting, and Fable5 saved 31 steps in one retest of combinations that were individually worse but jointly better. Models also built CPU-side mock labs to probe dynamics before touching GPUs, and GPT-5.6 Sol spun up a three-agent research team that crashed on shared-directory conflicts before fixing itself with read-only contracts.

The piece ties the results to OpenAI chief scientist Jakub Pachocki's roadmap (an autonomous research intern this September, a fully automated multi-agent research system by 2028) and Anthropic co-founder Jack Clark's call that recursive self-improvement is >60% likely by end of 2028: AI has absorbed 82% of the 99% perspiration, but the 1% inspiration — genuinely novel algorithms — remains a human moat.

Related event: Prime Intellect Releases Results of Largest Autonomous AI Research Experiment(16 posts)→

Original post →

More from AGI Musings

AGI Musings channel →