New paper argues RSI isn't near: agents fail to reproduce unpublished NeurIPS work
we_are_mammals · reddit · 2026-09-15
A new paper (arXiv:2607.27191) argues that recursive self-improvement (RSI) is not on the horizon, because current agents cannot do open-ended ML research.
- Method: take NeurIPS papers that were accepted but not yet published, have agents attempt the same research, and have the original authors grade the results.
- Agents tested included Codex/GPT-5.6 Sol and OpenClaw/Opus 4.8; none could do the work.
- The argument: no open-ended ML research capability means no recursive self-improvement.
The poster (not a coauthor) also laments that research posts in the subreddit rarely spark meaningful discussion.
More from AGI Musings
- Kevin Roose's Machine Gods pod: why scientists are mad at OpenAI solving their hardest problems — kevinroose · 2026-10-30
- Cryptographer Matthew Green: Anthropic's agent worm framing is unhelpful, we need a simulated one to know — matthew_d_green · 2026-10-30
- Ethan Mollick on AI and work: 'You're a hand-weaver in a world of power looms' — emollick · 2026-10-30
- Vercel founder Rauch: trust is the deepest moat for a post-AGI business — soleio · 2026-10-30
- Scattered Across AI Assistants: A New Kind of Cognitive Failure Nobody Has a Prior For — manosaie · 2026-10-30
- a16z partners: 10-20M people drive $150-200B in token spend, and AI workspaces will eat the browser — petergyang · 2026-10-30