🤖 AI 资讯

A new paper (I'm not a coauthor BTW -- I just found it interesting) argues, basically, that RSI¹ is not on the horizon^2, because current (at the time the study was done) agents cannot do open-ended ML research.

Specifically, they took some accepted, but unpublished papers from NeurIPS, and tried to get the agents to do the same work, which was then graded by the original authors. And the agents (Codex/GPT-5.6 Sol and OpenClaw/Opus 4.8) could not do it.

And since they cannot do open-ended ML research, they cannot recursively self-improve -- this is their argument.³

Link: https://arxiv.org/abs/2607.27191

I think I've regretted the last 10 or so times I posted any kind of "research" in this subreddit -- either people downvote it, or it gets upvoted, but there is zero meaningful discussion. This might be the last time I'm trying this.⁴

Footnotes:

1. RSI = Recursive self-improvement, a.k.a. superintelligence explosion. The concept was invented by I.J. Good in 1965. It does not mean "anything that speeds up AI research". Compilers speed it up! RSI means, basically, a nuclear chain-reaction, but for AI. The paper talks about "explosive AI progress" in the very first sentence of the abstract, and mentions "RSI" in the text.
2. Some people have objected to my use of "X is on the horizon". I consider it synonymous to "people forecast X", and the authors use the word "forecast". "Not on the horizon" does not mean "can never happen".
3. Quote: "This design also allows us to test a mechanism that informs many forecasts of recursive self-improvement: AI agents accelerate AI research because researchers delegate entire projects to agents and judge whether the returned results advance their work. Our evaluation closely matches this model, since authors handed an agent their own research question and closely evaluated the resulting output."
4. 3 years ago, many of you upvoted a bunch of very uninformed comments that accused some researchers of misconduct, until I explained that this stemmed from misunderstanding how training works, in practice: https://www.reddit.com/r/MachineLearning/comments/18bdcu7/r_sequential_modeling_enables_scalable_learning/kc60k7e/?context=3 Today, one of the top comments is "I read the abstract (...) Nowhere, absolutely nowhere, do they make the claim ...". It's completely absurd. (Also, the commenter doesn't understand what "RSI" means.) The hivemind is very disappointing.

https://preview.redd.it/kprgucsxaoph1.png?width=796&format=png&auto=webp&s=60fdf26d150d9588e90b3d08e6a1b7fd84192ba5