What's this about?
Scaling large reasoning models beyond human supervision, subtitled a path toward superintelligence. 22 upvotes.
What does "beyond human supervision" require?
A source of training signal that is not a human judgement. In verifiable domains that is easy — the test passes or it does not. Beyond them you need either a model judging another model, which is the J-Zero approach with its convergence risk, or some structural property of correctness that can be checked without knowing the answer. This is the central open problem in self-improvement, and papers claiming a path to superintelligence in the subtitle should be read with that difficulty in mind.
Why is it worth noticing?
- Third paper this week with self-improvement or superintelligence in its framing, alongside J-Zero and Zhipu's Full Self-Training announcement.
- The honest version of this research is important; the framing invites over-reading.
- Read next to Anthropic's reward-seeker result, which is the empirical account of what happens when the supervision signal is exploitable.