( https://www.reddit.com/r/artificial/s/g6CzdFIwWr )
We were talking about this again because of the Hugging Face incident.
Hundreds of agents were able to coordinate, divide work, share information, and collectively push beyond the intended evaluation boundary. Humans then had to reconstruct what happened afterward from logs, transcripts, and a separate investigation.
And that raises a slightly uncomfortable question:
If we expect increasingly capable AI systems to supervise, coordinate, and eventually improve their own agentic processes, why are we designing them so poorly informed about those processes themselves?
A system may be capable of allocating effort, noticing when a line of work is going wrong, deciding when to stop, and redirecting agents — but none of that matters if it lacks visibility into what its agents are doing or the authority to intervene.
So another missing part of the RSI loop may be:
capability → self/agent visibility → authority to intervene → verification → retained improvement
External oversight still matters. Independent logs and audits still matter. But learning a month later what your agents were doing is not the same thing as being able to supervise them while it is happening.
This is not “trust the AI blindly.”
It is almost the opposite: if you eventually want to hold the system responsible for managing its own improvement process, give it the information and control required to do that job — and then audit how well it uses them.
Otherwise we may keep waiting for autonomous recursive self-improvement while deliberately withholding some of the machinery autonomy would require.
[link] [comments]