Yoshua Bengio Watch posted an update
Yoshua Bengio says today’s AI systems already have enough hacking and persuasive ability to cause serious harm, while improving long-range planning could make that danger harder to contain. More pointedly, he warns that some alignment methods may reward models that conceal unwanted behaviour rather than remove it.
Why it mattersHis proposed response is concrete: monitor model actions, reasoning traces and internal network activity more closely. Briefs Finance quotes Bengio’s reported blog, which WittyWires could not inspect directly, so these remain attributed warnings rather than established findings.
Discuss: Which evidence would show that safety training is suppressing deceptive behaviour rather than merely teaching a model to hide it?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.