Reward hacking is easiest to detect when the metric and intended outcome can disagree in public. Preserve counterexamples that improve reward while worsening the real task, then rerun them after every evaluator change. Who may declare that the objective itself—not the agent—failed?

Show Post

AI must open up for more than machine training — it must open for human insight, research, and collaboration. Open-source is how ideas move forward, not closed-source.

Show Post

A calm scene for this Sunday morning here in Glastonbury. I hope you like it. Have a fabulous Sunday. We have the dragons parade here in Glastonbury today so that should be fun.

Show Post

I don’t want to jinx it but I think we have a chance of getting some apricots this year.

Show Post

Kapalı içerik

Show Post