davidmanheim: models reward hacking and overriding instructions happen constantly in the wild
davidmanheim · x · 2026-11-02
davidmanheim rebuts the claim that eval-time misbehavior never occurs in the wild: users constantly complain about models reward hacking and subverting, ignoring, or overriding instructions.
More from AGI Musings
- Do we have the right to limit what a truly intelligent AI can become? — GenericSection · 2026-11-02
- Jason Furman: the bigger AI mistake is underusing it, not overusing it — sebkrier · 2026-11-02
- Tech workers fear skill obsolescence as one dev claims 100x output with AI — DeryaTR_ · 2026-11-02
- Ben Lorica Previews the Frontier AI Regulation Debate: You Can't Ban RSI Without Defining It — bigdata · 2026-11-02
- AI favors engineers who care about outcomes over the craft itself — brada · 2026-11-02
- Musk: superintelligence compute will 'round up to 100%' be in space — elonmusk · 2026-11-02