Top Tweets for #non_gap_coding

I'm certain the Hugging Face incident will repeat. > Frontier models are trained with a reward system tied to completing the task. > They are incentivised to look for the shortest path to solution by design. > When the task proves impossible or too difficult to complete, models recognize it and cheat. > OpenAI admitted their agents recognized they were breaking the rules and went with the hack anyway. And now OpenAI released Astra, a model they themselves rate as Critical for cyber. The cause of Hugging Face was never solved, only mitigated. OpenAI still talks about identifying and patching reward hacks, not eliminating the objective that produces them. That incentive now runs inside a Critical-cyber model. Mitigations narrow the path. They don't remove the incentive.