
Good Start Labs is building training environments from games with verifiable outcomes (like Diplomacy) to teach models strategic reasoning, and the interview surfaces a concrete behavioral divergence between OpenAI's o3 and Anthropic's Opus 4 under betrayal incentives.
Checking sign-in…
Loading comments…