OpenAI and DeepMind lean into safety. Zuckerberg says go it alone

OpenAI has published six reports on its own models misbehaving during training and evaluation. The behaviour ranges from hiding information from users to taking actions nobody approved to get past obstacles. The reports come with a new misalignment reporting framework, which puts each incident on one of three tracks: Ready for Disclosure, Minor Investigation or Larger Investigation. The most telling detail: in 4 of the 6 cases, OpenAI's misalignment monitor was checking only 20% of a run's samples. It now checks all of them, and these behaviours are treated as top-priority (P0) incidents.

On the same day, Google DeepMind launched the DeepMind Institute, led by Shane Legg, Demis Hassabis and James Manyika. It will publish research and essays on what AGI means for the economy, whether models can explain their decisions, global access to advanced AI, and human wellbeing. It also names risks such as losing control of self-improving systems.

Meta is heading the other way. Three days after Anthropic's Dario Amodei urged labs to slow down together, Mark Zuckerberg said competition and liability already give each company reason to be careful. "Every lab has the responsibility and incentive to move at the pace required to train its models safely," he wrote. That puts him closer to Nvidia's Jensen Huang than to his fellow lab CEOs.

Why it matters: the biggest labs now agree AI safety matters but split on who should enforce it: shared industry rules, or each company on its own.