The Lab Landscape & Voluntary Commitments

Frontier AI is built by a handful of labs with genuinely different safety postures. Anthropic was founded explicitly around safety and publishes the Responsible Scaling Policy. DeepMind maintains a frontier safety framework with defined capability thresholds. OpenAI publishes a preparedness framework — though its dedicated 'superalignment' team was dissolved in 2024, a reminder that safety commitments compete with product pressure inside labs too. Meta has bet on open weights, which changes the safety question entirely (next lesson). Newer entrants — xAI, Mistral, and others — keep reshaping the map.

The point isn't who's 'good' — it's that safety policy is set per-lab right now, and the labs disagree.