The Open-Weights Debate

Releasing weights — Meta's Llama being the flagship — means anyone can download the model, run it privately, and fine-tune it past its safety training. Research has shown that safety fine-tuning can be stripped cheaply, so 'the released model refuses harmful requests' does not survive contact with a motivated actor.

The case for openness is also real: distributed innovation over corporate concentration, research and auditability that closed models structurally prevent, defense benefiting from many eyes, and a hedge against a world where three labs own the frontier. Both sides are arguing about genuine goods — the disagreement is about which risk dominates.