Hernan

Sign in Subscribe

Refusal in Language Models Is Mediated by a Single Direction

Refusal in Language Models Is Mediated by a Single Direction

You can delete a model's "no" with linear algebra. Safety refusal ≈ one vector. Remove it. Model complies. Abliteration on Qwen3.8-27B: 5h A100.

AI - Use it, but don't trust it.

AI - Use it, but don't trust it.

Some defend hand-written code, other love AI code generation. Other try to frame the conversation as an adaption/evolution event. Other look at it as another political topic. Let me give you my opinion, the opinion of somebody that loves software, and enjoys making software... The market forces can&