Translation Memory

AI safety beyond the frontier labs: uncensored local models

A recent Guardian report showed researchers asking frontier models to help plan violence and crime. Frontier labs (OpenAI, Google, Anthropic) responded that they’d already employed better safeguards since the research concluded. Ignoring the fact that you can still work your way around safeguards in many cases, they make no mention of local models.

I’ve experimented with highly capable local models - both from Google and from Alibaba - Gemini and Qwen - and found them to have similar safeguarding to hosted models. But then I checked the “Uncensored” versions of the same models, where researchers and enthusiasts take the base model and surgically isolate and remove the weights relating to guardrails and alignment. They report that the models answer 100% or 95+% of the unsafe questions asked of it, vs 1-5% before the further work.