⚡Megatron LLM Hub
←返回搜索
Twitter

Unprompted 会议要点:模型内部结构成新攻击面,运行时覆盖可防不可信模型供应链投毒。

本地跑第三方微调模型的实际风险与缓解手段,对应自托管场景。
llm供应链ai_agent
RT @msuiche: Today, was another great day at @unpromptedconf ! 1. Model Internals are the new Windows Internals. 2. Debugging used to be about basic blocks, it is now about transformer blocks. (We love runtime!) 3. AI is a race to control what people can think about. But the hacker mindset can't allow this. "My crime is that of curiosity." 4. Run time overlays like GLP prevent supply chain issues created by abliterated models from untrusted parties (potentially backdoored too). 5. Local models are useful now!