⚡Megatron LLM Hub
←Back to search
Twitter

质疑:开源模型追平后,恶意微调可在半年内造出攻击型模型。

给出了开源权重能力追平的时间窗估计,影响本地模型选型判断。
llm微调议题
so you can just take an open model as capable as openai's new model, which is solving math problems and finding new 0-days, and turn it into a maniac in probably 6 months when open models reach there? how easy is it for the bigger models?