Risks associated with AI technology, including prompt injection and the potential for AI to cause extinction, language models, with a new optimizer called MeZO proposed for fine-tuning large models and a paper investigating whether language models can identify their own "hallucinations." Additionally, a bias in the evaluation paradigm of using large language models to score the quality of responses generated by other models is uncovered, and two calibration strategies are proposed to address this bias.
Contact:  [email protected]
Timestamps:
00:34 Introduction
01:33 ChatGPT Plugins aren’t safe, Prompt Injections
02:48 Statement on AI Risk
04:40 AI is Eating The World
06:04 Fake sponsor
08:38 Fine-Tuning Language Models with Just Forward Passes
10:10 Do Language Models Know When They're Hallucinating References?
11:41 Large Language Models are not Fair Evaluators
13:50 Outro

Ver todos los episodios