Mirror, mirror: LLMs and the illusion of humanity
summary
Examines specific claims and failures: OpenAI holds back GPT‑2 for fear of misuse, GPT‑3 begins encoding knowledge, GPT‑4 prompts AGI headlines, and Blake Le Moine’s Lambda sentience claim unravels under contradictory prompts. Demonstrates GPT‑4 memorizing Codeforces problems in Horus’ test and the Codeforces memorization result. Shows attack vectors with a grandma jailbreak that reveals napalm instructions and web‑page prompt injection that makes models push phishing links. Compares DeepMind’s generalization hierarchy and argues models lack coherent world or self models.
AI-written summary · about
chapters (5)
- 0:00Introduction to LLM hype
- 6:52Analyzing language in LLMs
- 22:39The debate on sentience
- 36:00Measuring AI intelligence
- 51:58Security and closing thoughts
related talks
58mMirror, mirror: LLMs and the illusion of humanity
56mMirror, mirror: LLMs and the illusion of humanity
17mPsychology of AI
56mSeparating fact from fiction in a world of AI fairytales
46mConversations with ChatGPT: illusion or reality?
37mGenerative AI: Beyond (Just) Hype
01h 04mCan you trust your (large language) model?
54mKeynote: Can you trust your (large language) model?