Mirror, mirror: LLMs and the illusion of humanity
summary
Examines specific claims that LLMs are sentient or AGI by showing concrete tests and failures: vector math from word2vec analogies, the octopus grounding thought experiment, self‑reports that can be flipped by prompts, a case where GPT‑4 reproduces training‑time Codeforces puzzles but misses 10 post‑training puzzles, examples of hallucination (Black Mirror dates) and the grandma jailbreak and prompt‑injection exploits that bypass handcrafted guardrails, concluding models are powerful pattern predictors but lack coherent world and self models.
AI-written summary · about
chapters (15)
- 0:00Welcome
- 1:08LLMs and human-like illusion
- 3:57Claims of AI sentience
- 6:47Meaning in language models
- 11:39How LLMs learn language
- 16:33Limits of semantic understanding
- 19:01Sentience in artificial systems
- 25:42Evidence for model sentience
- 30:20LLMs lack coherence
- 33:09Assessing claims of artificial general intelligence
- 39:54Skill-based intelligence tests
- 44:08Measuring generalizability
- 49:24Risks of overestimating LLMs
- 52:11Prompt injection attacks
- 55:23Conclusion
related talks
58mMirror, mirror: LLMs and the illusion of humanity
59mMirror, mirror: LLMs and the illusion of humanity
17mPsychology of AI
23mGhost in the Machine: AI, Spirituality and a New Humanism
37mAI Hallucination Explained 🤯 Why AI Gets Things Wrong
32mWhy do LLMs hallucinate?
01h 18mThe Enigma of LLMs: on Creativity, Compositionality, Pluralism, and Paradoxes
33mAI is coming for your job