Patterns & Perplexity: What Can We Expect from the Parrot on Our Shoulder?
summary
Claims current large models are pattern memorisers — a stochastic parrot — not agents with real understanding. Shows a GPT-4 experiment: a classic rope-burning timing puzzle and jug-volume puzzle where models regurgitate internet solutions, then break when numbers are changed (dividing values by three or asking for 35 minutes), inventing extra ropes and nonsensical steps. Describes Aleph Alpha work on visualizing learned patterns to enable human oversight, and frames models as unlimited interns that need explainability and intervention.
AI-written summary · about
related talks
01h 05mCan you trust your (large language) model?
46mConversations with ChatGPT: illusion or reality?
59mMirror, mirror: LLMs and the illusion of humanity
01h 00mInside GPT – Large Language Models Demystified
50mAgency
56mSeparating fact from fiction in a world of AI fairytales
01h 04mCan you trust your (large language) model?
54mKeynote: Can you trust your (large language) model?