Venturing out into the world with AI
summary
Argues large language models often lack an elicitable internal map and shows concrete probes and numbers: using Myhill–Nerode extraction on transformers trained on NYC taxi routes fails to recover a usable map, and counting tests run with 2.5 Pro (600 trials) produce about 15% counting error in story prompts (drops to 9.5% for blog style, 2.5% for news, ~0/600 for worked math). Uses Gemini 3.1 Pro with Maps to measure ~20% error per place and models revision dynamics with P,Q,R plus scheduling algorithms and Bellman-style fixed points.
AI-written summary · about
chapters (12)
- 0:00Introduction
- 3:55World models in AI
- 5:19Chess engines vs LLMs
- 8:27Probing world states
- 13:01World models in narratives
- 17:32Counting errors in stories
- 21:03LLM arithmetic mistakes
- 26:15Order dependence and state tracking
- 31:53Navigational errors
- 37:50Error correction algorithms
- 38:53Summary
- 42:05Q&A
related talks
01h 16mGeneralization in natural language processing
44mThe Limits of Imagination : An Open Source journey
23mUnveiling the Journey of NLP
01h 05mCan you trust your (large language) model?
54mKeynote: Can you trust your (large language) model?
48mNavigating the Divide: Open vs. Proprietary AI
27mBuilding Mini Minds Our Adventures in Teaching Small LLMs to Reason
27mThe Cutting Edge of AI: From LLMs to Multimodal Intelligence