Jagged intelligence
Running the model & the world
The idea
Section titled “The idea”Capability is jagged. A model that handles olympiad-grade problems can fumble comparing 9.11 with 9.9, or miscount the letters in a word - because the world arrives as opaque subword tokens and behaviour comes from trained distributions, not from a general competence dial (Karpathy at ~2:01-2:07; an interpretability report he cites suggests the 9.11 slip may ride an unrelated verse-numbering feature - flagged secondhand).
The same lens explains self-description: ask a model what it is and it samples the most plausible answer from its training data, which is why unbranded models confidently claim to be ChatGPT (~1:41-1:46). Its account of itself is generation, not introspection.
The story, animated
Section titled “The story, animated”One dial fractures into an uneven bar chart - then watch the model sample a plausible name for itself.
Where it sits
Section titled “Where it sits”Worth having first: Tokens · Hallucination & tool use
Source-checked against Karpathy, “Deep Dive into LLMs like ChatGPT” (2025), ch. 1:41-1:46 and 2:01-2:07.