Artificial Intelligence · Reading
How LLMs See the World
Token-level seeing is a good antidote to treating models as minds. I reread this when a PRD starts assuming the model understands the user.
- Generative AI
Why I saved this
If you cannot explain the representation, you should not promise judgment. I use this explainer to keep agentic features honest: retrieval, tools, and evaluation, not vibes.
ByteByteGo has the diagrams. I am not copying them here.