“Gyarados has pretty teeth, dear”: Theatre, Pokémon, and software docs
summary
Claims that matter: large language models operate on tokens not words and produce text by sampling next-token probabilities, so correctness is accidental (demonstrated by an invented river‑crossing puzzle the model fails). Concrete numbers and artifacts: Pokémon Red speedrun records 1:43:40 any% glitchless and 1:17.577 any%, the Pokédex UI as an in‑game database, and effort values = hidden stats capped at 255 because of byte limits. Docs should publish schemas, data sheets, and model cards and include reproducible experiments and parsed algorithms.
AI-written summary · about
chapters (23)
- 0:00Introduction
- 0:43Amazon listings to social change
- 2:29Tech promises vs reality
- 3:56How ChatGPT works with tokens
- 5:21ChatGPT syntax without meaning
- 7:11Computers as theater and improv
- 9:39Ticking time bomb illusion
- 11:46Big data myth and AI
- 12:49Honest design and tech docs
- 13:32Brecht’s epic theater effects
- 15:20Literarization of software docs
- 16:43Reading guides and docs as genres
- 18:07Video games as theater
- 19:54Pokémon overview and mechanics
- 21:59Speedrunning Pokémon as code
- 24:05EVs and competitive play
- 25:31Docs against UX magic
- 26:55Data sheets and model cards
- 27:39Engaging technical docs
- 28:01Computers as bit machines
- 28:23Q&A
- 29:51Q&A
- 31:59Q&A
related talks
01h 05mCan you trust your (large language) model?
56mSeparating fact from fiction in a world of AI fairytales
34mTerrible tokenizer troubles in large language models
54mKeynote: Can you trust your (large language) model?
01h 00mInside GPT – Large Language Models Demystified
54mKeynote: Can you trust your (large language) model?
46mConversations with ChatGPT: illusion or reality?
21mBuilding at the Speed of Thought