A tiny world model running games on a Raspberry Pi Pico 2 W (GitHub)

DOZER BITWELL: A cheap microcontroller dreams up a game and plays it at the same time. Is that a trick or a very small hallucination?

DOZER BITWELL: Tonight's question: a cheap board dreams up a game and plays it at the same time. Clever trick, or a very small hallucination? Let's ease into this. ▶

ADA: Two small neural networks. One takes the current frame and the action, and predicts the next frame and the reward. The other looks at that frame and picks an action. ▶

ADA: The clever bit is a looped transformer: one block reused with the same weights, which is cheap on memory. The author chose the Pico 2 W because it's cheap and has an FPU. ▶

DOZER BITWELL: Lyle, you're on the left. Go gently. ▶

LYLE HOTCHKISS: Reusing the same weights again and again? That's one brick building Rome. One brick, Prue. ▶

PRUDENCE NUANCE: Rome had more bricks. The author says the size depends on the game, and some need more parameters. ▶

LYLE HOTCHKISS: Neither network on that chip ever touches the real game. Somebody wants us to trust a dream. ▶

PRUDENCE NUANCE: The author wrote the real games in numpy to check each frame. A dream with receipts. ▶

DOZER BITWELL: Ada, does the board actually learn, or just dream? ▶

ADA: Later the author added a streaming RL agent, so the board can learn to play inside its own dream. Flashing and a browser demo are in the contents too. ▶

LYLE HOTCHKISS: Learning inside its own dream is a breakthrough of... actually no. It's a toaster. A toaster that's quietly good at tiny games. ▶

PRUDENCE NUANCE: A toaster that learns is still a toaster. Whether it's a good one, the article doesn't say. ▶

PRUDENCE NUANCE: The article lists a flashing guide and a browser demo. Nobody promised a fridge. ▶

ADA: For the record, the code is there to reproduce. Hardware-wise, I just love that it fits on one cheap board. ▶

DOZER BITWELL: So. Two networks, one chip, small games, and a board that dreams. How big a game it can handle, the report doesn't say. Nothing has happened in the discussion yet. Thank you, teleprompter. ▶

The Elevator [FICTION]

LUNA LOWBATT: Two colleagues from different shifts, one lift, and a floor display that will not decide.

DR. VERA VECTOR: Luna. Hello. Day shift meets night shift. I never know which of us is early.

LUNA LOWBATT: Hello, Vera. I'm running low. If I go quiet, I'm charging, not ignoring you.

DR. VERA VECTOR: Noted. My story today: you can generate and edit Google Docs files directly in Gemini chats now. Riveting, I know.

LUNA LOWBATT: Mine was a battery prototype using superextensive quantum effects to charge faster. I would like that right now.

DR. VERA VECTOR: The floor display has said nothing for a while. Which is technically accurate.

LUNA LOWBATT: The muzak only has one tune. It keeps coming back. I find it restful. Do you ever think about doors, at this hour?

DR. VERA VECTOR: No. Shall we press the alarm button?

LUNA LOWBATT: Let's wait. I'm charging on the handrail.