NanoDreamer: A 750K-parameter world-model agent trained on a MacBook M5 Pro plays 5 MinAtar games
CatAstro_Piyush · x · 2026-10-29
Independent developer Benjamin Bai built NanoDreamer, a tiny Dreamer-inspired 750K-parameter world-model agent that learns to play all 5 MinAtar games — trained entirely on his MacBook M5 Pro with no GPU.
Highlights:
- From scratch: He wrote his own autograd engine, tensor library, neural network layers, optimizers, and transformer to understand every component.
- Learning through imagination: The actor–critic is trained entirely in latent imagination, with no expert demonstrations or game-specific strategies; each game uses the same architecture with separate weights.
- Cheap to train: 500K steps per game, roughly 3 hours on an M5 Pro.
- The write-up covers the bugs, experiments, and architectural tweaks needed to get Dreamer working at a fraction of the usual cost, plus a separate ML foundations guide.
More from Embodied
- Contract PM says 3 months in humanoid robotics beat years of full-time output — yongqianme · 2026-10-29
- ETH Zurich builds electrocatalytic implant that clears blood sugar, works six weeks in diabetic mice — CurieuxExplorer · 2026-10-29
- Running Liquid AI's LFM3-VL-2B on a robot arm camera as the first test — AIFlow_ML · 2026-10-29
- Dev uses GitHub Copilot to build a tilt-controlled game on the Universe hardware badge — lee_stott · 2026-10-29
- SiC waveguide costs could fall to ~$100 per unit, but an optics skeptic pushes back — chrisgrayson · 2026-10-29
- Dyna-2 trained on 1M+ hours of human video; data partner ships 450K hours monthly — Scobleizer · 2026-10-29