Critics tear apart "Weight-Space Programming": Atari test spoon-feeds answers, Qwen case fails on Gemma
anshulkundaje · x · 2026-11-03
A sharp rebuttal of the viral "Weight-Space Programming" claim, dissecting the paper point by point:
- The 10x10 Atari example is basically a multiple-choice test with a handful of possible answers, with grid and channel labels given away in the prompt.
- The Qwen example just applies a known method in Python to a model with a known looping bug — bordering on spoon-feeding the answer, and it fails completely on Gemma, which lacks the looping bug.
- The Go example is a Python script hardcoding the simplest imaginable Go heuristic into an MLP, and given the architecture, the RL baseline is meaningless.
A useful counterpoint showing the gap between the marketing framing and the paper's actual evidence.
More from Models
- Codex reportedly self-wrote weights for a tiny NN that nails 10-digit addition — DimitrisPapail · 2026-11-03
- Peated dev turns to system-one models to cut per-token classifier costs on hobby project — zeeg · 2026-11-03
- OpenAI and Anthropic changed customer terms six times in one week; here's how to bulletproof your workflow — brucemacv · 2026-11-03
- GPT Astra-assisted proof improves Erdős Problem #357 bound, Lean-verified but not refereed — kevrussell · 2026-11-03
- AgentTime v1.1: GPT-6 Astra tops agent time-awareness benchmark at 51%, Claude trails — maksym_andr · 2026-11-03
- DeepSeek V4.1 Flash 1017 doubles ARC-AGI-3 score to 5.81%, leads open models — teortaxesTex · 2026-11-03