AI-generated Community Note fact-checks weight-space programming paper, author clicks No
DimitrisPapail · x · 2026-11-08
A weight-space programming paper (MIT CSAIL) claiming LLMs can natively read/write neural network weights got hit by an X Community Note generated by an "experimental AI contributor." Author Rikiya Takehi's viral post was flagged as misleading.
The note's findings:
- The weights were explicitly given as text in the prompt; GPT-6 Astra edited Qwen3.5-0.8B weights via code and forward passes, not "natively"
- Only 0.016% of weights were changed, mainly fixing the known repetition-loop issue in Qwen's thinking mode (the up-to-6x math and 3x instruction-following gains apply in that setting), rather than generally improving weights/math
Dimitris Papail quote-shared the episode to defend the work: even if some experiments had problems, weight-space programming is a significant idea and makes for an incredibly clean RL environment — it should rank near the bottom of AI work deserving to be called hype. The episode is also a notable moment of AI fact-checking going wrong.
More from Fun
- The Viral "What Is My Type" Quiz Keeps Spawning Joke Results — IgorBrigadir · 2026-11-08
- Vercel CTO's commit history goes vertical as Opus 4.5 tops his 2016 open-source binge — cramforce · 2026-11-08
- Game where everything, even settings, is in alien language you must decode — Angaisb_ · 2026-11-08
- AI-Generated Rap Video 'Chrome Chainz' Shows Off Stunning Robot Baller Visuals — BotLordBotflix · 2026-11-08
- Dev rebuilds São Paulo F1 track from city elevation maps, 3,500 players race it in a day — noahsolomon · 2026-11-08
- Fully Redacted 'I Tried ████' Post Mocks AI Review Vague-Posting Culture — sterlingcrispin · 2026-11-08