Follow-up: model gap in MNIST weight programming, more data doesn't help
RobertTLange · x · 2026-11-05
RobertTLange follows up on his MNIST weight-programming experiment: the setup differs from the blog headline — agents program simple networks, not another LLM, and start from random init, so the popular activation-steering argument doesn't fully apply. The attached chart shows Opus 5.5 at 88.6%, Sonnet 5.5 at 85.5%, GPT-6 Astra at 84.9%, and Haiku 5.5 at just 50.3%: many models push the MLP near 90%, frontier models do better, but Sonnet 5.5 is competitive.
Related event: AI Agents Directly Edit Neural Network Weights in MNIST Experiment(2 posts)→
More from Models
- Ripple-65B-Preview, a mystery 65B model, trends on Hugging Face — ricursiveAI · 2026-11-05
- Is Grok intentionally botching translations for engagement? One viral mistranslation raises doubts — kingslyj · 2026-11-05
- 11,398 real-world evals: GPT-6 Astra stands out in 5-VLA policy comparison — DJiafei · 2026-11-05
- Researcher disqualifies Olmo 3 evals for cherry-picking comparisons, dodging Qwen baselines — JJitsev · 2026-11-05
- Sentry founder: $25/day quota can't even complete a reasonable session — zeeg · 2026-11-05
- OpenMined launches ScreamingFace: open-source model fusions beat best single models on live leaderboards — iamtrask · 2026-11-05