Microsoft's iWSP: Claude writes transformer weights directly, 166 params solve algorithmic tasks
dair_ai · x · 2026-11-17
Microsoft Research and collaborators introduce interpretable weight-space programming (iWSP): an LLM agent writes transformer weights and features explicitly in code with no gradient training, iterating 30–50 rounds of propose-test-revise, producing models readable line by line.
- All 26 runs on algorithmic tasks hit 100% accuracy on held-out data, often encoding known algorithms directly
- Model capability matters: smallest solving models had a median of 166 parameters (Opus 5.5); Sonnet 5.5 needed a median of 41,600 and Haiku 5.5 100,000
- Language tasks were harder — models plateaued near 190k params and trailed learned baselines
- iWSP also builds neural encoding models across vision, language, and audition, outpredicting interpretable baselines on visual cortex
Paper: arXiv 2611.18114, code on GitHub.
More from AGI Musings
- Garrison Lovely's 'Obsolete' makes FT's best economics books of 2026 — GarrisonLovely · 2026-11-17
- Cameron: Open weight models are America's bulwark against centralized, dystopian AI — markjeffrey · 2026-11-17
- diginomica: ethics committees aren't enough — de-biasing AI needs full-process governance — jonerp · 2026-11-17
- Alison Gopnik to speak on intelligence and consciousness in AI and children at Berkeley — AlisonGopnik · 2026-11-17
- Chollet: Advanced agentic coding is essentially machine learning, with ML's classic failure modes — fchollet · 2026-11-17
- What Anthropic is really listing: Azeem Azhar on AI capturing the intangible economy — Exponential View (Azeem Azhar) · 2026-11-17