Claude writes neural net weights from scratch: 959 programs, 100% accuracy on 10 algorithm tasks
ChengleiSi · x · 2026-11-03
- Inspired by Rikiya Takehi's "Weight-Space Programming," researchers flipped mechanistic interpretability around: instead of training a network then analyzing its weights, they had Claude, running as an autonomous coding agent, hand-set every weight of a blank transformer with no gradient descent at all.
- Setup: 13 tasks with known solutions, producing 959 weight-space programs in total.
- Results: all 10 algorithmic tasks — addition, multiplication, sorting, parity, GCD — hit 100% accuracy within three attempts, and Claude kept simplifying its solutions afterward.
- Because each program is written rather than learned, the role of every head and neuron comes with a built-in account, making these models a source of ground truth for interpretability research. The post includes a full weight visualization of a 174-parameter parity program.
- The reposter notes their own similar experiments suggest the original post may actually be underselling the approach — Claude can write entire, functional neural net weights from scratch.
More from coding & agent
- Benchmarking System 1 decision models: 40 yes/no questions in one pass, 159ms agent routing — ThatOneSnare · 2026-11-03
- Browser agents rarely fail at reasoning — modals, focus rendering and silent save failures break them — yantra-lucy02 · 2026-11-03
- Zed CEO shares weekly agent changelog automation summarizing 250 commits — zeeg · 2026-11-03
- 'Forget PRs, push direct to main': the emerging human+agent engineering workflow debate — hugobowne · 2026-11-03
- OpenAI Agents API ships automatic review, bringing Codex-style "approve for me" to agents — craigsdennis · 2026-11-03
- FastAPI Cloud goes GA after beta with thousands of apps deployed — honnibal · 2026-11-03