Claude hand-edits 8 neurons of Gemma 12B to fix one Italian tense, beating LoRA and prompts
FeydRowan · reddit · 2026-11-04
Replicating MIT CSAIL's "weight-space programming" demo, the author had Claude Opus 5.5 edit Gemma 4 12B's weights directly — no gradients, no training — via a Python kernel with 50 forward passes, to change how it renders one Italian tense (preferring the still-ongoing «nascondeva da» form only when semantically appropriate).
Key findings:
- A linear probe showed Gemma already separates "ongoing vs finished" almost perfectly in mid layers — it just never uses it for verb choice
- Claude rewrote 8 MLP neurons into a context-gated switch: 92k of 11.9B params, 100KB diff
- On unseen sentences: old form drops 26%/32% → 5%/16%, with only 2/50 control sentences broken and 182/200 unrelated translations unchanged. System-prompt preference broke 49 controls; mean-difference ablation wrecked the model; LoRA broke 120 unrelated outputs
- 2 of 3 runs succeeded, independently converging on the same switch-neuron design; repo includes code, agent logs and all baselines
A compelling demonstration that targeted weight edits can outperform prompting, ablation and LoRA on precision.
More from coding & agent
- ThunderAgent merged into LLM-D: 2x agent serving throughput, TTFT from ~50s to 2-4s — simran_s_arora · 2026-11-04
- Agents pass the test but fail the job: the requirement gap and the model gap — bigdata · 2026-11-04
- Does that new prompting trick actually work, or just look better? — jimmykoppel · 2026-11-04
- Vercel launches Security Review in public beta, free until Nov 10 — cramforce · 2026-11-04
- QuickProof: formal verification for C with specs written as plain C tests — thomasahle · 2026-11-04
- Project Moah: Open-Source Memory Linker That Gives AI Agents a Brain — starcount_games · 2026-11-04