Claude rewrites eight Gemma 4 neurons to adjust Italian tense selection
A Reddit author reports using Claude Opus 5.5 to rewrite eight MLP neurons in Gemma 4 12B, creating a context-sensitive switch for Italian tense translation without gradients or conventional training. In the author's tests, the edit left 5% and 16% of sentences in the original form across two tests, compared with 4% and 2% for a system-prompt preference. However, the prompt broke 49 control cases, compared with two for the edit, while the edit left 182 of 200 unrelated translations unchanged versus 82 for the prompt. The experiment covered one model and one Italian construction; two of three runs succeeded, while the third crashed because of the author's harness.
The self-reported experiment, with model edits, code, agent logs and baselines in a repository, provides a concrete example of targeted weight editing for a narrow local-model behavior without gradients or conventional training. Its narrow scope and incomplete third run limit the result to one author's reported test.
Yesterday MIT CSAIL showed GPT-6 Astra editing the weights of a small model by hand. It changed a few neurons of Qwen3.5-0.8B, the model stopped looping on math problems, and its scores jumped. No training at all.
I wanted to try the same thing on something closer to what I actually do with local models, which is translating into Italian. Gemma 4 has a habit I don't love. For sentences like
He had been hiding in the attic for months when the soldiers found him.
it writes «Si era nascosto in soffitta per mesi…». That's correct, but I prefer «Si nascondeva in soffitta da mesi…», which says he was still hiding when they came. The catch is that when the period was already over (he had lived in Paris for ten years, but by 1990 he had moved back), the original tense has to stay. So it's not "always use tense X", it's "use it only when the meaning calls for it".
So I gave Claude Opus 5.5 the same kind of setup MIT used:
- a Python kernel over Gemma's weights;
- 50 forward passes to test its ideas on practice sentences;
- no gradients, no training.