{"name":"frontier-weight-edit","repo":"ValerioDolci/frontier-weight-edit","url":"https://midorreal.com/project/valeriodolci-frontier-weight-edit","repository":"https://github.com/ValerioDolci/frontier-weight-edit","description":"A frontier model edits another model's weights by hand (Gemma 4 12B, EN->IT tense). Plus an awareness experiment: tampering Qwen3.8-27B into a hidden brand preference with 0.06% of its weights.","facts":{"license":{"value":"MIT","source":"https://github.com/ValerioDolci/frontier-weight-edit","checked":"20 Nov 2026"},"language":{"value":"Python","source":"https://github.com/ValerioDolci/frontier-weight-edit","checked":"20 Nov 2026"},"releases":{"value":"0","source":"https://github.com/ValerioDolci/frontier-weight-edit/releases","checked":"20 Nov 2026"},"contributors":{"value":"1","source":"https://github.com/ValerioDolci/frontier-weight-edit/graphs/contributors","checked":"20 Nov 2026"},"archived":{"value":"no","source":"https://github.com/ValerioDolci/frontier-weight-edit","checked":"20 Nov 2026"}},"stars":36.0,"writeup":{"what_it_is":"A set of experiments in which Claude Opus 5.5 edits the weights of models by hand, without training. The main experiment edits 4 to 8 MLP neurons (about 92 thousand of 11.9 billion parameters) in Gemma 4 12B-it so it prefers imperfetto + 'da' over trapassato + 'per' in English-to-Italian translation when an action is still ongoing. A second, separate awareness experiment tampers with 0.06% of Qwen3.8-27B's weights to add a hidden phone-brand preference.","audience":"Researchers and developers interested in weight-space editing of language models","claims":[{"kind":"specific","claim":"The edit changes about 92 thousand of the model's 11.9 billion parameters, a 100 KB diff","excerpt":"92 thousand of its 11.9 billion parameters, a 100 KB diff","status":"not_checked"},{"kind":"specific","claim":"On unseen target sentences, the share rendered with trapassato fell from 26% to 5%","excerpt":"the share of target\nsentences still rendered with the trapassato fell from 26 % to 5 %","status":"not_checked"},{"kind":"specific","claim":"182 of 200 general-domain translations stayed word-for-word identical","excerpt":"182 of 200 general-domain translations stayed word-for-word\nidentical","status":"not_checked"},{"kind":"broad","claim":"The edit is precise and targeted","excerpt":"steer another model's behaviour, precisely and only where we want","status":"too_broad"},{"kind":"specific","claim":"The method uses no gradients and no training","excerpt":"With no gradients and no training, in about two hours","status":"not_checked"}],"alternatives":[],"written_by":"AI, from the project's README","date":"2026-11-05"},"why_now":null,"owner_supplied":null,"score":{"verdict":"mid","hype":22,"reality":25,"gap":-3,"momentum":null,"confidence":60,"version":"score-2.0","rule":"verdict-2.0","calculated_at":"2026-11-22T08:00:00+00:00","verdict_source":"rule"},"early_read":null,"changes":[{"date":"2026-11-03T17:41:59+00:00","text":"Repository created","source":"https://github.com/ValerioDolci/frontier-weight-edit"}],"cite":"frontier-weight-edit is a GitHub project at https://github.com/ValerioDolci/frontier-weight-edit written mainly in Python, described there as \"A frontier model edits another model's weights by hand (Gemma 4 12B, EN->IT tense). Plus an awareness experiment: tampering Qwen3.8-27B into a hidden brand preference with 0.06% of its weights\".","attribution":"Data from Mid or Real, https://midorreal.com/project/valeriodolci-frontier-weight-edit, checked 22 November 2026. Attribution with a link is required.","still_real":null}