AIToday
Large Language ModelsOpen-Source AIZenn AI/MLPublished: Nov 7, 2026, 19:01 JST

GPT-6 Astra edits Qwen3.5-0.8B weights to add zundamon endings

Send on LINE
GPT-6 Astra edits Qwen3.5-0.8B weights to add zundamon endings

3 Key Points

  1. What happened

    Using Codex with GPT-6 Astra, the author autonomously edited Qwen3.5-0.8B's weights, hitting the zundamon-style ending on 77 of 100 questions and 36 of 50 new ones, changing about 0.486% of weights.

  2. Why it matters

    The author says the result came from weight editing rather than prompting or fine-tuning, and that the whole process could be left to an AI.

Not sure about something? Ask the AI

Questions and answers are published on this page.

Summaries like this, every morning.

FAQ
What exactly did GPT-6 Astra do in this experiment?
It wrote code to inspect the target model's weights and internal activations, formed hypotheses about behavior, then modified weights and checked the results, according to the author's account.
How much of the model was changed?
The author says about 0.486% of the model's weights were changed, and only one weight tensor, mlp.down_proj.weight in the 20th block, was modified.
Can I try the edited model?
Yes, the author says the model created in this experiment is available on Hugging Face.
Send on LINE

AI news that matters for your work, delivered every morning.

Pick your industry and the AI tools you use, and get news related to your work every day.

Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →

Ask AI

Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.

Questions and answers are published on this page.

Related Articles

Next articleOpenAI closes $11 billion first tranche at $1.25 trillion valuation