AI Just Reversed Engineered AI, What Does This Mean?
A team at MIT had GPT-6 Astra edit the weights of Qwen3.5-0.8B directly, with no training and no gradients. It changed 0.016% of the weights, and the authors report MATH-500 going from 7.9% to 48.1%. In this video I break down what weight editing actually is, why most of that jump comes from stopping the model's thinking loops rather than teaching it maths, and what the prompt and the logs show. I also cover the history behind it, from ROME and the refusal direction to last week's Qwenputer, what researchers like Neel Nanda, Ethan Mollick and Lucas Beyer said, and what this could mean for small models, safety and the open-weight supply chain. All results are the authors' own and haven't been independently replicated yet. COURSE & COMMUNITY https://switchdim.s.gy/yt-ai-edits-ai-weights SOCIALS X: https://x.com/robshocks LinkedIn: https://www.linkedin.com/in/robertoshaughnessy/ Instagram: https://www.instagram.com/therobshocks/ REFERENCES The research Sparks of Weight-Space Programming (blog): https://rikiyat.github.io/blog/weight-space-programming/ Announcement post (Rikiya Takehi): https://x.com/rikiyatakehi/status/2117313742891327994 Full experiment logs: https://github.com/RikiyaT/lm_understand-logs Prior work ROME: https://arxiv.org/abs/2202.05262 MEMIT: https://arxiv.org/abs/2210.07229 Language models can explain neurons in language models (OpenAI): https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html Scaling Monosemanticity (Anthropic): https://transformer-circuits.pub/2024/scaling-monosemanticity/ Refusal in Language Models Is Mediated by a Single Direction: https://arxiv.org/abs/2406.11717 Qwenputer (Divyansh Agrawal): https://divagr.com/blog/qwenputer Reactions Hacker News discussion: https://news.ycombinator.com/item?id=50341174 Neel Nanda: https://x.com/NeelNanda5/status/2117375052618539485 Lucas Beyer: https://x.com/giffmana/status/2117346161866514524 Ethan Mollick: https://x.com/emollick/status/2117338841749029221 WORK & COLLABS info@switchdimension.com CHAPTERS 0:00 An AI edited another AI's weights 0:46 What editing weights actually means 2:33 Why the 6x is mostly a loop fix 3:54 Did the prompt give it away? 4:33 Didn't we do this already? 5:22 What's actually new 5:46 Hype check: what Hacker News asked 6:20 What it could be useful for 6:44 The security risk in open weights 7:25 Not an AI rewriting its own brain (yet) #GPT6 #Astra #WeightSpaceProgramming #Qwen #MechanisticInterpretability #AIResearch #LLM #OpenWeights #AISafety #OpenAI #MIT #AIDevelopment #MachineLearning #SwitchDimension
