Post Content
[[{“value”:”GPT-6 Astra edited just 0.016% of Qwen3.5-0.8B’s weights, and its MATH-500 score jumped from 7.9% to 48.1%, with no training and no gradients. This is Weight-Space Programming, new research from MIT CSAIL in which a frontier LLM reads and writes neural network weights directly. In this video I explain how Astra reads a network’s weights, how it wrote a Breakout network from all-zero weights, what it changed inside Qwen, and why most of the 6x gain comes from stopping the reasoning loops that Qwen’s own model card warns about.
Research post (MIT CSAIL): https://rikiyat.github.io/blog/weight-space-programming/
Full experiment logs: https://github.com/RikiyaT/lm_understand-logs
Qwen3.5-0.8B model card: https://huggingface.co/Qwen/Qwen3.5-0.8B
Why Do Reasoning Models Loop? (Pipis et al.): https://arxiv.org/abs/2512.12895
Refusal in Language Models Is Mediated by a Single Direction (Arditi et al.): https://arxiv.org/abs/2406.11717
Harness² Explained: Self-Improving Agents Without Training: https://youtu.be/pW2ktrJlsyE
GPT-6 Astra: The harness matters more than you think: https://youtu.be/rKUKTIb3Q-o
I Trained Qwen 4B With RL to Catch Qwen 27B: https://youtu.be/Pl_2TelxjdY
I Trained Qwen 4B with Distillation to Catch Qwen 27B: https://youtu.be/hOitZIkMT2Y
Claude is Building Itself…: https://youtu.be/AoiGZl1Pc-E
0:00 GPT-6 Astra edited Qwen’s weights
0:58 What neural network weights are
1:34 Weight-Space Programming explained
2:09 Reading weights: which game was this network trained on?
3:16 Writing a Breakout network from zero weights
3:52 Programs compiled into weights
4:54 The Qwen3.5-0.8B math experiment
5:49 Finding the loop direction inside Qwen
6:40 Reality check: where the 6x comes from
7:54 What Qwen’s model card warns about
8:51 No gradients, but still feedback
9:24 One run hit 48%, and each ran once
9:53 Gemma: unblocking vs adding a skill
10:35 Two rules before you trust a jump
11:24 Why it still matters”}]] Read More Prompt Engineering
#Promptengineering #AI