skinny.

Latest / Neural Newscast

MIT Uses GPT-6 Astra to Edit Qwen Weights Directly [Model Behavior]

Researchers from the MIT Computer Science & Artificial Intelligence Laboratory have recently published a study exploring "weight-space programming." The team used OpenAI’s GPT-6 Astra model to inspect and directly modify the weights of smaller models, specifically Qwen3.5-0.8B, without the need for traditional gradient-based training. By targeting specific layers responsible for repetitive reasoning, Astra was able to boost Qwen's math scores sixfold. This episode of Model Behavior examines the technical implications of direct model repair, the limitations of weight-only analysis, and the…

Log in to save or queue this episode.

The skinny

The skinny isn't ready yet — notes appear once the transcript is processed.