Four MIT CSAIL researchers published a blog post on Monday that collected more than 2,100 points on Hacker News within a day. Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab handed OpenAI’s GPT-6 Astra the weights of a small open model and asked it to edit them directly. No fine-tuning, no gradient descent. They call it weight-space programming. I wrote about Astra back in October.
What the numbers say
The target was Qwen3.5-0.8B. After the edit:
| Measure | Original | Edited |
|---|---|---|
| MATH-500 | 7.9% | 48.1% |
| GSM8K | 13.8% | 67.1% |
| IFEval instruction following | 21.0% | 62.3% |
| Responses hitting the token limit | 93.0% | 33.1% |
That last row is the key. Qwen used to repeat text until it exhausted the permitted output length, which is why it never finished most math problems.
The patch sits in two layers
According to the authors, Astra removed a numerical direction associated with repetition from selected components in layers 15 and 19. That touched 131,072 weights, roughly 0.016% of the model.
In some experiments Astra received the weights as text. In others it got a sandboxed Python environment for inspecting larger networks, plus practice problems and a limited budget of model calls.
Getting there was expensive
The strongest attempt ran across 50 evaluation requests, 158 conversational turns, 80 Python runs and 20.3 hours, including 4.7 hours waiting out a service outage. The patch is tiny; the search for it was not. So the paper does not show that direct editing is cheaper or faster than training. It shows that it works at all.
Takehi writes that models may come to treat raw weights like readable, editable code: diagnosing failures and patching them directly. In the reactions on X, Yuchen Jin (Databricks) and Chandan Singh (Voyage AI) welcome exactly that prospect, while Maxime Fournes of Pause AI warns that reliable weight surgery would mostly accelerate the acceleration.
The gain is a repair, not a new skill
A sixfold jump in a math score sounds like a model that suddenly learned arithmetic. It is not. Qwen could already partly solve these problems; it just never reached the end because it got stuck in loops. Astra removed an obstruction rather than adding an ability.
That makes the work more interesting, not less. A model that can locate the spot in somebody else’s network where a known malfunction lives, then fix it with 131,072 numbers, is a debugger for neural networks. I would like a version of that to point at a model doing odd things in production. Until then it is a lab result with a very good log attached.