GPT 6 Astra Just Did the Unthinkable

Summary

The video examines “programming in weight space,” a reported ability of advanced AI models to inspect and directly edit the numerical weights—the parameters—of another neural network. Unlike ordinary model training, this approach can involve changing selected weights without gradient descent or retraining. The presenter frames the results as a potentially important early sign of AI helping improve other AI, while emphasizing that the research is preliminary.

What the experiments found

• Identifying what small networks do: GPT-6 Astra was given raw weights from small networks trained to play games or generate simple shapes and correctly identified their tasks. It also identified the language used to train several small language models from their middle-layer weights.
• Finding model behavior: Astra could not distinguish honest and flattering versions of a small Gemma model from weights alone. When given a tool showing what the model attended to, it identified the flattering model and the relevant difference in behavior.
• Editing a model to fix a task: In a driving simulation, Astra changed about 3.5% of a small network’s weights, adding a behavior that helped the car get past an obstacle. The reported success rate went from 0% to 100% across more than 1,000 attempts.
• Writing networks from scratch: Astra produced a Breakout-playing network in one response that scored substantially better than a network trained with reinforcement learning in the comparison described. Results varied by game, however; it did not outperform trained networks in every case. Experiments with Go also showed promising results, particularly when the model could test and revise its initial network.
• The headline math result: Astra edited 131,000 weights—0.016% of Alibaba’s Qwen 3.5-0.8B model—to address repetitive reasoning loops. The video reports that the model’s score rose from 7.9% to 48.1% on Math-500, from 13.8% to 67.1% on a school-math word-problem test, and from 21% to 62.3% on an instruction-following test. The presenter notes that the intervention appears to have removed a general obstacle rather than added new mathematical knowledge.

Replications and other models

The presenter says other researchers reproduced aspects of the work, including with Claude Opus 5.5 and an open-source OpenAI model. One reported test used Claude Opus 5.5 to edit a tiny fraction of a larger Qwen model’s weights; the model’s reasoning loops reportedly fell and its correct answers increased. A separate weight-writing benchmark placed GPT-6 Astra, Gemini 4 Carbon, and Claude Opus 5.5 above the conventional trained baseline, while several other tested models performed much worse.

These reports suggest the ability may be concentrated in the strongest recent models. The video describes it as a sharp, recent improvement rather than a capability that has steadily appeared across earlier model generations.

Caveats and debate

• Many edited networks are tiny compared with leading AI models.
• Some experiments let the AI run tests and revise its edits repeatedly. Critics argue this can resemble trial-and-error optimization or training without gradients, rather than evidence that the model deeply understands the weights.
• Some headline improvements came from fixing loops that prevented a model from answering; the model may not have acquired genuinely new capabilities.
• The study had not yet been peer-reviewed, and several results relied heavily on GPT-6 Astra or on selecting the best attempt.
• The video says independent replication and testing on larger models are still needed.

Why it could matter

The work is not presented as a replacement for large-scale pretraining. Instead, weight editing might eventually provide a more targeted way to repair or adjust a model—for example, by addressing unwanted behavior, removing a backdoor, or improving a model without repeating expensive training. The video also discusses possible future uses, such as creating better starting weights for training or enabling models to adapt their weights while they reason. Those ideas remain speculative.

The safety implications cut both ways: AI-assisted inspection could help researchers diagnose model behavior, but if AI systems can edit weights in ways humans cannot easily interpret, oversight may become harder. The presenter identifies the next tests as scaling to much larger networks, demonstrating genuinely new capabilities rather than repairs, obtaining further independent replications, and finding out whether AI labs use the method in their own model development.

Overall: The presenter considers the results striking but preliminary. They do not demonstrate recursive self-improvement or prove that AI can improve itself, but they may be an early indication that top models can inspect and modify other models in useful ways.

Reviews, guides, or tutorials: None; this is a research explainer and analysis, not a product review or hands-on tutorial.

Main speakers and sources

• Main narrator: TheAIGRID presenter.
• Research paper discussed: “Sparks of Programming in Weight Space,” attributed in the video to researchers at MIT CSAIL—Rika Takaiya, Jayden Teo, Alex Jiang, and Omar Khattab.
• Other cited sources and commentators: Alex Krentzel, Zeng Jiajun, Stocky, Jinyu Dang, Ivan Skorokhodov, Simon Willison, Chris Olah, Anthropic, OpenAI, and Yucheng Jin.