What changed
MIT researchers have shown that GPT-6 Astra, developed by OpenAI, can directly understand and modify the internal settings, or weights, of other AI models without traditional training. This approach, termed 'Weight-Space Programming,' allows the model to act like a mechanic adjusting individual parameters within a model's numerical structure.
What this means in practice
In experiments, GPT-6 Astra was able to identify which computer games a trained network was designed to play by analyzing its raw weights. It also recognized the basic shapes produced by an image generator, though it sometimes confused filled shapes with outlines. In another test, Astra was able to determine which of two fine-tuned language models was trained to always agree with users, though it required additional feedback on what the models paid attention to during text processing.
Astra also generated weights for models meant to play three simple computer games without any training, outperforming models trained with standard methods in some cases.
For example, in the game Breakout, Astra's model scored an average of 517 points compared to 25 for the trained baseline.
In the board game Go on a 9x9 grid, Astra's model with about 1.5 million parameters outperformed a model trained on roughly 2,000 million games.
In a practical experiment involving the language model Qwen3.5-0.8B, Astra improved its math skills by modifying just 0.016 percent of all weights, significantly increasing accuracy on the MATH-500 benchmark and instruction-following ability. However, it could not fully eliminate thinking loops, as a third of responses still hit the maximum text length.
Limitations and availability
The researchers acknowledge that the capabilities are still limited, and a benchmark for systematic measurement is next on their list. The experiment logs are available on GitHub, and the edited Qwen model is on Hugging Face. The supplied announcement does not specify whether the approach is available for general use or restricted to research purposes.
