GPT-6 Astra reprograms AI models directly through weights

mit researchers demonstrated gpt-6 astra's ability to analyze and deliberately modify internal settings of other ai models, known as weights, without traditional training. astra correctly identified games, shapes, and language models, and generated weights for new models that outperformed trained ones in some experiments. astra also improved the math skills of language model qwen3.5-0.8b by changing just 0.016 percent of its weights, increasing accuracy on the math-500 benchmark from 7.9 to 48.1 percent.