GPT-6 Astra is OpenAI’s most capable model, designed for demanding reasoning, coding, research, and computer-use tasks. In a new MIT experiment, it did something unusual: it inspected another AI’s internal “weights”* and changed them directly.* Weights are the numerical settings that decide how a model thinks and responds.
Normally, improving an AI means more training, data, and computing power. Here, Astra changed only 0.016% of Alibaba’s small Qwen model’s weights, fixing its habit of repeating itself in loops. That small edit raised Qwen’s MATH-500 score from 7.9% to 48.1%, without conventional training or gradient updates.
This matters because it suggests AI may one day repair, upgrade, or add skills to other models on request. It could also help find hidden flaws, remove unsafe behavior, or make smaller models more useful without expensive retraining. But it is still an early, narrow result: the same method did little for Google’s Gemma 3, which did not have one clear, fixable weakness.
If models can, reliably improve models AI progress could become much faster—but not automatically exponential. Real gains will still depend on better data, testing, safety checks, and human oversight. The biggest change is that “training” may no longer be the only path to making AI smarter.
WHEN AI CAN UPGRADE AI, THE SPEED LIMIT OF PROGRESS MAY DISAPPEAR.
Sanjay Sahay
Have a nice evening.