[GPT-6 Begins "Hand-Writing"" Neural Networks: Turning Game Rules into Weights, Achieving 20x Higher Scores in Breakout] Beating AI News: Neural networks often contain millions or even hundreds of millions of parameters, which have traditionally been adjusted through repeated training algorithms. However, an MIT research team has discovered that GPT-6 Astra can directly analyze and modify these numbers, and even write a neural network from scratch without training. Researchers provided GPT-6 with the rules, source code, and a neural network with zero weights for a simplified Breakout game, allowing it to automatically fill in the parameters. No training data was provided, and no play was allowed. GPT-6 calculated the ball's landing point and incorporated the logic for prediction and paddle movement into the network weights. The final game score reached 517.4, approximately 20 times that of a model with the same structure trained for 500 million steps of reinforcement learning. However, in two other games, it did not surpass the model trained for 500 million steps. Researchers then had GPT-6 directly modify the weights of Ali's Qwen3.5-0.8B model. Without retraining the model, by changing only 0.016% of the parameters, it increased MATH-500's math accuracy from 7.9% to 48.1%, and instruction compliance improved nearly threefold. It turned out that Qwen frequently got stuck in a thought loop, repeatedly reasoning without providing an answer. Before the modification, 93% of math test responses would exhaust 32,768 tokens; after the modification, this dropped to 33.1%. GPT-6 essentially fixed the problem hindering the model's ability to answer questions correctly. However, when switching to the Gemma 3 270M with only 270 million parameters, GPT-6 failed to significantly improve math performance. This smaller model rarely gets stuck in a thought loop; the main problem is that arithmetic and reasoning are inherently error-prone, and cannot be easily solved by modifying a small number of weights. Researchers call this method of directly reading and writing parameters "weight-space programming". [Original link]