GPT-6 Begins 'Handwriting' Neural Networks: Turning Game Rules into Weights, Achieving 20 Times Higher Scores in Brick Breaker

Beating AI News Flash: Neural networks often have millions or hundreds of millions of parameters, which in the past were mainly adjusted repeatedly through training algorithms. However, a research team at MIT has found that GPT-6 Astra can already directly analyze and modify these numbers, and can even write a neural network from scratch without training.

The researchers gave GPT-6 the rules of a simplified brick-breaker game, the source code, and a neural network with all weights set to zero, and asked it to fill in the parameters itself. No training data was provided, and trial play was not allowed. GPT-6 calculated the ball's landing point on its own and wrote the logic for prediction and paddle movement into the network weights. The final game score reached 517.4, about 20 times that of a model with the same structure trained through 500 million steps of reinforcement learning. However, in two other games, it did not surpass the model trained for 500 million steps.

The researchers then had GPT-6 directly modify the weights of Alibaba's Qwen3.5-0.8B. Without retraining the model, it changed only 0.016% of the parameters, raising the MATH-500 math accuracy rate from 7.9% to 48.1%, and nearly tripling the instruction-following score.

It turned out that Qwen often fell into thinking loops, reasoning repeatedly without giving an answer. Before the modification, 93% of math test responses exhausted 32,768 tokens; after the modification, that dropped to 33.1%. GPT-6 essentially fixed the problem that prevented the model from answering questions normally.

But when switched to Gemma 3 270M, which has only 270 million parameters, GPT-6 was unable to significantly improve math performance. This smaller model rarely fell into thinking loops; its main problem was that arithmetic and reasoning themselves were prone to errors, which could not be easily solved by modifying a small number of weights.

The researchers call this method of directly reading and writing parameters "Weight-Space Programming."