According to CoinMeta, GPT-6 began to "handwrite" neural networks. The research team found that it was capable of directly analyzing and modifying neural network parameters without even the need for training. By giving GPT-6 a simplified version of a tile-matching game and a neural network with all weights set to zero, it ultimately achieved a score of 517.4, which is about 20 times that of a model that had undergone 500 million steps of reinforcement learning training. The researchers also had GPT-6 modify the weights of Alibaba's Qwen3.5-0.8b without retraining the model; by changing just 0.016% of the parameters, they increased the mathematical accuracy from 7.9% to 48.1%. This method of directly reading and writing parameters is known as "weight space programming."