Breakout: 517.4 points… GPT-6-generated neural network performs 20 times better than the 500-million-step model

A neural network written by GPT-6 Astra scored 517.4 points on a simplified Breakout game using only the game's rules and source code, roughly 20 times higher than a reinforcement learning model trained over 500 million steps with the same architecture, MIT CSAIL researchers said in a study description published on the 2nd. The approach, which the researchers call "programming in weight space," involved giving the model a neural network with all weights set to zero and no training data or opportunity to run the game beforehand. The result suggests large language models may be able to produce functional neural network weights directly, bypassing costly training runs. Watch for whether the method generalizes beyond Breakout to more complex environments and whether MIT CSAIL releases a full paper with reproducible benchmarks.