Science Researchers Claim LLMs Can Edit Neural Weights Directly WireByte A new study titled 'Sparks of Weight-Space Programming' claims frontier models like GPT-6 Astra can interpret and modify raw neural network weights without gradients. The authors report significant performance gains in math and instruction following tasks by editing tiny fractions of weights. However, expert analysis suggests the methodology may be misleading, arguing that the results rely on complex underlying mechanisms rather than simple weight inspection alone.