That’s what I’ve been arguing with a fellow programmer recently. Right now you have to tell these programmer LLMs what to do on a function-by-function basis, because it doesn’t have enough capacity to think on a project level. However, that’s exactly what can be improved by scaling the neural network up. Right now the LLMs are limited by hardware, but they’re still using off-the-shelf GPUs that were designed for a completely different use case. The accelerators designed for AI are currently in the preproduction phase, very close to getting used in the AI data centers.