Gang Fang's Blog

With Software 2.0, we say the broader the scope of responsibility DL takes on, the better the end results are. Using this as a prior, should we train a model to produce bits that a CPU can execute directly instead of code in a modern programming language? What is the point of producing code if machines can close the whole loop of software-engineering, where humans write and read HUMAN-READABLE code?

But the immediate challenge I see is LLMs won’t be able to help because the label for this supervised learning task is not language anymore, so how should training be done in this case? It’s a possibility that the label is now the behavior of the machine that executes the independent variables.