This post is a call to action. I've been thinking about AI a lot recently (hard not to these days). Last night I had an idea that sounds interesting and impractical. I don't have the means or desire to pursue it, but someone out there certainly does. Hopefully this can motivate them.
The idea: Make an AI compiler. Training data would be trivial to obtain, and doing this at scale might yield a network that can faithfully map code to machine code. What I like about this idea is that there's no fundamental reason a highly effective neural network compiler couldn't exist. But regardless of how well it works on test inputs, without rigorous output guarantees, it's still a terrible compiler. I think it would be interesting, and utterly pointless, to see this done at scale.