Discussion about this post

User's avatar
Adam Chlipala's avatar

Yeah, stay tuned for details of "TPU for formal verification"!

As for the potential of further-scaled deep learning to provide all the intelligence we need forever, I think it's hard to argue that, in domains where we can provide more structured methods, it isn't possible to reduce costs and increase performance by enough to provide major advantages, for instance in any domain where the party that decides more quickly has a real edge. I'm putting my money on at least neurosymbolic methods (so including important parts beyond deep learning) becoming essential to compete in *low-latency* automated theorem-proving and automated program synthesis, domains that I know well.

Abel Nieto's avatar

This is a very interesting and ambitious research programme, and I would love to learn what the "TPU for formal verification" would look like. The obvious counter to this (the one the AI maximalists would offer) is that we don't need to bother with any of this; rather, we just wait for a couple of more iterations of the current transformer-based systems to land (or the "world models" style systems) and offload all of the research efforts to them. What's your thinking around this hard takeoff scenario?

No posts

Ready for more?