[fastneutron]

The idea of neurosymbolic systems has been in the air a long time, but every time I look at the commentary of an article like this I’m surprised at number the “OMG why didn’t anyone think of this?” type of comments.

For a while I got the impression that an ideological undercurrent of “DL vs GOFAI” had gotten in the way of more widespread exploration of these ideas. Tao’s writing here changed my view to something more pragmatic, that being the formalization of the symbolic part of neurosymbolic AI requires too much manual intervention to easily scale. He is likely onto something by having an LLM in the loop with another system like Lean or Athena to iterate on the formalization process.