Construct the neural nets as lisp list structures. Now you can write functions and macros to operate on the internals.
The self-modifying nature of lisp allows ideas like optimizations, recognizing sections that have rich internal structure but sparse connections as a hint that there is some "clustered idea", etc.
The "it's just matrices" crowd have not really explored the deeper ideas.
This article demonstrates the main trait of 90s Lisp programmers - they are extremely condescending even when they turn out to be wrong. Since modern AI (the kind that actually works) doesn't need knowledge graph systems nor image computing.
I'm pretty sure computer algebra systems and automatic theorem provers work quite well. The CPU you wrote this comment on was developed using the other kind of AI that supposedly doesn't work: https://en.wikipedia.org/wiki/ACL2
Construct the neural nets as lisp list structures. Now you can write functions and macros to operate on the internals.
The self-modifying nature of lisp allows ideas like optimizations, recognizing sections that have rich internal structure but sparse connections as a hint that there is some "clustered idea", etc.
The "it's just matrices" crowd have not really explored the deeper ideas.