Gary Marcus

I’ll see you you’re silly challenge and raise you. Show me one of the major platforms that runs the world: Google, Apple, Facebook, Amazon, that doesn’t use neural nets.

if you can rebuild any of those major platforms w end to the end nets and no symbol-manipulation let me know.

Symbols work well only in narrow domains were there's few ambiguities. A neural network outperformed a symbolic chess playing system (AlphaZero vs Stockfish). Perhaps you need to show me a symbolic system that can do state-of-the-art image or speech recognition.

for someone like me who advocates hybrid models, this is just a non sequitur; different tools for different jobs

@GaryMarcus that seems besides the point. Perhaps I've misunderstood, but there's quite obviously a difference between a bespoke solution (autonomous cars or software platforms) and understanding intelligence & whether solutions can emerge w/o innateness

Point is, this discussion seems to be shifting the innateness goalposts. NNs can learn symbol-like representations and abstractions, but the cited example domains don't use them because we want bespoke solutions rather than general AI.