Like most of Wolfram's scientific work, it's hard to pin down whether he's a genius or an eccentric crank. He built his own theory of the universe from first principles. It all sounds plausible, but most of the large-scale assertions are only empirically verifiable with the necessary compute power (his law of computational irreproducibility says effectively that). It makes for really cool cellular automata graphics but I'm not convinced that it's necessarily a new kind of science with broad explanatory power at scale. But it might? Maybe I'll try to read NKS again, I "only" got 200 pages into it the first time.

I'm generally impressed with his software work though. Later in the podcast, he talks about training an LLM to write Wolfram code and automatically execute it, with the ability to refine the intermediate code. I've always appreciated Wolfram Alpha as a powerful computation tool and combining it with an LLM is a pretty creative approach. You can express a complex analysis in plain language but, instead of getting a full response likely to contain hallucinations, you get some code in a high-level programming language to then generate the response. Even if you just execute it and accept the result, at least you have a copy of the "logic" used to generate your result. Neat approach to escape the black box problem of LLMs, at least for the kind of questions that you'd normally write in Wolfram.