Post #958350

2026-03-30 19:39 UTC

I&#39;ve read a couple of breathless articles about Donald #Knuth&#39;s use of #LLM systems, and found (unsurprisingly) that he has a paper about it, if you prefer your news without someone else&#39;s hot takes: https://www-cs-faculty.stanford.edu/~knuth/papers/claude-cycles.pdf (Sidebar: apparently he&#39;s working on *another* volume of #TAOCP? Good lord. I still haven&#39;t worked through the exercises in Vol. 1.) #AI #Programming #ComputerScience

Replies (2)

@Kadin2048@mefi.social 2026-03-30 19:57

But if you want *my* hot take: anyone who reads that paper is going to be lost without a pretty strong background in modern math, at least a solid undergraduate level. The part that #Claude did seems pretty mechanistic; it literally tries a brute-force approach, but all of its attempts in general could be considered a sort of brute force, something that I'm confident Stappers or Knuth could have done by hand if they wanted… it just does them faster. More importantly, though, is that it takes someone of Stappers' or Knuth's caliber to recognize that the solution is, in fact, valid and generalizable. Although the model does stop for one partial solution, there was a significant amount of steering getting it to that point. Also, being able to ask the question in a rigorous way, or realize that it's even interesting, is itself a creative act that the machine can't do. I think there's an analogy here to compilers and compiled languages.

Open ##1725142
@alexmu@social.vivaldi.net 2026-03-30 21:18

@Kadin2048 he's not using llms. Someone used claude to solve a problem posed by Knuth in volume 4. Knuth was impressed by the result.

Open ##1725143