Login
You're viewing the functional.cafe public feed.

Replies

  • Nov 4, 2026, 5:56 PM

    The inability of people on this webbed site to simply enjoy a cool hack that answers a practically relatively pointless but interesting question is disappointing

    1. The question is exactly the question posed in the title of the article. The quotation I posted is out of this context; I chose it because it’s funny and punchy.

      No, IBM could not have used 1996 techniques on 1996 consumer-grade hardware. The question is whether the algorithmic advances since 1996 are good enough to get results on 1996 hardware. This is the author’s explicit premise.

      The point is not that IBM engineers were stupid for building an expensive supercomputer. In 1996 they had two choices: they could have worked on doing 30 years of research into classical AI techniques for chess solving and written software as efficient as we have today, or they could have taken what is effectively an embarrassingly parallel problem and thrown a lot of individuals CPUs at it. They chose the latter.

      The result is interesting in part because it’s surprising: the conventional wisdom is that hardware, especially in the 1990s, got faster so much more quickly than you could optimize software that it was better to just throw more hardware at a problem; the other conventional wisdom is that software gets slower in some kind of proportion to hardware getting faster. But 30 years of software optimization have made the specialized hardware thrown at the problem in 1996 retroactively obsolete, despite conventional consumer hardware itself getting so much more powerful in that time.

    2. Yes, the author admits to using an LLM to write code. In case you hadn’t noticed, I am not exactly a purist here even if I don’t use the things myself and am concerned about the consequences of their use.

      This is one-time, throwaway, non-open-source code to answer a specific research question. (If one of your concerns about LLM use is code quality: have you seen the usual quality of research code?) The modern chess engines used were not vibe coded by the article’s author; only the tools used to get 2026 code to run on a 1995 computer were. And yes, I am fully aware of the environmental etc. consequences of LLM use and training. But as long as the technology exists, this – not only single-user but single-use code – is about the least harmful application of it I can possibly think of.

    💬 1🔄 0⭐ 1
  • 💬 0🔄 0⭐ 0
  • Nov 4, 2026, 11:05 AM
    @dpk
    "Throughout this article, when I say "I wrote", what I really mean is "I asked an LLM to write". I used a combination of Fable 5.1, Opus 5.5, and GPT-6 to implement the code here."
    Oh well, I suppose the author can get by without my attention, or anybody's attention, you may want to remove this post.
    💬 0🔄 1⭐ 0
  • Nov 4, 2026, 11:13 AM

    @dpk it was an interesting read in the beginning (although it did raise a few questions I hoped will be answered later) but then I got hit with

    > (Throughout this article, when I say "I wrote", what I really mean is "I asked an LLM to write". I used a combination of Fable 5.1, Opus 5.5, and GPT-6 to implement the code here. (This article, though, is human text.).)

    Ehhh.
    Definitely cannot trust anything there then.

    💬 0🔄 1⭐ 0
  • 💬 1🔄 1⭐ 0
  • 💬 1🔄 1⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • 💬 1🔄 1⭐ 2
  • 💬 1🔄 0⭐ 2
  • 💬 1🔄 0⭐ 1
  • Nov 4, 2026, 6:08 PM

    @whitequark I think my view is that if I had to make a choice between, on the one hand, this result existing in the world but LLMs being used to facilitate it being found, and on the other, the result not existing at all, I would prefer the former. This is not the case with almost any other use of LLMs for coding.

    I do doubt that the vibe coded software was actually necessary, though. It’s still possible to get a (fairly?) modern Linux kernel for a Pentium box, and I imagine Stockfish would probably just compile and run normally on it?

    💬 1🔄 1⭐ 0
  • 💬 0🔄 0⭐ 1