3 ms·

> this PR is 90% AI!", _ignoring any previous failed attempt_ Someone pointed out that AI has a gatcha / intermittent reinforcement effect to it, which explain

• ben_w 1y ago
  You may well be correct that people are too optimistic about it being fixable, but I would say the "why" is a little more complex than that. It is also possible for artisans to produce functioning goods in low volumes, which is why "hand made" has often been seen a positive quality sign rather than a negative quality sign, and AI can do "artisanal" (same root as artificial). With mass production we want all three of "good", "cheap", and "fast", and we've been able to get it. But even that AI can do artisanal is not sufficient for AI to succeed, and likely still won't be sufficient even when AI is near the best you can hire in a human, because AI is a memetic monoculture: while the errors of many human artisans can be somewhat uncorrelated and therefore the failure modes different, if everything comes from a single AI mind, it all fails in the same general kind of way. Likewise the quotation about science advancing one funeral at a time: diversity beats even very smart geniuses. Memetic monoculture is what I think can't be fixed with current approaches, because current approaches need too much information for it to be possible to give each LLM* (even from different providers) a statistically distinct training set to learn from. * The image generators do seem to be able to get meaningfully distinct output styles. At least superficially, I'm not an artist, so I may be missing some common thread they share beyond the "indistinct blurry background item" that I count as an error rather than a style. But in this topic, even that error would count as a common failure mode.
  • pjc50 1y ago
    > if everything comes from a single AI mind, it all fails in the same general kind of way I'm reminded of https://qntm.org/mmacevedo https://qntm.org/mmacevedo ; even if we don't get brain uploading, we have a similar sort of thing going on with synthetic "personalities" of LLMs. > diversity beats even very smart geniuses. Some of HN will get very mad about that statement but it is often true. Of course, the best combination is a diverse collection of geniuses, which is why midcentury openish borders US did so well. > The image generators do seem to be able to get meaningfully distinct output styles Generally they can plagiarize in the style of any human artist you can name whose work is on the internet. Artists are, not unreasonably, upset about this.
    • ben_w 1y ago
      > I'm reminded of https://qntm.org/mmacevedo https://qntm.org/mmacevedo Mmm. My inspiration was from Alastair Reynolds putting indoctrinal viruses into the Revelation Space series, listening to an audio book of that several years before mmacevedo was published. At the time, I was thinking of how AI might enforce group-think, rather than write the content itself directly: https://benwheatley.github.io/blog/2019/12/30-18.46.50.html https://benwheatley.github.io/blog/2019/12/30-18.46.50.html > Generally they can plagiarize in the style of any human artist you can name whose work is on the internet. Artists are, not unreasonably, upset about this. Indeed, but so can LLMs of text styles. I meant more of the "voice" seems to be different between them — there seems to be more of a difference between getting SDXL and SD 3.5 (or DALL•E 2 and whatever it's called now it's baked into ChatGPT) to mimic, say, Hieronymus Bosch, than there is between getting GPT-4o and Claude Sonnet 4 to do a poem in the style of a Shakespeare soliloquy. On the other hand, this may just be my limited appreciation for the arts, as per XKCD: https://xkcd.com/1015/ https://xkcd.com/1015/
• marcosdumay 1y ago
  > The entire history of industrial process control is about trying to always produce good, or at least identical, results. Oh, but the entire history of gambling is trying to produce highly variable results with exactly the characteristics you mentioned. That's what gets people addicted.