744
Generative AI Is an Engineering Disaster. A shockingly inefficient trillion-dollar project. (www.theatlantic.com)
submitted 3 months ago by to c/technology@lemmy.world
182 comments fedilink hide all child comments
you are viewing a single comment's thread
view the rest of the comments
[–] 38 points 3 months ago (19 children)

In 1897, they built the first music synthesizer. It worked, but it took up the basement of an entire city-block sized building, so it was essentially useless. After a few decades of development, it could fit in a suitcase, and be carried around.

Data Centers are like that 1897 synthesizer. Sure, it works, but at what cost? It clearly isn't ready for prime time. Go back to the drawing board, tweak the problems, including regulations, and maybe in a couple of decades, we take another run at the new and improved version.

  • permalink
  • fedilink
  • source
  • hideshow 19 child comments
    •
  • [–] 36 points 3 months ago* (12 children)

    Data Centers are like that 1897 synthesizer. Sure, it works, but at what cost? It clearly isn’t ready for prime time. Go back to the drawing board, tweak the problems, including regulations, and maybe in a couple of decades, we take another run at the new and improved version.

    The issue with AI is not a technical or development problem. It's not even a regulation problem. It's a capitalism problem. Infinite growth will still be as unscalable in a hundred years as it is now no matter how good and mature the tech is.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 12 child comments
    •
  • [–] 15 points 3 months ago (11 children)

    There are also hard mathematical limits stalling AI growth. Frontier models haven't improved in like a year despite being fed money by basically the entire global economy. Diminishing returns on steroids basically. They're already at the limit of what they can make, and going further gives a much smaller improvement in the model, and now I hear there might not be enough human written material on the internet to train them.

    It also looks like hallucinations are inherent to LLMs and you can't get rid of them. It's a side effect of the model. What commercial applications are there then, if you can't guarantee the output? It's worse than a human for most things since it doesn't know truth from lie and will confidently say both as if they're fact. It also looks like prompt injection isn't something you can fully guard against either.

    What's the value proposition when you can't trust the output and the model might give a massive refund or discount to a customer and the courts rule the AI speaks on behalf of your company?

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 11 child comments
    •
  • [–] 7 points 3 months ago

    I feel like calling „hallucinations“ a side effect isn’t really describing the issue properly. They’re not a side effect, nor are they hallucinations. It implied that there’s somehow something that distinguishes „correct“ output from „incorrect“. There isn’t, it’s all just output. The output resembling actual factual reality is statistical chance.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 5 points 3 months ago

    It’s worse than a human for most things since it doesn’t know truth from lie and will confidently say both as if they’re fact

    It works for most executives and sales folks.

    Baseless confidence is the recipe for business success, which is why they love these AI chatbots.

    Bigger problem for the business leaders is how sycophantic they want to be to the user. If an insurance company used it for claims, it might actually approve a claim, and that would be unforgivable for them.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 1 point 3 months ago (3 children)

    Haven’t been improved in a year? By what metric are you basing that assertion?

    IP theft is probably the main one I’ll concede the point on - that damage was done long ago so they haven’t “improved” on it.

    But whether it’s reasoning or generation or building things… that’s a crazy take. Unless you consider them like a chatbot companion? I wouldn’t really know much on that front, I’ll concede.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] 6 points 3 months ago (2 children)

    It's been marginal improvements for like 18 months now. I don't know if you remember what they promised that long ago but current frontier models just ain't it.

    If you don't believe me, then why? Put your argument in numbers.

    Anthropic and openAI have both spent nation-state levels of money training these models and they only seem marginally better compared to the last ones? Maybe larger context and better reasoning but they still hallucinate, they still make the same mistakes and pitfalls.

    Even with tokens getting dramatically cheaper inference on mythos or other frontier models is so expensive they need to start replacing skilled professionals and right now they just can't. Productivity doesn't seem to go up from AI use, if anything net productivity for an org goes down from people outsourcing human cognition onto their colleagues.

    "How about instead of me summarizing this report i use Claude and then my colleague spends the cognitive effort deciding if Claude lied or not"

    The guy using AI for everything looks super productive and the people stuck dealing with the work he's "getting the AI to do for him" look like under performers when they're actually load bearing in this new setup.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 3 points 3 months ago (1 child)

    Tokens aren't going to get cheaper. Tokens must get more expensive, and soon. AI companies are making big losses even if you ignore the stratospheric debt for the immemse quantity of hardware.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 1 point 3 months ago

    Yeah you're right they just appear cheaper due to circular financing and some memory improvements. I was trying to steel man my argument into the strongest possible point that still fails to make sense.

  • permalink
  • fedilink
  • source
  • parent
  • [–] -3 points 3 months ago* (2 children)

    It’s worse than a human for most things since it doesn’t know truth from lie and will confidently say both as if they’re fact.

    I think that's where you're wrong. It's really not worse than a human. It's smarter than like 90% of the population already. No it's not perfect but it doesn't have to be. Humans literally hallucinate and lie, all the time. AI hallucinations is an athropomorphism, AI metaphorically hallucinates, humans actually literally do.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 4 points 3 months ago* (1 child)

    The 10+% of the time it's wrong you can't blame it, sue it, imprison it, or apply leins against it if it causes real world damages to people or the company that operates it

    "For entertainment purposes only" is still the level of liability AI companies operate at. Air Canada just had a ruling that anything their bot says is the speech of Air Canada and they're bound by it. So far there's always ways to prompt inject, and you sometimes don't even need to do that. Just guiding the conversation in ways that are difficult to pin malice on to is enough.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 8 points 3 months ago* (last edited 3 months ago) (2 children)

    In 1897, they built the first music synthesizer. It worked, but it took up the basement of an entire city-block sized building, so it was essentially useless. After a few decades of development, it could fit in a suitcase, and be carried around.

    As per the 1890s, it probably also had to be lubricated with orphaned child blood.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 4 points 3 months ago

    It's not even that it isn't 'ready for prime time', largely to the extent it works, it works with not so crazy requirements.

    The problem is that they don't settle for what it is, they try to overextend it. In software development for example, in the cases where it works at all, you are 95% of the possible successes within 3 or 4 iterations. Problem is they demand that extra 5% which takes an order of magnitude more. Note that '100%' here is the max success possible with AI, not 100% success in general, that number varies greatly with context. So it might be in a certain scenario more like going from 19% AI curated to 20% AI curated, at huge incremental expense.

  • permalink
  • fedilink
  • source
  • parent