[–] 1 point 7 months ago

The productivity increases are real for office work. Absolutely not for 'generative' tasks - using them to do anything that takes a modicum of thought is a recipe for frustration and burned time - but they make pretty good search engines for unstructured information.

Now, you don't need a whole-ass model for that, a vector database would do just fine, but since everyone rushed to hook the models in to all their repos and internal wikis it is a convenient omni-search tool.

Now was that worth all the bullshit over the last 4 years as the bubble inflated? Not at all. But it is a real use case and has real value.

I even have leaders now pushing everyone to keep their docs up to date and to make sure tables have metadata fields and such so that agents can make best use of the information. This is particularly hilarious to me because they didn't care to make this investment when people would benefit, but doing it now makes my life easier anyway.

  • source
  • parent
  • context
  • [–] [S] 1 point 7 months ago

    More productive is debatable, but even if it never improves I doubt it's going anywhere. This is anecdotal, but every housewife I know uses it religiously for recipes/meal planning/house project planning/etc.

    Also a big part of it being unsustainable right now is training costs, if companies stop trying to make better models it's very feasible to continue offering the current models in financially sustainable way. But that only works until a rival company comes out with a bigger better model, so everyone is having to dump loads of cash into training constantly.

  • source
  • parent
  • context
  • [–] [S] 5 points 7 months ago (1 child)

    Also companies would have to hurry to regain workers they replaced with ai when ai bubble bursts.

    Pretty sure this won't happen, a lot of the tech companies are reliant on investors to keep running. Before AI, the investors were cutting off money, and it was causing mass layoffs across the tech sector.

    AI hype for investors throwing money at companies again, which has largely stopped the layoffs. It's also why every company was throwing themselves into AI even if their company model didn't really benefit. Saying they were AI focused got the investor money flowing again.

    So if AI bubble pops, not only will a lot of AI companies layoff people, but the venture investors will probably tighten up on their investments, causing all the not yet profitable tech companies to shut down or rapidly enshitify to try to turn a profit. Meanwhile we'll see far more layoffs across the tech industry.

    And with how interconnected economies are, it will probably trigger a pretty large economy wide recession.

    Overall I expect it to be a bad experience for everyone.

  • source
  • parent
  • context
  • [–] 4 points 7 months ago* (2 children)

    datacenters failing would likely give positive surge to component economy, as all the wasted capacity would be freed. Also companies would have to hurry to regain workers they replaced with ai when ai bubble bursts.

    I can see nothing but positive effects from it, though they will likely be turned into bad ones by the rich. Or maybe one of the companies survives it and gains complete dominance they exploit slowly.

  • source
  • [–] 4 points 1 year ago (1 child)

    None of this is going to happen.

    Multiple articles have come out in the past week with techbros blathering about putting giant stacks of power-hungry compute in fucking orbit, and not a single one of them has explained how a) any of it is going to be cooled and b) how any of it will be even remotely (as in, within 2 orders of magnitude) cost-competive with data centers on the ground.

    This is a scam seeking to part extremely rich fools from their money.

  • source
  • [–] 1 point 1 year ago* (last edited 1 year ago)

    Right off the bat their writing is exaggerating the realities, which makes me not want to bother reading any more.

    • Dial-up connected virtually every time if you chose an ISP who'd provisioned enough modems for their customers to use.
    • The "first time in history, you can exchange letters with someone across the world in seconds" was over a decade prior to that - it just cost an arm and a leg for the service to do so (something like $7-10/hr in early-mid 1980s money, if I recall correctly). And that's just talking about commercial services available to the general public, as the universities, large businesses, and government/military that created the early networks had that ability before even then.
    • One-minute video did NOT take "hours" to download unless it was in an absurdly high quality for the time - something pretty unusual because of the common speed (and storage requirement) constraints.

    Yeah, maybe I'll look at this later when my bullshit alarm settles down.

  • source
  • [–] [S] 1 point 1 year ago

    Some guides would be great!

    I personally find updates to the big mainstream models interesting as well, both as a sign of where things are going, and also with how they can be worked into personal workflow stuff. For example I was reading about people using Gemini API for more powerful OCR for mass document digitizing, and it was interesting to me to read about what it's relative strengths/weaknesses are compared to traditional methods.

  • source
  • parent
  • context
  • [–] 2 points 1 year ago*

    Oh, yeah, and I’m not sure anything can be done about downvotes. To be blunt, there’s a large section of Lemmy that will just stop reading and downvote anything that has the word “AI” in it.

    At the same time, you won’t find many people here get excited about updates to “tech bro” focuses like ChatGPT or whatever.

  • source
  • [–] 3 points 1 year ago (1 child)

    There’s some overlap with more active communities like localllama I think.

    But yeah, as other said, focus on local running! Or at least using local interfaces with APIs and such, pretty much anything but mainstream stuff like Grok and ChatGPT.

    I have some knowledge here. Should I write up guides to exllamav3 and ik_llama.cpp?

  • source
  • [–] 3 points 1 year ago

    Newly subscribed member here. Personally I'd be interested in hearing about use cases/experiences for self-hosted LLMs and other AIs, hardware accelerators like google coral etc.

    ML is also interesting (just set up paperless-ngx and it is amazing how well it learns with the meager resources it takes).

    Also as IT professional, it would be interesting to hear experiences about training and certifications around AI and ml.

  • source
  • [–] 3 points 1 year ago*

    You're going to have to ban drive-by downvoters.

    Personally, I'd like to see a primary focus on systems that can be run locally, and systems which run on lower capability hardware, but that's just me.

    When it comes to stuff people can't really run at home, maybe also encourage posting on systems that aren't LLMs or art regurgitators, for example machine learning & "AI" in areas like weather forecasting, drug design, materials science, etc.

    (edit: I suppose not all of those uses are generative... 🤷)

  • source
  • The best explanation of "AI" I've managed to come up with (as someone who programmed in a former life, so it's still kinda techie) is: pattern-matching on an insane amount of steroids with some LSD thrown in for its "magical" effects.

  • source
  • view more: next ›