Most of 2026 went by with inflated prices for DDR5 memory, SSDs, and GPUs but next year may finally bring relief from gamers and tech companies producing consumer electronics, at least according to one market analysis report on SK Hynix coming out of Seoul, Korea. BNK Investment & Securities analyst Lee Min-hee shares a detailed […]

  • b3nsn0w@lemmy.world
    link
    fedilink
    English
    arrow-up
    1
    ·
    1 hour ago

    we’re already getting there. open weight models are only about 3-4 months behind now, and for a lot of tasks a small model is perfectly fine these days, they’ve gotten more than good enough recently.

    the hyperscalers are already focusing on cost optimisation too, microsoft just released some tiny coding models, and google’s main general purpose model is absolutely tiny now. (they’ll tell you argon is the main model, but it’s clearly hydrogen imo.) for everyone else, qwen 4 is about to drop too.

    and when the bubble crashes, which is pretty much inevitable now, and the money for bigger and bigger training runs dries up, you can expect open weight models to close in that gap too. the future of ai is a piece of software you run on your own computer, not a cloud service. there will still be some cloud stuff going on, as it has been with everything else too, but it won’t be mandatory.