Wherever I wander I wonder whether I’ll ever find a place to call home…

  • 0 Posts
  • 7 Comments
Joined 1 month ago
cake
Cake day: December 31st, 2025

help-circle

  • I think it’s called an inferencing chip. I read about it a few months ago.

    Basically, the way it was explained, the most energy-intensive part of AI is training the models. Once training is complete, it requires less energy to make inferences from the data.

    So the idea with these inferencing chips is that the AI models are already trained; all they need to do now is make inferences. So the chips are designed more specifically to do that, and they’re supposed to be way more efficient.

    I kept waiting to see it in devices on the consumer market, but then it seemed to disappear and I wasn’t able to even find any articles about it for months. It was like the whole thing vanished. Maybe Nvidia wanted to suppress it, cause they were worried it would reduce demand for their GPUs.

    At one point I had seen a smaller-scale company listing laptops for sale with their own inferencing chips, but the webpage seems to have disappeared. Or at least the page where they were selling it.


  • Maybe I didn’t mean blockchain, cause I’m still not really certain what it is. I mean like the fediverse itself, or a mesh network, where a bunch of hobbyist self-hosting their own servers can federate as a system of nodes for a more distributed model.

    Instead of all the compute being hoarded in power-hungry data centers; regular folks, hobbyists, researchers, indie devs, etc., would be able to run more powerful simulations, meta-analyses, renderings, etc., and then pool their data/collaborate on projects, and ultimately create a more efficient and intelligently guided use of the compute instead of simply “CEO says generate more profit! 24/7 overdrive!!!”

    At the very least, a surplus of cheap RAM would expand the computing capabilities of everyone who isn’t a greedy corporation with enough money to buy up all the expensive RAM.


  • So there would be an enormous surplus and a lot of e-waste. That’s a shame, but that’s going to happen anyway. I’m only saying that the silver lining is that it means GPU and RAM would become dirt cheap (unless companies manufacture scarcity like the snakes they are).

    Industrial applications aren’t the only uses for it. Academic researchers could use it to run simulations and meta-analyses. Whatever they can do now, they could do more powerfully with cheap RAM.

    Gamers who self-host could render worlds more powerfully. Indie devs could add more complex dynamics to their games. Computer hobbyists would have more compute to tinker with. Fediverse instances would be able to handle more data. Maybe someone could even make a fediverse MMO. I wonder if that would catch on.

    Basically, whatever people can do now, more people would be able to do more powerfully and for cheaper. Computations only academia and industry can do now would become within reach of hobbyists. Hobbyists would be able to expand their capacities. People who only have computers to tinker with now would be able to afford servers to tinker with.

    “Trickle-down” is a bullshit concept, as everything gets siphoned to the top and hoarded. But when that cyst bursts, and those metaphorical towers come crashing down, there’s gonna be a lot of rubble to sift through. It’s going to enable the redistribution of RAM on a grand scale.

    I’m not pretending it’ll solve everyone’s problems, and of course it would have been better if they had left the minerals in the ground and data centers had never grown to such cancerous proportions. But when the AI bubble bursts and tech companies have to liquidate, there’s no denying that the price of RAM would plummet. It’s not a magic bullet, just a silver lining.