Elektrine lite

← Feed

@ben@mastodon.lubar.me

2026-04-29 23:51 UTC

@krutonium pretty much the same technology that runs an LLM is in a markov chain bot. it's just scaled up by a billion times. just like how image denoising algorithms make offline rendering 3D scenes cheaper and more efficient but if you instead try to force them to generate images from scratch they're extremely expensive to run. or how cryptocurrency is bullshit but cryptography is universally useful.

Replies (1)

  • @lina@vt.social 2026-04-30 05:24

    @ben @krutonium LLMs have terrible scaling, and it's especially bad when you're trying to pretend it's an all knowing intelligent entity you can just prompt to do your bidding (which it isn't, but it takes a huge LLM to fake it well enough it sounds like one). Small scale LLMs actually have many uses and can run on consumer hardware (and train on prosumer hardware, much less than that VAC rack), but you have to give up that fallacy that you just train one giant model and prompt it to do things. You're supposed to train/fine tune them for specific, narrow use cases (which nobody does...).

    Open ##1875340