Elektrine lite

← Feed

@norikawa@efdn.club

Post #968570

2026-04-07 00:44 UTC

Thinking about I decided to go and learn about how LLMs actually work and the structure behind them and somehow it was even less impressive than what I thought was going on and that wasn't a particularly glowing view of them to begin with

Replies (2)

  • @norikawa@efdn.club 2026-04-07 00:47

    It all boils down to what are basically word frequency tables, but where it's complicated by having each word be associated with others words (so like, it's tables of how frequently certain words show up together) And then slap a whole bunch of neural network layers on top of it. And then also add some more layers to counteract the other layers being bad I guess??? "Hey let's just add more layers and call it a breakthrough" is basically all that's ever been done since they stopped working on improving anything to do with those word frequency-ish tables

    Open ##3147648

  • @The_T@glaceon.social 2026-04-07 00:46

    @norikawa@efdn.club don't feel obligated to but I'd be curious to hear a simplified explanation of just how inconsequential they are

    Open ##3147653