Elektrine lite

← Feed

@cthos@mastodon.cthos.dev

Post #1223851

2026-03-18 04:12 UTC

@dhobern And "Reasoning" models literally just run themselves in a loop, feeding the output from one iteration back to itself a number of times with the aim to improve a result. I don't actually know how they determine when to stop looping, I presume it's when it hits / outputs a stop token.

Replies (1)

  • @wnd@fosstodon.org 2026-03-18 19:02

    @cthos @dhobern if it is at all like any of the language word or sentence encoders I've used (BERT, ELMo, SBERT et al), you terminate at a point either when bored because you have run out of loops or the output from your loss function (e.g. cosine similarity between vectors) is below a given threshold. (I am also aware that I am unlikely to beat this wordsalad at an point later this week, and it is only Wednesday).

    Open ##1223856