Elektrine lite

← Feed

@troed@swecyb.com

2026-07-19 19:52 UTC

@mikalai@privacysafe.social I really do mean emergent. LLMs weren't even researched for "AI" to begin with. The "Attention is all you need" paper that spawned today's LLM architecture is from 2017 - yet it wasn't until ~2022 that the emergent behavior of LLMs started to lift eyebrows. https://arxiv.org/abs/2206.07682 @david_chisnall@infosec.exchange @hfalcke@mastodon.social

Replies (3)

  • @mikalai@privacysafe.social 2026-07-19 20:29

    @troed@swecyb.com Right. I forgot that they give their very particular definition of "emergent". Can't you see, that when a word with common sense, loaded common sense is used, but, redefined, ... it feels like placing your address beyond the boundary of an array. Get yourself a hook into the rest of the system. Is this from UK? Do gentlemen do this? @david_chisnall@infosec.exchange @hfalcke@mastodon.social

    Open ##3941748

  • @mikalai@privacysafe.social 2026-07-19 20:13

    @troed@swecyb.com I also want to point out that same people use LL Model name, while it is a Condensate. It helps them. Same code on bigger pool of data gives better output. This isn't an emergent. Geometry of actions was set by code. Inputs have changed. @david_chisnall@infosec.exchange @hfalcke@mastodon.social

    Open ##5006539

  • @mikalai@privacysafe.social 2026-07-19 20:17

    @troed@swecyb.com Also, my understanding from other papers, scaling hypothesis has stopped. That's it. When you time it, you undetstand why a particular push towards startups/looking_for_use happens -- there is nothing to gain at basic level .... if one follows same path. Hence, Yann LeCun. @david_chisnall@infosec.exchange @hfalcke@mastodon.social

    Open ##5006540