Conversation

Notices

1. Embed this notice
   Aaron (hosford42@techhub.social)'s status on Wednesday, 29-Nov-2023 07:11:45 JST Aaron

   Nothing to see here, folks. Move along.

   Unpacking the hype around OpenAI’s rumored new Q* model
   https://www.technologyreview.com/2023/11/27/1083886/unpacking-the-hype-around-openais-rumored-new-q-model/

   • Embed this notice
     David Nash (dpnash@neurodifferent.me)'s status on Wednesday, 29-Nov-2023 07:11:43 JST David Nash

     in reply to

     • James Britt

     @hosford42 @jamesbritt If I understand this correctly (you have more experience in this topic, so please let me know if I don’t): this sounds like a way to make an LLM become even more fluent and “natural”-sounding than the current crop of GPTs…but, like all LLMs, still confined by the extent of its training data.

     If so, I’m utterly unsurprised that the usual suspects were alarmed by how “intelligent” it sounded. Verbal fluency seems to be the overwhelming measure they’re using, and that is known to be hugely problematic…

     …especially in the context of #ActuallyAutistic people. https://karawynn.substack.com/p/language-is-a-poor-heuristic-for

   • Embed this notice
     Aaron (hosford42@techhub.social)'s status on Wednesday, 29-Nov-2023 07:11:44 JST Aaron

     in reply to

     • James Britt

     @jamesbritt lol I might have done the same if I weren't already familiar with what is most likely its predecessor, A*.

     Which tells you a little bit about what Q* probably is: A fancy search algorithm, and that's it. (That's also alluded to by Yann LeCun's tweet that's embedded in the article I linked.) Basically, they are adding lookahead instead of just generating one token at a time, to generate content that's more human-seeming because it's more tightly constrained within the probability distribution of human language usage. To even suggest this is some magical AGI breakthrough that deserves a moral panic is ludicrous, and can only be justified by the execs involved (1) not really understanding how the technology they're using even works and (2) being high on the feelings of power they get from the magical thinking this leads them to.

     Brendan Jones repeated this.
   • Embed this notice
     James Britt (jamesbritt@mastodon.social)'s status on Wednesday, 29-Nov-2023 07:11:45 JST James Britt

     in reply to

     @hosford42

     "... created a new model, called Q* (pronounced Quasterisk)"

     How I read it in my head.