Conversation

Notices

1. Embed this notice
   Robbie Coleman :verified: (erraggy@hachyderm.io)'s status on Sunday, 28-Jun-2026 23:39:33 JST Robbie Coleman :verified:

   • Thomas 🔭🕹️

   @thomasfuchs
   And they didn't consume barrels of diesel and produce tons of carbon either.

   • Embed this notice
     The Sleight Doctor 🃏🍉 (apostateenglishman@mastodon.world)'s status on Monday, 29-Jun-2026 01:29:23 JST The Sleight Doctor 🃏🍉

     • Thomas 🔭🕹️

     @thomasfuchs But a baby does inherit the genetic "memory" of humankind, giving their brains complex innate neural structures adapted for language. The baby learned so fast because much of that learning had already been done by countless generations of ancestors. 🤷🏻♂️

   • Embed this notice
     The Sleight Doctor 🃏🍉 (apostateenglishman@mastodon.world)'s status on Monday, 29-Jun-2026 02:41:58 JST The Sleight Doctor 🃏🍉

     • Thomas 🔭🕹️

     @thomasfuchs Specific languages are never genetically inherited; that wasn't the point. It's the underlying 👉🏼biological machinery👈🏼 for processing and learning any humanly comprehensible language - yes, with equal ease and speed - that's encoded into our genes.

     All human languages share common grammatical features and "deep structure" because our brains are hardwired for human language.

   • Embed this notice
     The Sleight Doctor 🃏🍉 (apostateenglishman@mastodon.world)'s status on Monday, 29-Jun-2026 03:22:51 JST The Sleight Doctor 🃏🍉

     in reply to

     • Thomas 🔭🕹️
     • Erotic Mythology 🌈

     @AimeeMaroux @thomasfuchs No, only that babies learn language so fast because they have a genetic head start. LLMs have to hoover up everything to build a statistical model of a language that a baby already has built-in - at least, to a large enough extent for linguists to discern grammatical commonalities and deep structure that's shared by all human languages.

   • Embed this notice
     Erotic Mythology 🌈 (aimeemaroux@mastodon.social)'s status on Monday, 29-Jun-2026 03:22:52 JST Erotic Mythology 🌈

     in reply to

     • Thomas 🔭🕹️
     • The Sleight Doctor 🃏🍉

     @ApostateEnglishman @thomasfuchs Biologist here. I'm not quite sure what your argument is. That developing the capacity for speech is equivalent to stealing every text humankind ever put into writing in order to reproduce human speech patterns?

   • Embed this notice
     Craig Stewart (pmb00cs@mastodon.online)'s status on Monday, 29-Jun-2026 03:54:25 JST Craig Stewart

     in reply to

     • Thomas 🔭🕹️
     • The Sleight Doctor 🃏🍉
     • eclexic

     @theeclecticdyslexic @ApostateEnglishman @thomasfuchs my point was solely in response to the idea that there is nothing like evolutionary pressure for LLMs during their development. Evolution is brutal, you survive to pass on your features, or you die. The development of LLMs will include that brutality towards models that don't meet the desired standards. The idea that there is something special about how evolution works because it's natural irritates me. It's simple brutality.

   • Embed this notice
     eclexic (theeclecticdyslexic@mstdn.social)'s status on Monday, 29-Jun-2026 03:54:27 JST eclexic

     in reply to

     • Thomas 🔭🕹️
     • Craig Stewart
     • The Sleight Doctor 🃏🍉

     @pmb00cs

     I would say it's more a problem of what metric you prune on... and also no amount of pruning will turn lead to gold.

     The underlying structure is definitely wrong. I don't think you can prune this thing into having a meaningful internal representation.

     Admittedly, we outsource a lot of cognitive load to these textual relations, and LLMs have tapped lots of it. But it doesn't go far enough without a useful representation underneath it.

     @ApostateEnglishman @thomasfuchs

   • Embed this notice
     Craig Stewart (pmb00cs@mastodon.online)'s status on Monday, 29-Jun-2026 03:54:28 JST Craig Stewart

     in reply to

     • Thomas 🔭🕹️
     • The Sleight Doctor 🃏🍉
     • eclexic

     @ApostateEnglishman @theeclecticdyslexic @thomasfuchs 'there's no equivalent of an evolutionary "pruning" process to weed out falsehood or mediocrity, the way natural selection does with maladaptive genes'

     What the hell do you think the people training these models do with them when they aren't up to scratch? They prune the shit out of the results that aren't up to scratch. And yet model collapse has still been shown to be inevitable, as has the generation of false but plausible looking output!

   • Embed this notice
     The Sleight Doctor 🃏🍉 (apostateenglishman@mastodon.world)'s status on Monday, 29-Jun-2026 03:54:30 JST The Sleight Doctor 🃏🍉

     in reply to

     • Thomas 🔭🕹️
     • eclexic

     @theeclecticdyslexic @thomasfuchs True, there's no equivalent of an evolutionary "pruning" process to weed out falsehood or mediocrity, the way natural selection does with maladaptive genes. Both will only accumulate and spread over time and poison future iterations. A chicken quickly dies if you feed it only its own eggs, with no external source of nourishment. Increasingly, LLMs are also consuming their own outputs as their slop proliferates across the internet.

     Model collapse is inevitable.

   • Embed this notice
     eclexic (theeclecticdyslexic@mstdn.social)'s status on Monday, 29-Jun-2026 03:54:31 JST eclexic

     in reply to

     • Thomas 🔭🕹️
     • The Sleight Doctor 🃏🍉

     @ApostateEnglishman @thomasfuchs the problem here is when LLMs learn from countless generations of ancestors... you get model collapse.

   • Embed this notice
     albion my way (albi@f.cz)'s status on Monday, 29-Jun-2026 06:49:47 JST albion my way

     • Thomas 🔭🕹️

     @thomasfuchs
     I bet they don't even read and analyze all your emails
     babies are literally suckers

   • Embed this notice
     Wraithe (wraithe@mastodon.social)'s status on Monday, 29-Jun-2026 12:30:57 JST Wraithe

     • Thomas 🔭🕹️

     @thomasfuchs I was about to say… Don’t put it past babies, they’re sneaky