Conversation

Notices

1. Embed this notice
   Rich Felker (dalias@hachyderm.io)'s status on Tuesday, 28-Apr-2026 01:47:21 JST Rich Felker

   in reply to

   • El Duvelle
   • Artemis
   • zenkat

   @zenkat @elduvelle @artemis It's not even your entire "conversation". Since context is costly, once length of interaction grows beyond a small threshold, the system uses the LLM to "summarize" it (something it can't actually do; it just produces summary-shaped slop) and replaces all but the newest part of the interaction with the "summary" it vomited out. So you get compounding lossiness. 🤡

   • Embed this notice
     zenkat (zenkat@sfba.social)'s status on Tuesday, 28-Apr-2026 01:47:22 JST zenkat

     in reply to

     • El Duvelle
     • Artemis

     @elduvelle @artemis Great analysis. As a software engineer and a biologist, I think your intuition is spot-on.

     The biggest difference between LLMs and biological intelligence is that LLMs *don't learn from experience*, at least not immediately in the moment like we do. They have distinct "training" and "inference" modes. When you are interacting with Claude, you are in "inference" mode. It gives likely answers based on its training, but *it's not learning anything new".

     The model has no memory of is interactions with you. Literally! All of the persistence you see in a session is really just a fancy trick -- they just feed your entire chat session back into the LLM, and ask "what is the next most likely token"? Again, and again, and again. Every time, your *entire conversation* is fed back into the model, just to predict the next token.

   • Embed this notice
     El Duvelle (elduvelle@neuromatch.social)'s status on Tuesday, 28-Apr-2026 01:47:24 JST El Duvelle

     in reply to

     • Artemis

     @artemis
     I would say this is a pretty good description, and what is most annoying (to me, as a neuroscientist) is that we could be looking into how actual intelligence works (in humans and other animals) and how to imitate it in computational models, but most of the funding that could be used for this is now being redirected towards generative "AI" instead.. well, hopefully this is just a phase..

   • Embed this notice
     Artemis (artemis@dice.camp)'s status on Tuesday, 28-Apr-2026 01:47:25 JST Artemis

     in reply to

     I already knew this form of "AI" is bullshit & not intelligent in the least, but now I fear it less than I used to.

     The AI companies want you to believe breakthrough after breakthrough is coming, that they will continually innovate but...

     ...that's actually pretty unlikely. That's not really how shit works. Throwing money at a project doesn't make it possible for that project to perpetually improve. There may be a limit.

   • Embed this notice
     Artemis (artemis@dice.camp)'s status on Tuesday, 28-Apr-2026 01:47:26 JST Artemis

     in reply to

     This happens with lots of new technologies. Eventually, you will have gotten just about as far as you can with a particular one.

     Without something else new (which does sometimes happen, but sometimes not), the thing simply is what it is & is only capable of whatever tasks its design allows for.

     You can't extrapolate infinite progress from past progress, *especially* not expecting the same rate of growth as when the technology was first being created.

   • Embed this notice
     Artemis (artemis@dice.camp)'s status on Tuesday, 28-Apr-2026 01:47:28 JST Artemis

     in reply to

     That is my layperson's understanding of the current chatbot AI situation anyway.

     The outputs seem surprisingly plausible in most instances, but the makers of these things have just about hit the limits of what this technology can do (at least without someone coming up with a genuinely new breakthrough).

     It seems that these companies are now just fiddling with prompts (or having the bots do it themselves) & calling it "development".

   • Embed this notice
     Artemis (artemis@dice.camp)'s status on Tuesday, 28-Apr-2026 01:47:29 JST Artemis

     in reply to

     This is barely a simulacrum of "intelligence". And most importantly, it doesn't seem like there is a lot of obvious possibility for genuine progress on these things, as it is no longer a matter of innovation but of adding prompts inside prompts inside prompts.

     Performance in specific circumstances can be "improved" for the user experience, but it's not actually that the thing is "smarter" but that the flowchart of steps before the output is given to the user has been made more complex.

   • Embed this notice
     Artemis (artemis@dice.camp)'s status on Tuesday, 28-Apr-2026 01:47:30 JST Artemis

     Something has changed in the way I view AI since Claude code leaked.

     I should note: I am depending on some other people's analysis, because a) I haven't looked in depth myself and b) even if I did I wouldn't know what I was talking about.

     But my understanding is that Claude & the like are filled with all these recursive loops of checking, "creating" an AI agent for a specific tasks, like reviewing the output before it's given to the user to "check" for things.