Login
You're viewing the mstdn.social public feed.
  • myrmepropagandistfuturebird@sauropods.win
    Apr 29, 2026, 10:04 AM

    The voice LLM never says "I cannot create a working timer" it just pretends that it can.

    Sam explains that this issue could be fixed by "implementing timers" which means someone would write some code to intercept these kinds of requests and treat them differently.

    You know, they'd need to write one of the core functions of a voice chat personal assistant software.

    But the magic of LLMs was supposed to be that they'd never need to do any of that work, right? 2/2

    💬 4🔄 12⭐ 39

Replies

  • 💬 3🔄 0⭐ 21
  • myrmepropagandistfuturebird@sauropods.win
    Apr 29, 2026, 10:09 AM

    Also, I think husk.irl did a good job showing how the LLM fails, but he didn't show the most scary example.

    This would be if someone really ran a mile. And it takes them 9:00 or 12:00min but they are told it was 10.

    10 min is a very reasonable answer to the question. It's a typical mile time for a person who runs regularly and is good at it for an amateur. Most people using the tool earnestly will never spot the error or the lies until it's too late.

    💬 1🔄 5⭐ 28
  • 💬 1🔄 0⭐ 2
  • 💬 0🔄 0⭐ 2
  • 💬 0🔄 0⭐ 1
  • Philippa Cowderoyflippac@types.pl
    Apr 29, 2026, 1:31 PM

    @futurebird I don't think anybody's ever linked me a publication of his? Which isn't a strong statement that he's something else, but...

    (my own list is exactly one extended abstract, but there are also plenty of the Scottish PL community who'll happily be all "yep, we know Philippa!")

    💬 0🔄 0⭐ 0
  • A Flock of Beaglesburnitdown@beige.party
    Apr 29, 2026, 10:27 AM

    @futurebird it's not failing, it's performing exactly as intended. it's a Large Language Model, not a Large Math Model. it can't deal with numbers because it was never supposed to do that, and also computers don't know what numbers are anyway. they don't know anything and are not capable of knowing anything.

    💬 0🔄 0⭐ 0
  • dominicbarometz@treehouse.systems
    Apr 29, 2026, 10:09 AM

    @futurebird the fact that "I can't do that" and "I don't know" weren't foundational features is malpractice and it's depressing that the people who sell it don't see that.

    "But that's hard", they may reply. Yes, it is. That's exactly why leaving it to the user is the wrong choice!

    💬 2🔄 0⭐ 4
  • FeralRobotsFeralRobots@mastodon.social
    Apr 29, 2026, 10:29 AM

    @barometz @futurebird
    Also 'it's hard' is not why they didn't do it.

    They didn't do it because they always wanted to tell people 'yes.' Because they knew just bullshitting people would get more engagement than 'can't do that.'

    💬 1🔄 0⭐ 5
  • RalfMaximusralfmaximus@mastodon.social
    Apr 29, 2026, 1:14 PM

    @FeralRobots @barometz @futurebird

    But this stuff USED to work! I remember regularly asking my phone to "set a timer" and it'd just do that. Only a few years ago.

    After they implemented Gemini on my phone and forced a shift to the "AI assistant" it stopped working.

    💬 0🔄 1⭐ 1
  • Lealeadore@sunny.garden
    Apr 29, 2026, 8:08 PM

    @barometz @futurebird

    IME it's not even possible to have an "I don't know" aspect to these things because an LLM is unable to know what it knows or what it can and can't do.

    Any time someone gets an "I can't do that" response is because the company specifically programmed a custom-made filter to catch and ignore certain prompts that they don't want it to answer. Whack-a-mole style.

    💬 0🔄 0⭐ 0
  • FoolishOwlfoolishowl@social.coop
    Apr 30, 2026, 3:55 PM

    @futurebird One of the things that puzzles me is that it seems like they often seem to have LLMs do tasks directly instead of spinning off a separate, specialized tool that's purpose built for the task, like a timer or a calculator.

    Not only is that how software in general works, but it would also resemble more current models of how the mind works -- which are probably influenced by modern software.

    I'd had the impression that "agentic AI" worked somewhat like that. Which makes it more puzzling.

    There's this bizarrely regressive quality to techbro thinking, where they seem to revert to ideas of Europe's early modern period, like naive empiricism, the mind as a blank slate, and so on.

    💬 0🔄 0⭐ 0