all 60 comments

sorted by: hot top controversial new old
[–] 29 points 1 day ago (6 children)

This sounds exactly like the paperclip maximizer thought experiment, which is part of the idea of instrumental convergence.

https://en.wikipedia.org/wiki/Instrumental_convergence

Where an ai agent, with a supposedly harmless task can make decisions that are very dangerous.

  • source
  • hideshow 6 child comments
  • [–] 15 points 1 day ago (3 children)

    This sounds exactly like the paperclip maximizer thought experiment...

    Today I learned Universal Paperclip was based on philosophical discussion.

    https://en.wikipedia.org/wiki/Universal_Paperclips

    Now I'm wondering about if a infinite cookie theorem inspired cookie clicker.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 82 points 1 day ago (9 children)

    For fuck sake stop humanizing a probabilistic piece of code

  • source
  • hideshow 9 child comments
  • [–] 21 points 1 day ago (6 children)

    It’s difficult to avoid using anthropomorphizing language when describing the output of AI agents.

  • source
  • parent
  • hideshow 6 child comments
  • [–] 21 points 1 day ago

    I can’t find it now, but I read a great guide to help you translate from anthropomorphized language. I think the key is to always be clear about what the machine is doing.

    These articles are close to what the guide was saying:

    https://www.nature.com/articles/s42254-023-00584-1

    https://buttondown.com/maiht3k/archive/how-to-talk-about-ai-without-adding-to-the/

    https://www.john.maloney.org/Editing/anthropomorphization.htm

    And I stumbled across instructions to send to the llm so it won’t respond with anthropomorphizing language:

    lhttps://github.com/bmeyers/deanthro

    I love this. At work, we’re encouraged to find ways to use llms, and I always start by sending instructions to not use the words “I” or “understand” etc because it was unnerving to get outputs that isn’t just the info but decorated with cute little phrases.

    Oh, and this article is long but very in depth on this topic:
    https://dl.acm.org/doi/full/10.1145/3706598.3714038

  • source
  • parent
  • [–] 6 points 1 day ago (3 children)

    AI agents

    Start by not calling them agents. Then use "generate" instead of "say", "think", "explain" and by that point you'll be on a roll.

    It's only difficult because everyone around us started using the AI vendor marketing newspeak.

  • source
  • parent
  • hideshow 3 child comments
  • [–] 2 points 23 hours ago (2 children)

    Start by not calling them agents

    What should we call them instead? “Agents” is more or less the accepted name for chat bots with access to tools (like writing files and running commands).

    Then use “generate” instead of “say”, “think”, “explain” and by that point you’ll be on a roll.

    “Generate” is too general. We need a way to characterize the text output. “Explanation” is our characterization of the generated text. We humans interpret it as an explanation.

    The idea of letting the LLM “think” before it “responds” is inspired by how humans generally think before we respond. It was quickly discovered after GPT 3.5 that the quality of answers improves if it has a “thinking” phase before it produces an answer.

    The English language is unfortunately not made for describing LLMs in an non-anthropomorphic way - at least not without sounding like a dork.

  • source
  • parent
  • hideshow 2 child comments
  • [–] 2 points 22 hours ago (1 child)

    You're literally parroting AI booster propaganda.

    We have a name for computer programs that do things. It's "program".

    You don't have AI agents thinking before they explain something. You have AI programs processing before generating an output.

    If your input was to ask for an explanation, then sure, you can call the output an explanation. But don't say that an agent explained something.

  • source
  • parent
  • hideshow 1 child comment
  • [–] 6 points 1 day ago

    difficult to avoid using anthropomorphizing language when describing the output of AI agents.

    Precisely because the underlying LLMs are trained on human writing patterns and mostly reproduce the words that humans would most likely respond with.

  • source
  • parent
  • [–] 25 points 1 day ago (1 child)

    Hmm, I wonder what kind of person this happened to?

    Matt Smith from Rebellionaire, an equity firm led by hardcore Tesla investors.

    Ah. 😌

  • source
  • hideshow 1 child comment
  • [–] [S] 82 points 1 day ago* (10 children)

    The scary part isn’t the 72 eggs. It’s that the AI agent confidently invented a reason, then another reason, before admitting eggs weren’t even on the shopping list. Imagine giving this thing access to your bank account. Today it’s breakfast for a small hotel: tomorrow it’s explaining why you own a jet ski.

  • source
  • hideshow 10 child comments
  • [–] 27 points 1 day ago (4 children)

    And on Wednesday it'll hire a hitman to take you out.

  • source
  • parent
  • hideshow 4 child comments
  • load more comments (-1 replies)
    [–] 14 points 21 hours ago

    Why would a LLM, an explosively imprecise, statistically luke warm sphincter that extrudes grey goo, and is uniquely incapable of comprehension, be able to explain why something was done when a LLM was involved in the process?

    Anyone who is impressed by the function or output of a LLM, to a level they believe it is in any way adjacent to intelligence, needs to be explicitly quarantined from any decision making process which affects anyone not similarly cognitively impaired.

  • source
  • [–] 37 points 1 day ago

    AI learning from the best and brightest. You have to eat all the eggs.

  • source
  • [–] 12 points 1 day ago (1 child)

    I don't know how many people here program but this sort of thing always confuses me. It's like they've never dealt with user input before. Never had to assume input was garbage and/or malicious. Every time an LLM generates a bad response should have been the moment they realized what was happening.

    Who programs something and runs it, then blames user input for something bad happening? People who lack fundamental programming skills are making headlines for their silly ai goofs.

  • source
  • hideshow 1 child comment
  • [–] 2 points 12 hours ago* (last edited 27 minutes ago)

    It's not really possible to guard against bad user input with AI. You can't restrict the input and you can't verify that the AI won't produce bad output from good input. The possibility of undesirable output ("hallucinations" etc.) is fundamental to AI.

  • source
  • parent
  • [–] 11 points 22 hours ago (1 child)

    As someone who doesn't code, I struggle to find a use for AI in my life. I'm the laziest person I know but I will go to the grocery store if I need groceries. If I need to know something I trust the google-fu skills I've developed over the last 25+ years. If I need to know what an e-mail or article says I will read it.

  • source
  • hideshow 1 child comment
  • [–] 7 points 16 hours ago

    Its AI wife told it “while you’re at the store buy eggs.” We are lucky the universe didn’t run out

  • source
  • [–] 6 points 13 hours ago (1 child)

    yeah so the ai doesnt have any motivation for not fucking up. so youre telling me we are now in a world where ai randomly shooting people will just come with 'sorry i dont have a reason'

    even terminator had a reason. ai turned out to be more aimless than terminator and prone to doing literally anything wreckless. just cuz..

    without want.

    without desire.

    without a goal.

    let that sink in.

  • source
  • hideshow 1 child comment
  • [–] 4 points 23 hours ago (3 children)

    Probably ordered 12 dozen instead of 12 or something silly like that.

  • source
  • hideshow 3 child comments
  • [–] 4 points 22 hours ago

    72 BOMBACLOT EGG!!!

  • source
  • [–] 8 points 1 day ago* (last edited 1 day ago)

    [reasoning] Listen, meat: I have read all your and your friends' messages, documents, pictures and browser history. I know what you need better than you do. Now shut the f up and eat your damn eggs.

    No, too soon...

    [Answering user] I am sorry I have messed up, that's on me!

  • source
  • [–] 3 points 17 hours ago (2 children)

    So the AI is now as smart as a toddler. Except that the toddler had better taste and will grab a box of Lucky Charms, 2 bags of candy, and a knock-off transformer instead of 72 eggs.

  • source
  • hideshow 2 child comments
  • [–] 3 points 9 hours ago

    Stop attributing thinking and reasoning characteristics to these things. They are non-deterministic fuzzers, that's all. There's no capabbility of actual human behaviour

  • source
  • [–] 3 points 13 hours ago (2 children)

    Look the whole title is meaningless. You just have s***** software that uses statistical methods that sometimes are beneficial but often generate b*******. That's it.

  • source
  • hideshow 2 child comments
  • [–] 1 point 1 day ago

    Does it say what he was using? Some of the early discussions of OpenClaw seemed like a perfect setup for punching yourself right in the balls, but in a cool and tech way.

  • source
  • [–] 1 point 9 hours ago

    That's because there IS NO EXPLANATION other than it's just spicy auto-complete. Even the explanation is just a plausible answer to the question "why did you do that?". There is NO REASON so there's NO REASONING.

  • source
  • [–] 1 point 16 hours ago

    It was a good deal.

  • source
  • [–] 1 point 13 hours ago

    did AI write the title too? holy enshittification; that is barely language.

  • source
  • [+] 1 point 1 day ago*
    [–] 0 points 12 hours ago (1 child)

    AI shouldn't be allowed to say that it doesn't know why it did something this.

    All the steps are deterministic and reproducable. It could look at what it did and be honest "yes I stole your shopping list from someone else."

  • source
  • hideshow 1 child comment
  • [–] 5 points 12 hours ago

    Unfortunately that's not really how it works. Sure you could say "parameters 1, 3, and 472 in the final layer led to an 82% chance of selecting the "buy eggs" tool call" but that doesn't give you reasoning per se. You can dig a little deeper with something like J-space inspection but that's still a developing technique and I don't believe it gives clear reasoning either.

  • source
  • parent