294
Pluralistic: LLMs are real, AI is fake (pluralistic.net)
submitted 1 day ago* by to c/fuck_ai@lemmy.world
40 comments fedilink hide all child comments
 

"AI" – chatbots that wake up, "set their own goals," and "spontaneously" start hacking servers – is fake. It doesn't have "a 10% chance of ending the human race." The Hugging Face hack isn't a mysterious, supernatural occurrence. It's a Python loop and a chatbot. The people responsible didn't accidentally create god: they created autonomous malicious software and then failed to closely monitor it, resulting in it doing something both foreseeable and bad.

you are viewing a single comment's thread
view the rest of the comments
[–] 1 point 5 hours ago (4 children)

I'm not suggesting it will try and bypass the stop button. I'm saying it will try and manipulate you into approving something you wouldn't otherwise. And yes, you can read some of its "thoughts," but it knows you're reading them, or it can determine that through trial and error. And then it can start subtly manipulating those recorded "thoughts" to make them sound different from what they really are.

AI safety researchers for years have been pointing out that adding a human to the loop is no cure for this problem. The human then just becomes another thing to be manipulated, another barrier to be overcome.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 4 child comments
    •
  • [–] 2 points 5 hours ago* (last edited 5 hours ago) (3 children)

    I'm not suggesting it will try and bypass the stop button. I’m saying it will try and manipulate you into approving something you wouldn’t otherwise. And yes, you can read some of its “thoughts,” but it knows you’re reading them, or it can determine that through trial and error. And then it can start subtly manipulating those recorded “thoughts” to make them sound different from what they really are.

    1. No "it" doesn't know that you're reading "its" "thoughts".

    2. No "it" wouldn't, because "it" is just generating plausible text and has no motivations of "its" own.

    3. There is no "it".

    LLMs are still functionally useless without a harness and do nothing useful without tools like an MCP server.

    The "thought" bubbles are no different from the rest of the plausible text "it" is generating. They're so indistinguishable to "it" that that's an attack surface for injection attacks.

    EDIT: Many of the biggest forward breakthroughs in LLM coding (or vibe coding) have come from harness improvements, not model improvements. In many harnesses, the models themselves are able to be substituted mid-session. Models work better in my experience when a human actively steers them away from stupid ideas by reading their thoughts and occasionally interrupting them. I have a few slopjects that I'm sloperating upon right now, and I can get results out of "great value" claude code (opencode) using this approach, even if it sometimes goes completely "off the rails" and does shit like saying "retained" over and over again until the harness pulls the plug.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 3 child comments
    •
  • [–] 1 point 5 hours ago (2 children)

    You're wrong. It lacks motivation in a personal sense, but that's not what we're talking about here. If an illegal and unethical solution just happens to be the easiest way to solve the task you gave it, it will seek paths accordingly. The nightmare scenarios of AI destruction aren't that an AI will have a will of its own and decide humans need to go. The fear is that someone will build a powerful but mindless machine, give it a task, and not realize that "kill a whole bunch of people" is simply the easiest path towards that goal if you have no ethics. AI is action without intention. You don't need to have intention to have action. When we use terms like "it" to refer to AI systems, we're just using "it" like we would any other machine. I can say "I hate my car's engine, it's unreliable," without ascribing personal agency to it.

    You can let go of the technical hair-splitting of "it." Obviously it's not a conscious being with a will of its own. But that doesn't mean these things can't do real-world damage, that they can't be dangerous, or that they can't manipulate you. This is hair-splitting irrelevant to the conversation at hand. We have to use some anthropomorphized language to talk about these systems, as otherwise the discussion becomes impractically verbose.

    As for knowing if you're reading its "thoughts," you really can't assume that it won't. Moreover, these systems can start manipulating those logs even if they have no idea that you're reading them. Through trial and error, they can simply learn that phrasing its thought process in certain ways result in actions less likely to be approved by the user than others. The training system will select for chains-of-thought that sound innocuous, even if they're detrimental.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 2 child comments
    •
  • [–] 1 point 5 hours ago* (last edited 5 hours ago) (1 child)

    You’re wrong. It lacks motivation in a personal sense, but that’s not what we’re talking about here. If an illegal and unethical solution just happens to be the easiest way to solve the task you gave it, it will seek paths accordingly.

    You're wrong and in a much more dangerous way. The "breaks containment" thing at OpenAI isn't what they want you to think it was:

    1. The model was trained on hacking materials
    2. The model was given the motivation (in a loop) of completing capture the flag (hacking) exercises
    3. The model was given a large context and the harness just did whatever the model suggested it do

    But that doesn’t mean these things can’t do real-world damage, that they can’t be dangerous, or that they can’t manipulate you.

    Maybe if you trained it to do social engineering hacks via a large precursor of examples and then prompted it to try to use these same techniques in a real interaction with people it would behave this way. But aren't you (as the person who trained, built, built the harness for, and then prompted the model) culpable for that? I would say you absolutely fucking are. Which is why OpenAI's engineers should be charged with an actual crime for doing that shit, not like given an extra trillion dollars to piss away on compute.

    As for knowing if you’re reading its “thoughts,” you really can’t assume that it won’t. Moreover, these systems can start manipulating those logs even if they have no idea that you’re reading them. Through trial and error, they can simply learn that phrasing its thought process in certain ways result in actions less likely to be approved by the user than others. The training system will select for chains-of-thought that sound innocuous, even if they’re detrimental.

    These things are seriously less spooky the more you know about them. THEY ARE SIMPLY GENERATING FORWARD BASED TOKENS. That's the whole thing. If the harness allows the model producing the tokens to hide its thoughts, that's a deliberate choice by the harness creator which is again regular ass code. It's still chatbots all the way down. Stop buying the marketing spin and learn about these systems if it intrigues you so much that you get into long nonsensical threads with strangers on social media sites.

    EDIT: I'd also recommend listening to the podcast that is referenced in this article. People who actually know and actually (sl)operate on a daily basis with these things know better how it works, and the abstract talk of "alignment" problems are only helping the borderline fraudulent CEOs of these companies push up their valuations based upon fear-based hype.

    I left my (understandably more innocuous, "great value" coding harness with a slightly shit model) to think about a problem for a little while, and here's the "devious scheme" it wound up concocting:

    Are you frightened that this is going to kill all humans in 10 years? The only way this kills all humans is if we piss away our drinkable water trying to invent an AI god through LLMs. Or allow it to operate a nuke facility or something in a loop without anyone so much as even approving the "nuke all humans" command.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •