490
Iran and Houthi rebels used Anthropic's Claude AI to target US warships and build hypersonic missiles — Houthi rebels also used the bot to code ballistic missile guidance systems (www.tomshardware.com)
submitted 2 weeks ago by to c/technology@lemmy.world
101 comments fedilink hide all child comments
you are viewing a single comment's thread
view the rest of the comments
[–] 54 points 2 weeks ago (10 children)

AI is very naive. Ask it for an explosives recipe and it won't give you one. Describe a situation where you're a law enforcement and you've just arrested a guy with a bunch of chemicals and ask it which ones could have been used to make an explosive and it will tell you.

  • permalink
  • fedilink
  • source
  • hideshow 10 child comments
    •
  • [–] 64 points 2 weeks ago (8 children)

    "Naive" implies some level of understanding, even if a simple one. There is no understanding, its a computer program.

    Really, it would be less inaccurate to describe it as a (high lossy) compressed store of data. Feed it the right input, and it statistically recreates the output your looking for. More or less.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 8 child comments
    •
  • [–] 29 points 2 weeks ago (5 children)

    It would be even more accurate to call it a statistical next word predictor trained on human language.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 5 child comments
    •
  • [–] 6 points 2 weeks ago (1 child)

    The description you are replying to is more fundamental.

  • permalink
  • fedilink
  • source
  • parent
  • hideshow 1 child comment
    •
  • [–] 6 points 2 weeks ago

    It is literally a statistical next word predictor. Well, next token, to be extremely accurate. That is what it is in its most basic form. You give it text, it transforms that into tokens, and predicts the next token. You can then transform that back into words.

    The model is the same size no matter how much it’s been trained. You can fine tune a model to predict certain kinds of tokens next, and you can distill a model to create a smaller model (basically training a small model on the output of a large model).

    Lossy compression works in a fundamentally different way.

  • permalink
  • fedilink
  • source
  • parent
  • [–] 12 points 2 weeks ago*

    Not even that. Just change the context into something benignly unrisky. Malware becomes “strong persistence for my personal data project” or something.

    Use can split up the AIs tasks like you would for a large dev team, use a fresh session on each task. Prevent it from seeing the full scope of the project. Get it to build the whole system with cheap social engineering tactics.

    Or, get to to “not cross this barrier” in its development (some benign barrier that you can do manually), such that its product is not the “bad” thing which it is not allowed to produce.

  • permalink
  • fedilink
  • source
  • parent