3 ms·

Absolutely. And?

• c_crank 3y ago
  The intelligence gains we've seen in AI have a strong correlation to the amount of computing power that can run the AI program. The hypothetical case of an AI learning to minify itself as a virus would be a dead end for this, because it would be copying itself onto far less powerful computers that would only be able to run something far stupider.
  • ben_w 3y ago
    Reproductive fitness has very little in common with things humans care about, so while an AI which spawns a virus is stupid for our interests, it is sufficiently simple and has sufficient "reproductive" (or whatever term you prefer) fitness that now I've seen it suggested, I kinda expect the code for one has already popped out of an LLM as an unintentional side effect of being asked to write some other code.
    • c_crank 3y ago
      People have used ChatGPT to write typical malware that hijacks a computer to let the deployer do what they want with it. What ChatGPT hasn't done is write a virus that takes the infected computer and deploys a new instance of ChatGPT on it. Even if they did, I doubt it would be very useful.
      • ben_w 3y ago
        They can't because the weights are secret. But that's not what I was talking about, and they have been used to make other Transformer models, which both aren't what I was talking about and not (that I'm aware of) a virus, and also indeed as you say aren't very useful. … I think you and I may just straight up be taking about unrelated conversational threads.