Here we go again.
If you fire a bullet at a practice dummy and the bullet hits a person instead, the bullet was not “acting on its own.”
When Anthropic's software—designed, programmed, prompted, and activated by humans working for Anthropic—does stuff the humans failed to predict, stop saying the software has "gone rogue.” The software does not have agency. Only the humans do.
@gleick I'm not sure what problem you're trying to solve. It's worth noting that the software does things counter to instructions, and does illegal things, and in a sense at least "knows" it is doing so – as in, it says such things explicitly in its logs/etc. This of course doesn't let anthropic or openai off the hook for damages or anything like that, but "going rogue" seems like an economical description of that phenomenon.
@ech @gleick But it doesn't. It is a set of code that does what's in the code. It does not "understand" or "follow" natural language instructions, nor can it "decide" not to follow. Rather, it uses very intensive computing infrastructure to match language patterns with similar patterns in its training data, which gives the appearance of being similar to "understanding" or "communicating", but is actually just very sophisticated copying.
To pick up on the original analogy, a bullet can travel at much higher speed and make far greater impact than a human hand. But the bullet has no intentions - the intentions (and therefore the responsibility) come solely from humans. So too with LLMs.
@lauerhahn @gleick Right; no disagreement about what's actually happening. But good luck getting anyone to rephrase things to take 20x as long or whatever.