[albatross79]

The heart of the matter is a system that produces correct output in a wide variety of scenarios, including ones outside our scope of experience or understanding. That would be a one definition of AI. LLMs are not that, and I don't understand the desperate insistence that they are. Just accept what it is. A good auto complete is useful, a crappy AI is a marketing scam.

[logicprog]

> a system that produces correct output in a wide variety of scenarios, including ones outside our scope of experience or understanding

Where does this definition come from? I certainly don't agree with it, and I am not sure who does, besides yourself and Marcus. Also it seems that you're saying AI does, in fact, mean 'perfect' AI, basically?

> A good auto complete is useful, a crappy AI is a marketing scam.

LLM agents do a lot more than auto complete now, and using them less like auto complete and more like 'AI' (via agents) has actually made them more useful and less crappy! Also, I don't think framing how modern RLVR'd LLMs operate now as as auto complete even makes a whole lot of sense in the first place.

[albatross79]

It's a common sense definition that covers the broad features of AI as most people would have imagined it before LLMs came along. If you have a better one I'm open to it.

Also, agents are just LLMs with harnessing. Attaching a plow to a horse makes it more useful too, but it doesn't change the intrinsic nature of a horse.