It might be reasonable to call AlphaZero “goal-seeking”. The DeepMind people are actually working on stuff that can cope with diverse environments and pursue open-ended strategies.
That’s not what anyone means by “AI” in this context.
You can give an LLM a goal in the prompt, and set of possible actions which can be taken by calling APIs, and a state which can be called via API, and ask it what to do, then do it, in a loop.
I've done it. It's trivial. It's far from trivial to get it to do something especially sophisticated, but... it's not hard to see a couple generations away it will be possible to do some real damage.
(I want to course correct the thread back to an inquiry frame) .. very relevant that there is not one definition of "AI" at all, and certainly many flavors and architectures of setups right now that all fall under an umbrella term of "AI" -- agree
DeepMind is a peculiar example, since they started by solving video games explicitly (very narrow and goal oriented)
Yeah and there is no shortage of interview footage where Hassabis made a clear, compelling argument around why Atari was a better starting point for the long road to true digital intelligence than click prediction.
There are a number of reasons why I knew OpenAI was a scam including the fact that a metric fuck ton of my former colleagues work there and they all skew a way, but the biggest reason is that I took one look at the Instruct paper and I was like “differentiable loss function on getting humans to click: feed ranking on crack. they’ll make a lot of money but it’ll never work out”.
That’s not what anyone means by “AI” in this context.