That's very different and specific from "anything that requires manual dexterity and thinking on the fly"
Yup, and robotic surgery has been around for a while. The really cool thing though is LLMs are actually good at modeling (or assisting to) some tasks requiring dexterity, I'm not sure that would have been predicted a few years ago. Obviously using deep learning to control a robot was an obvious step but using a language model... well it's got an internal world model and it helps.
edit - to be clear it's the language model writing the code and its awareness of the world (which it only knows of via text) that's impressive, it's inherent world model has enabled this sim to reality:
"I’m excited to announce DrEureka, an LLM agent that writes code to train robot skills in simulation, and writes more code to bridge the difficult simulation-reality gap. It fully automates the pipeline from new skill learning to real-world deployment. "
Like this is an insanely cool demo from Nvidia/UPenn - trained in sim and it works right away in the real world too:
If we leave AI to be completely autonomous and make decisions for itself, we'll end up in the Matrix or with Skynet, but nobody is suggesting that, it's obvious there will still be "some" humans in work making decisions/orders for the AI
AI in video games is also not true AI like we're getting here -
https://en.wikipedia.org/wiki/Artificial_intelligence_in_video_games it doesn't learn, much like current tech like autopilot doesn't learn, or robots in a car factory don't learn
Yeah, there's still got to be human involvement in most things at some level. It depends on the AI to some extent, some of RL overlaps a bit with control systems and AI agents monitoring stuff, controlling things, carrying out basic repeated tasks is already a thing. For example, I saw a presentation by a vertical farm few years ago and lots of it was controlled by RL agents monitoring and optimising the growing conditions. Obviously that's not going to turn into skynet, you can run basic agents on a pi or something.
Re: Games AFAIK classic ML/AI
typically isn't learning (updating parameters) when deployed for inference tasks such as in games but there is no reason why it couldn't be in some contexts, it's called online learning or online machine learning - certainly can be used to adapt the behaviour of NPCs or adjust things to a players skill level. Typically lots of "AI" has been quite basic though.
Beyond the classic stuff for sure; if deploying LLMs the underlying model isn't being updated via each interaction, though I guess there is some scope for "learning" in so far as there is a context window (and also a developer could perhaps try some other tricks like saving some details of earlier interactions beyond a context window, maybe updating some external parameters that are then fed to the LLM at the start of a new interaction etc.. so in a sense some degree of "learning" can occur).
