What is the difference between AI and robotics?
In my head, they are rather vaguely separated; “A Thousand Brains” pointed to some cultural differences (apparently they talk to each other less than you might think i.e. the concept of ‘reference frames’ is currently limited to robotics), but in my head they’re under the umbrella of ‘artifical machines’. And here I mean artificial as in man-made, engineered. And machines in the more cybernetics sense, not the steampunk/top-down sense.
Currently, AI points to the field of neural networks and LLMs; machines that take in digital data and spit out digital data. Robotics may also take in digital data, but perhaps a notable distinction is that they interact with the physical world.
I think you can go pretty far with just the abstract (says someone who generally prefers reading about something than doing it), but I’ve noticed that there are times where intuitive/experiencial data is much richer. I have not delineated where those contexts tend to be yet. While I think I also ‘hallucinate’ like AI, I also have a subconscioius system that interfaces with physical world data and moves, which theoretically ‘grounds’ my hallucinations into the more useful ones (can some hallucinatory conditions be mechanistically explained by a decoupling with physical world data?).
What are the limitations of AI? Perhaps it parallels the limitations of a computational biologist – you get data, and your value is in your ability to extract signals and make models/hypothesis. You can make a predictive model, like an epigenetic clock. You can ask for more data to refine your working models (it can’t do that yet, but I don’t see why not – how do you ask good questions? To make a guess of what is absent?), but you are in the end limited in knowledge and ability of how that data is obtained/created.
Reminiscent of Judea Pearl’s argument in “The Book of Why”, it is hard to identify causal models amidst the vast possibility of statistical models if you can’t ‘do’. And although he argues you can try to infer causlity through observational data (his do-operators), I suspect it can only compensate so much. Toddlers get so much more ‘do’ data.
I was at first inclined to write AI versus robotics, but perhaps it should be AI and robotics – maybe coarsely mapping to the prefrontal cortex, and the rest of the brain.
Anyway, this is also one of those things not yet fleshed out, but just to note for where my thinking is at this time.