I'm fully in agreement on this one. Having other modes of data to learn from, i.e. physical sensations, certainly helps improve learning and understanding, but

I'm fully in agreement on this one. Having other modes of data to learn from, i.e. physical sensations, certainly helps improve learning and understanding, but I don't think it's by any means required for understanding. In fact, the example referred to in that thread of a language model being unable to understand how to construct a weapon to fend off an angry bear, I would argue, is entirely untrue. If I remember, when I get home, I will pose the scenario to GPT-3 and see what it says; but from my experience with it, I give it a more-than-90% chance of providing coherent and useful output comparable to a person. (Because honestly, I've seen it handle even more complex questions with ease.) But yeah, I also fully agree that consciousness, sentience, etc. are not very interesting for an AI to have. I mean, they're kind of the goal, but they're also nebulous concepts that tend to define a continuum, when they define anything concrete at all. I'd argue AIs have had those properties for years, if not decades, and we're just approaching the level of those properties comparable to an average human. BTW, one of my favorite examples of "actually, this is more like human learning than you think" comes from a famous French word problem for students (often misreported as being Chinese). People say LLMs are just predicting the most likely next tokens of language, which is true, but they often fail to realize how much humans do the same thing when we learn. They point to the failure cases of LLMs being asked questions that have no answer (either because they're nonsense, or because they don't have enough information) and giving ridiculous answers anyway, and claim people don't think that way. So the French boat problem goes like this: "If a ship had 26 sheep and 10 goats onboard, how old is the ship's captain?" Obviously, the correct answer is, "I don't know, there's not enough information to determine this." But when posed on an exam for students, the majority of them wrote a concrete answer, often some arithmetic manipulation of the numerical symbols (i.e. numbers) in the problem, even though there's no correlation between those numbers and the age of the captain. They did exactly the same thing the LLMs do. What's even more telling, if you prompt an LLM with explicit instructions to say when it doesn't know an answer, it performs much better at these kinds of ignorance-recognition questions, by saying it doesn't know when it can't know. In other words: humans have very similar failure cases to these LLMs, but we point to the LLMs' failures as evidence they're not thinking like humans. Which is wishful thinking, I'd say.