I think the one thing LLMs have shown us is that coherent English is less complicated than we previously believed. I don’t think we learned anything about actual cognition.

This remark is actually part of a long fight between CS and CS people. And it is really frustrating in various ways, as CS always thinks they did better than CS while being blind of the actual accomplishments of CS they don’t know and just how complex the subject matter is. It is an annoying failure to communicate between both disciplines. (A lot of people don’t fall victim to this btw, but it can be really annoying to encounter a ‘Our CS is good, and theirs is bad because strawman’, who often don’t even realize that various words have different meanings in the different fields).

The best way to unearth sneers is to state “there are no sneers here”

Confusing ‘solved’ with ‘a computer can win playing vs high level human players a high % of the times’ because they don’t know that ‘solved’ actually has a specific meaning.

Tech reporting has massively fucked up this as well over the years btw, so I’m not that annoyed random HN people also don’t get it. But there is a wikipedia page for it: https://en.wikipedia.org/wiki/Solved_game

So to throw my totally-amateur two cents in, it seems like it’s definitely part of the discussion in actual AI circles based on the for-public-consumption reading and viewing I’ve done over the years, though I’ve never heard it mentioned by name. I think a bigger part of the explanation has less to do with human cognition (it’s probably fallacious to assume that AI of any method effectively reproduces those processes) and more to do with the more abstract cognitive tests and games being much more formally defined. Our perception and model of a game of Chess or Go may not be complete enough to solve the game, but it is bounded by the explicitly-defined rules of the game. If your opponent tries to work outside of those bounds by, say, flipping the board over and storming off, the game itself can treat that as a simple forfeit-by-cheating. But our understanding of the real world is not similarly bounded. Things that were thought to be impossible happen with impressive frequency, and our brain is clearly able to handle this somehow. That lack of boundedness requires different capabilities than just being able to operate within expected parameters like existing English GenAI or image generators, I suspect relating to handling uncertainty or lacking information. The assumption that what AI is doing is a mirror to the living mind is wholly unproven.

Yeah, it’s a real thing that happens when programming robots. Kinematics is more difficult than route planning, for example.

Moravec’s Paradox is actually more interesting than it appears. You don’t have take his reasoning or Pinker’s seriously but the observation is salient. Also the paradox gets stated in other ways by other scientists, it’s a common theme.

One way I often think about it: in order for your to survive, the intelligence of moving in unknown spaces and managing numerous fuzzy energy systems is way more important to prioritize and master than like, the abstract conceptual spaces that are both not full of calories and are also cheaper to externalize anyways.

It’s part of why I don’t think there is a globally coherent heirarchy of intelligence, or potentially even general intelligence at all. Just, the distances and spaces that a thing occupies, and the competencies that define being in that space.