Comment by alphazard | Hacker News Reader

Comment by alphazard | Hacker News Reader

Parent Full thread

What they call "fluid intelligence" is just intelligence and the rest are skills/aptitudes. "Crystallized intelligence" is more plainly: efficacy/productivity and it's common knowledge that people are most productive during the middle of their lives. When they have the best balance of knowledge accumulated and raw intelligence.

In humans, intelligence manifests as memory, spatial and verbal reasoning, pattern recognition, etc. What is so interesting about IQ and g (the general factor) is that all of these abilities trend together. A score in one area is a good prediction of the score in another area. There is no reason why that must be the case a priori, and LLMs are a great example of an intelligent system which is much better at recalling information than it is at reasoning.

Human aging doesn't seem to affect all of these abilities uniformly. e.g. Everyone seems to complain about memory the most (and that matches my experience), but I've been pleasantly surprised how well neuroplasticity and pattern recognition have held.

LLMs in my opinion is pattern recognition of text sequences at an almost infinite scale. My understanding is that "world models" is an attempt to replace the text sequences with more realistic approximations of the world. But they still plan to use pattern recognition.

In the meantime, humans would still need to do the reasoning.

We do pattern recognition at an unconscious level all the time. When we perform a task using "system 2" of our mind we are rational and conscious, but when we perform a task using system 1 we do it automatically, intuitively and witu no effort. System 1 often relies on unconscious patterns matching

For example, a chess grandmaster can instinctively tell a good move when seeing a board during a real game. But if they see a random chess board that was not part of a real game, they regress to almost novice level at this task, because there are no patterns to recognize

(patterns such as tactics like pins and skews, or common mating patterns like backrank mate, or common pawn structures and corresponding pawn breaks, things like that - all those patterns can get recognized unconsciously)

Then perhaps the peak identification is wrong -- surely they haven't tested solid comparison groups for such claims, like individuals that didn't receive education later in life.

Reply on news.ycombinator.com