Hacker News Reader

There's been a lot of discussion in these comments regarding the author's definition of intelligence. What I haven't seen people address is the author's correct (in my opinion) criticism in the amount of data needed to train these LLMs versus what the human brain of a child is exposed to. Unless of course part of a child's knowledge is passed down through DNA/RNA/...

View on HN

It's true, and it is a weakness of LMs relative to humans, but it's kind of obviously true and something that's being actively pursued as a topic of research, so it's not a remarkable criticism.

It's also kinda apples-to-oranges - a human child also has thousands of hours of other rich sensory input. Multi-modal models are not as well-developed as LMs.

Reply on news.ycombinator.com