Hacker News Reader

I'm not here to argue about Chomsky's basic claims (which I do not find useful), but I am here specifically to argue against the idea that children easily learn language with somehow way less data than ChatGPT. Those who claim otherwise simply have not spent a couple years with a baby.

View on HN

GPT3's input is something like 45TB of data. It's a significant fraction of all text humans ever produced, to the point researches are concerned we may run out of text within the next few years.

Normal healthy language exposure for children is around 100KB/day, we can round up to 40MB/year.

No, I have not spent ~1000000 years around a baby.

ChatGPT is trained on some healthy fraction of the internet. You're arguing that the amount of information a baby consumes to learn language is comparable to the size of the internet?

A baby almost certainly consumes far more data before it speaks than has been fed into GPT-3, yes.

Reply on news.ycombinator.com