Hacker News new | past | comments | ask | show | jobs | submit
An entire generation of web developers learned by viewing the source of webpages they visited at a time when HTML and JavaScript were still nascent. I don't think setting the bar for "massive" is the way to win this argument. It's valid because the biggest difference between an LLM reading Stack Overflow and remembering content and me finding answers there and learning something is that the LLM can do it much, much faster.
The difference is, just as a child doesn't need to consume the entirety of a language before being able to communicate with said language, a human developer doesn't need to consume all of Stack Overflow, Open source projects etc. in order to start producing functional work.
I'm confused - I thought you said the response was child-like, not that a child can do the work. I'm not advocating that an LLM has child's level of experience, just that humans learns how to do things from looking at previously done work, so why is that not an allowed method for an artificial intelligence? How else is it supposed to learn?

This also ignores all of the great artists who learned by sitting at the feet of other great artists. Sometimes they extend the art of their teachers, sometimes they diverge.

A child or a human developer cant do all the thousands of different things an LLM can do either.
loading story #50508600
So if future models are able to generalize better from much smaller amounts of data, you think that all the current concerns disappear? It doesn’t seem like a very good argument if it’s purely based on the amount of data needed to train a model.

What matters are things like where the data comes from, how similar the output is to the input, how close the result is to existing (copyrighted) work, and so on.

Simply using copyrighted work, even all copyrighted work, to train a model doesn’t, by itself, trigger any legal consequences. In that respect at least, the comment you called “child like” was correct.

loading story #50508698