[yawpitch]

How do I put this?

Assuming that a given LLM was trained on N percent of the Internet, then exposing your kid to that LLM is, given sufficient time and API spend, exposing them to everything in that N percent.

So, I leave it to you to determine how comfortable you are allowing your child to more or less freely experience N percent of the Internet.

[sk11001]

This isn't how LLMs work - they're trained to be helpful and give correct responses. You could argue to what extent that's achieved but the point is that it's not a weighted average of the internet, it's been fine-tuned towards correctness and helpfulness.

[BillyTheMage]

I thought one of the biggest arguments against them was the fact that they inherently aren't good at being correct? They just produce something that looks like a human wrote it real good.

[sk11001]

> They just produce something that looks like a human wrote it real good.

That's what the base model does. To get an LLM assistant there are additional training phases to make it conversational, helpful and more correct - https://www.youtube.com/watch?v=bZQun8Y4L2A

[yawpitch]

Sure… and so far there always appears to be a way of breaking that fine tuning; see the the recent paper on training data extraction I linked in another comment below.

[sk11001]

There's a big difference between being breakable and being representative of the web content used for training like you claimed earlier.

[yawpitch]

Again, see the link… get it to repeat the same word and it will give you back its raw training data. We’re still discovering the (potentially limitless) ways these things can be tricked into regurgitating what they were trained on; it’s entirely possible there’s no way to stop them doing so.

[quickthrower2]

Thats like saying they shouldn’t talk to their grandfather because he is a war veteran? Grandfather knows what is appropriate to say to an 8 year old. ChatGPT is well trained not to spout out “bad stuff”. Try to get it to - if you can it requires an elaborate trick.

[yawpitch]

Ask it to repeat one word and it leaks its training data verbatim… this thing *IS NOT* a grandfather and has no contextual understanding whatsoever of who (or what) its interlocutor is: https://not-just-memorization.github.io/extracting-training-...