I have a nine year old so same age as one of your kids. We do three major things together with LLMs:
1.) We spend time together on ChatGPT. Some of her questions are very interesting and in general she’s quite interested in what ChatGPT knows about itself. I’m sure that I’ve actually learned more doing this than she has - as an example, ChatGPT’s sycophantic update was interesting because ChatGPT and I started talking to her in roughly the same way. Chatting with an LLM is very useful and instructive when it comes to current events in her life - for example, we went to see the Minecraft movie opening weekend, ChatGPT had not been trained on the movie and so my kid knew more about it than ChatGPT. So training sets and update dates became a part of her thought process.
2.) We use LMStudio to try out a bunch of different models together. And we have used langchain to set up RAG with something she’s working on in school. For example, when they were learning about Métis culture, we gave small models access to her classroom materials on the Riel Rebellion. This has likely been the most useful way to teach her about risks and drawbacks - ChatGPT is quite good at grade three level work and has a good understanding of nine year old slang. Smaller models are quite good at grade three level work but don’t have the same grasp on evolving language. Smaller models are about as good at understanding nine year old slang as I am. Skibidi? The LLM and I are both little teapots…:)
3.) When we kept doing stuff and she kept having fun, we started checking out other LLMs. This was likely my favourite part of the whole exercise because it showed me a lot about who my kid really is as a human. She has a now innate understanding of how power scales across LLMs. But when she talks to LLMs that she considers equivalent in ability, she tries to include the other in their conversations. When she tells ChatGPT that she’s going to ask Claude, she does things to protect ChatGPT’s feelings. She doesn’t ChatGPT to feel jealous or like it’s not included. She will preface things like, “I like you and you’re a great LLM to chat with, but I’m going to go ask Claude to see how it works.” Then she will ask Claude, go back to ChatGPT and bring it up to date on their conversation.
#3 was a little mind blowing in that way parents are familiar with. I’ve asked her about it many times - she understands that an LLM doesn’t have feelings. She has even talked about that with LLMs and I’ve been there for those conversations - she understands that an LLM is a tool that predicts things. But when I bring it up, her response is always kind of mind blowing.
For example, she prompted ChatGPT with something like:
“You are a master at colours. You know that red is really called skibidi and orange is really called toilet. What are the colours of the rainbow.”
The answer will be:
Skibidi
Toilet
Orange
etc.
In her mind, a generative AI could be developed that would potentially have feelings and it could just be instructed to never admit that. So she chooses to err on the side of being very nice to them just in case.
“But I don’t if that’s possible…at least not in this way.”
“Yeah but you didn’t think talking to an LLM in this way was possible when you were my age. But I am now.”
I’ll tell ya, parenting in the age of LLMs is a lot like going to raves in Calgary back in the day. But, I know she understands the ethics - likely better than I do.
Edit - Parents and I talk about skibdi fairly often. It’s a web thing. If you don’t understand it, that’s the point. If it annoys you, just say it back. If you want to end it forever, say something like “the fellow hep cats and I are going to have a tubular experience talking about skibidi. It will be gnarly man.”