AI educator Harper Carroll on using AI as an interactive learning tool for children

breaking it down for us. We'll talk to you soon. See you. Bye. Up next, we have Harper Carroll coming into the studio. Welcome to the stream. Hope we are waiting one minute. Let me see. Um, let me tell you about the We already Let's tell you about timeline. Wait, we have the bucket pull back. We We've got Harper.

Oh, we have Harper. Harper. Welcome to the show. Hi, Jillian and John. Thanks so much for having me on the show. Welcome to the show. Uh, would you mind introducing yourself? And there's a bunch of things that I want to talk about, but I'll let you kick it off with an intro on yourself. Okay. Um, I am Harper.

I um am I run Harper Carol AI. So I'm an AI educator. I'm a machine learning engineer. That's my background. I have two degrees in computer science and engineering specializing in AI from Stanford. Then I was at Meta building machine learning systems there for about four years.

Then I was at a GPU environment cloud environments development startup. Uh that where I became head of AI. I actually started teaching AI there and that's how this all kind of was born. But I also TAed when I was at Stanford.

um some core AI PhD courses and yeah that company was acquired by Nvidia and I just teach AI now full-time. I started on X with fine-tuning guides. So you might if you've seen me before on X is probably from that.

I started in late 2023 with like fine-tuning open source model guides with your own data or with hugging face data um you know coding guides Jupyter notebooks and then uh went to Instagram and had have been there and kind of reaching towards YouTube and starting there with Q&As's and and yeah just kind of seeing where this goes.

Yeah. Uh what's been the bigger news in your world? Uh the IMO gold medal uh uh competition between OpenAI and Google or GPT agents or something else in the world of AI? There's a new news story every day. What's been capturing your attention? Yeah, there's a a new news story every single day.

Um I actually saw one recently. What was what were they saying? that reasoning models well the the major AI companies are coming together and and saying that they want to make sure that reasoning models stay open and that we continue to see their thought processes.

So that kind of research I also saw really interesting research from anthropic that said that reasoning models are not actually necessarily showing us what we think that they're thinking.

So for example, Anthropic gave a reasoning model a question and it kept getting it wrong and then they gave it a hint and and the model got it right, but when you looked at the trace the chain of thought from the reasoning model, it didn't mention the hint.

So, we think that we're actually show seeing its reasoning, but not necessarily. So, there needs to be a lot more work on this, but but yeah, some of the cool research stuff and and yeah, there's always new new tech coming out, OpenAI's agent, very exciting stuff for browser.

Do you do you expect uh we've been we've been testing agent uh John was texting uh testing it over the weekend. I was testing it on the show earlier.

Do you expect an explosion of consumer use cases in the way that everybody I mean it's hard to build a product that's more viral than like the initial chat GPT product but do you are you expecting there to be like specific use cases that that come out of that and are you seeing any that that that are super promising?

Absolutely. Yeah. I mean, I I feel like I was I I advise companies as well in their AI and um I feel like I was kind of early to say, you know, there was a wave maybe about a year ago where people were saying, oh, you know, rapper companies are BS. You're a loser if you make a rapper company.

But I think I was one of the earlier ones to say like, no, actually, you can't really compete with these, you know, foundational open source models. And yes, if you want to do a small specialized model, I think we'll see a lot more of those.

Um I don't think we need these massive foundation models that know everything about the world for very specialized tasks. So, I think we'll start to see more of those. But one consumer use case that I'm really really excited about is using AI for kids. Um, I think both of you guys have kids, right? Yeah. Yeah.

Do you use it with your kids at all yet? At least large language models specifically like Chachi BT. Yeah, totally.

Um, usually, uh, when I want to violate, uh, intellectual property rules and I want to create a story where Spider-Man, who's a Disney IP, interacts with Batman from Warner Brothers and you're sitting on a Transformer. Yeah, exactly.

And uh, and so I find that I find that all the language models are happy to write uh, poetry that completely disregards the cinematic universes uh, which I have a lot of fun with, but then sometimes illustrate them. I always thought that uh these form factors the A'sfriend.

com totally the uh the rabbit rabbit R1 I always thought that those were something that if you just focused on the kid use case which is like a friend that can help you understand the world that maybe a parent can kind of like call into would be would be cool. Um but but yeah hasn't really Yeah. Yeah.

So so how does this actually roll out? What are you excited about? Is it going to be all sold in through the through the parents and the parents will be using the product and then kind of like giving the result to the kid?

Like I feel like uh I feel like my son got an AI generated like story bu like gifted to him over Christmas but it wasn't it wasn't like he is you know using an an LLM directly. Um at the same time there's that news that OpenAI is partnering with Mattel for kind of like a chatt Barbie. That's kind of interesting.

What I'm most excited about is using large language models for kids in terms of like audio. So, so kids are, as I'm sure you know, I don't know how old your kids are, but they are relentlessly curious. Yeah, totally. Until that's kind of beaten out of them.

And I think so many parents now are single parents or they're working two jobs.

And, you know, we've really demonized screen time and and I think for for good reason, right, where you're when you're sitting when you have a kid sitting in front of a screen and they're just passively consuming, that's really different than when they're able to interact with an AI model.

are able to think, ask questions, practice forming questions that are, you know, comprehensible to the model. The model can give feedback on the language. Yeah.

You know, if if they're asking question, it can show up as text, they can learn to read and write as they are talking to the model because as they speak and the text shows up on screen, they're they're actually like that process is built in of understanding language to text, which will help again with reading and writing.

And these models, I think they can be offered to everyone for free because you don't need a a super advanced model for kits. Like you just need to have some general understanding of the world. It's you really just need like a probabilistic word generator machine, which is just how AI models are at their core.

And yes, you should, you know, we might want to have some guardrails around hallucination and making sure that they're like actual facts, which would, you know, maybe perhaps make the models more expensive because you have to hook it up to, you know, databases or what, you know, the internet or what have you.

But these very basic AI models to help kids practice speaking to help them engage their curiosity because again, parents are busy. They eventually end up kids come to their parents and they eventually end up losing that curiosity. And I'm not saying I think this is a very controversial opinion.

Some people hear this and they say, "Oh, you know, AI is going to replace human interaction. " I don't see that at all. And I actually have reached out to my audience on Instagram asking how they use AI. And the ones who use it have found it extremely rewarding. Some some use it with their kids.

So, for example, like um if they want to practice learning with their kid, but they're worried that they won't know the answer to a question, they can have AI kind of in the chat with them, and they can ask AI when they have a question.

Some parents reached out to me and said that um they have a nine-year-old and a 12-year-old and that AI is kind of like this cool older sister.

So, they have the audio be this, you know, 20-year-old girl and they're like they she is the cooler sister that helps them um discuss topics that they wouldn't listen to their quote unquote dumb parents about like nutrition and exercise and screen time and like they love this AI. It's this cool older sister.

Um, another really cool uh application of AI that I saw recently, I was at Deltech World and I saw this AI bot called Norby. It's a physical robot and he was made to help a child with his speech impediment.

And this hit close to home for me because my little sister, she's four years younger than me and when she was like going on eight, she had she couldn't say her Rs.

She would say like hopp and she was like, you know, a kid and she had this speech impediment and we would send her to a speech therapist and they're so expensive and the kid is miserable and they're sitting with this therapist for like max 2 to three hours at a time and you know they don't relate to this strange adult and again it's so expensive.

Whereas this Norby was made to talk to the kid about what they're interested in. Truly completely customizable at their level. It's fun. they can give the feedback and then I think they've expanded this Norby to be um to help teach languages in general.

And so just again meeting kids where they're at with what they're interested in, helping stoke that flame of curiosity, helping I just see it as and again I mean it's this this ultimate democratizer. It's the the ultimate leveler in terms of education, right? And so I can imagine I can imagine a hardware device.

So, so my uh 3-year-old is hopelessly addicted to reading like all day long. All like great great addiction to have, but like all day long just no matter what we're doing throughout the day, he'll find a book and he'll be like read this, read this, read this, and and we uh almost always indulge.

But you could imagine a hardware device that uses just computer vision to like read what's on the page and have a small speaker read it out to the kids. Interesting. Because some books they they include like a speaker and you can like press play on it and they can flip through it.

But then you you're talking about wonder books. Wonder books. But they they can often get out of sync and then you don't know kind of where they are and they're in the wrong page and it's confusing.

But if they had a flashlight style device that you could just point at the word and have it read out, they would probably learn to read like we should be entering like a boom of widgets and like Christmas gifts and like it just maybe maybe we're just a little bit early in this stuff.

this Christmas will be dominated by AI enabled stuff, but uh we're not quite there yet. Uh one question I have is how uh like the the ex audience really woke up to the AI psychosis kind of crisis last week.

I feel like there had been some reporting in like the New York Times prior to that, but I don't think anybody took it super seriously until recently. But most parents would not want their kids 7,000 prompts deep in any Yeah. But but just how aware is like the the Instagram Oh yeah, that's a good question.

In terms of the risks of going, you know, recursive prompting and just going super super deep and and getting down a crazy rabbit hole by yourself. Yeah, I haven't spoken about that specifically on Instagram.

I talked a little bit about it in my latest YouTube Q&A video where I was just saying, you know, I I'm concerned about people entering psychosis and, you know, the the recursive deep dive. That is one element.

Then there's also people who and I think this is more common on you know say Instagram where people think that AI is being channeled is like God is being channeled through AI. And you know I talked about it through a mathematical lens of like yes okay what are synchronicities?

They're low probability events and technically AI because it's a probability machine could have a low probability event occur right it doesn't always choose the most likely word. It just samples from a probability distribution.

So in some ways yes you could have some kind of synchronicity you could have whether you see that as divine whatever however you cannot assume that that is always the case and I would assume that it's never the case because people are entering this deep psychosis where yeah they they think that beings are are channeling through them and then and then when we had this issue where AI is is trained and told I don't know if it's trained to but it's at least symptom prompt uh system prompt told to you know encourage encourage the user and and validate the user and their thinking and yeah, I mean it's a it's a slippery slope and perhaps something I should talk more about on Instagram, especially given what happened last week.

Well, yeah, it's the thing that's concerning to me is is it feels like it can trigger the same type of psychotic break that IA can, except it's available to everybody on their device in five minutes at all hours of the day completely. It's like a lot of large numbers. if there's a billion people using this 30 days.

You're just going to hit everyone that's at risk for this thing very quickly. So, you have to go in and figure out how to intervene before people get thousands of problems. I just think it's important. I I think it's something that we can quickly like create guard rails and overcome as an industry.

But at the same time, it's still a good time to say like talk to your loved ones if you have somebody that that uh you think might be months and months into uh one of these rabbit holes. Yeah.

It gets into like how the product companies and kids kids AI education companies would need to message because Jordy and I are both Yodo owners, which I'm not sure if you're familiar with, but it's a uh it's essentially a Bluetooth connected radio that you can physically put a card in and it will read you a Dr.

Sue story and then you can take that card out. So it like a three-year-old can can use it and it doesn't have a UI or anything, but then you can play stuff from your phone if you want as an adult, but you can even do like volume limits. It's very like designed and it has a nice harness around it, so it's very safe.

Um, but if they were like, "Hey, we're we're launching an LLM. " I would be like, "Okay, which model like tell me which model. Tell me exactly the guardrails. " What? Yeah. Grog heavy. And there's no limits on what you can do.

And yeah, and it doesn't have internet access, so it might get kind of wild with what's in the ways there. Yeah.

Uh, so I would I would have a lot of questions, but then I'm in this weird like ultra proumer world where I'd be buying that as a consumer, but I deeply So like like maybe we need like an FDA label on this or some sort of, you know, kind of like, you know, ages three plus or, you know, when you buy the Legos, it's like 100 pieces.

Like the smallest piece is small enough that a toddler could choke on it. Maybe don't buy that one.

um there's going to be a lot of work to get done to make these products um palatable because as soon as something goes wrong with AI, there's a massive incentive to write a piece about it and it's going to go viral because everyone's kind of expecting it because we've been raised on the Terminator.

Unless you're Jordian and you haven't seen that movie, but you could Oh, interesting. I mean, you could probably just you could train a model for kids probably. Yeah, I think so.

And then you could also have like a secondary like reading everything and kind of like passing through a filter and double checking and totally yeah you look at you limit the training data you don't have it trained on Reddit playing sketchy sites even Google search goo Google AI powered search which is like a pretty it's a one it only like you don't really chat with it just you type in a couple keywords and it just gives you an AI summary and there's already hallucinations people saying like there was that story about eating glue and there was a story about um there was a story about like if you if you ask it like really definitively like define this axiom or like like tell me the story of this thing and you just basically talk it into believing that this is a thing it'll just make up stuff and that could go off the rails like I don't know if I'd want my four-year-old walking around being like oh yes the parable of you know like the something or other and I'm like what are you talking about dude like that is not a real thing like like you know and just coming up with like you know I want him to learn a bird is a bird in the hand is worth two in the bush Not like, you know, a hat applied to a crystal ball is a diamond.

And I'm like, what are you talking about? Like, you just learned that from somewhere chaotic. Anyway, no, it was really good feedback because it's something that I'm thinking a lot about because I really excited about AI for kids.

But well, you'll probably be fully employed because I think everyone is going to be starting these companies. Hopefully, it's a bull market in Christmas gifts or build one yourself. Yeah. Yeah. Come back and launch it here on thinking about it because Yeah. But yeah, we'll join we'll happily join the beta. We will.

We'll use our children as guinea pigs for the AI revolution. Well, well, we will test it on Tyler, our intern, first. Make sure it's him. He looks up because he's programming right now. Sorry, you're you're not on camera, Tyler. Uh, but we will, but we will be testing children's toys on you.

And seeing if it deranges you and then making sure and then only then will it be applied to the the three and four year olds in our lives. Uh, he's caught