Hey, Megan, let's do a podcast.

Great idea. What should we talk about?

So I was looking at LinkedIn the other day, and it kind of seems like when you went to technology, like maybe you got a new boyfriend or something.

How did this happen? It started over breakfast. Technology was in Los Angeles. So at about 06:00 a.m. All of the east coast bodies had been up for hours, and the coffee shop finally opened. So there's a bunch of people, all my friends, hanging out, and as I look around, I hang out with a bunch of really nerdy fun people. So we were just talking and catching up. It's been a while since we've seen each other. And I said, yeah, I've been meaning to just spend some time, some weekend and make myself a boyfriend.

GBT. And my good friend Josh Cavalier says, well, I can do that. And I said, how about this week? And he said, it's on. And the next thing you know, Sarah Mercier says, well, I could build one. And next thing you know, we had, I think we had three people just at breakfast. And I realized, oh, my gosh, we have a hackathon on our hands. So we all went for it. It was hilarious.

So, as a learning designer, I'm really curious, how was the whole hackathon constructed? What are the rules? How did you decide how it was all going to go down?

So that was actually a big part of it. And there's lots on LinkedIn. You could see how everybody's. And we'll write up an article, but how everybody built their bots and stuff. But the building of the hackathon itself was something I haven't done in a few years and haven't done competitively. So that was fun. We ended up, we had five teams. Each team had the challenge to make an AI boyfriend that I would interview.

Each team got four pages of training data. They all had the same information about me. Now, all the teams were led by somebody whom I'm a friend with, so there was some familiarity already. And then they got four pages of things I like, things I don't like, my hobbies. Therapist put in a blurb. My best friend put a blurb. One of another good friend at the conference put in a blurb. He looked up, so Chris King put in a blurb, and then he looked up on chat GPT, and chat GPT had something to say about me, which was hilarious.

We put that in the training data, all sorts of stuff. Each team could interview me for 15 minutes or email me, and they knew that they would have five minutes to present their solution. And so this was, we started Wednesday morning, and then we presented at 445 on Thursday. Actually, ATD was really amazing and flexible because I went to them at about 11:00 on Wednesday morning and said, hi, we're doing this thing, and can we make this an official conference thing? And they said, totally, yes. Put it in the conference app, the whole nine yards. I used chat GPT to write the description that went on the conference app. So it was all full level. I had a rubric that I used chat GPT to create the rubric, because I explained to chat GPT, here's what I'm doing.

How should I set it up? Actually, chat GPT ended up being an amazing thought partner. As I pulled this together, it thought of some things I didn't. And then we had a dating game, like the old timey game show. Dating was the format for the big reveal. So I asked each bot, I wanted to keep it fair, as fair as I could. And that was, in and of itself, unfair. So that was interesting. So my plan was that I would ask each AI agent the same questions so that we could see the same answers compared, right? So the questions were, how would you support me on a busy week? How would you engage with me as a thought partner? So thinking with, okay, I had to do some serious thinking about what I really wanted, right? Which probably more than I ever have.

So that was valuable for me personally, but also for the bot design. Because if I can't be, I mean, just from putting a learning hat, if I can't be specific about my needs or the needs of the use case, it's impossible to build a bot that would support that, right? I had those two questions I wanted to know, what would you ask me to get to know me better? And that got some interesting responses. And then there was one question that the first bot that went was Vince Hahn's mobile coach bot. And just the way mobile coach works, he could suggest the next question, I could ask questions, or the bot could suggest the next question. And so one of the suggestions was, make me blush. Yikes. Yeah, totally yikes. Right? Like, completely.

I'm in front of, like, 70 people who came to watch this, and I'm pushing a button that says, make me blush. Right? Awkward. So then I asked each of the other bots to say something that would make me blush. And I have to say, the AV team, who are otherwise the unsung heroes of conferences anyways, they had to switch out between my presentation to open it up. And explain to everybody what we did. And then five different presenters with their different phones or laptops. We had some Macs, everybody had a different phone and a different connector and all this stuff. And those guys were, they made it work.

That's awesome.

So that was huge.

It's fascinating to think about five different approaches to the same exact problem using the same training data that you fed to it. It really shows kind of the power of the design in an AI solution. So I'm curious to know, what were these solutions like? Tell me about them.

Yeah, I'll introduce you to them. It was kind of. So, I mean, we talk about a b testing and usually we don't get the opportunity to a b test. And here we had what ABCde testing. So I mentioned Vince Hahn's mobile coach bot. His name was Ben. Everybody had a name. And mobile coach is a rules based platform that uses natural language processing and generative.

So there was a set of rules and then behind it was the four pages of training data. What that means is it's a super controlled environment, so Ben is not going to go off the rails. Ben's pretty straight laced. He did really well sticking at those boundaries. And like I said, it had those next most logical. Or I could still enter what I wanted to, but it would also kind of forward the conversation a little bit by suggesting what I could ask them. Sarah Mercier from build capable. Her bot was named Nathan Hale with hers.

And then Josh and Myra also had, they came with pictures and images. So we had a lineup of very attractive thoughts to talk to. This gets awkward, more and more awkward as we built around, she had an OpenAI GPT, Josh Cavalier's kai. So k AI kai, get it? It was interesting because they had similar interfaces, different voices. So Sarah's had eleven labs, different voices, different approaches and different personalities on top of that. But by time you got to Kai and Nathan, got to keep my guys straight here, by the time you got to them, you could start to see the training data poke through. So they were each very clearly, very aware already of my hobbies and stuff, which it made it interesting. It also made it a little bit not predictable because they weren't predictable, but it was like, got it.

Those are my guardrails around that conversation. Right? Which from a learning design perspective in the real world, that's useful. Myra Rolden. Okay, she made me blush. This was hilarious. She warned me in advance that it might get inappropriate and I just looked.

At her funny and she's.

No, no, really, in testing it got inappropriate. So we were actually ready to pull its plug if we had to. So she used candy AI, which I have no experience with. It is basically like a dating AI bot, and it definitely had a very distinct personality. This bot was very intent on picking me up. I felt like I was at a bar and being hit on where everybody else I was asking them, and it was dating game, and this guy was like, he wanted to go places and he wanted to buy me drinks, like completely ignoring the training data that said, I didn't drink. It was very interesting and fun, but aside from just that, it was just kind of cool to see that additional layer of personality on top of it. And then Mike Ruska from problem Solutions, his bot's name was Steve, and his was a combination of OpenAI.

We used the AI mobile app and the voice. He had asked me before what voice I use on my chat GPT mobile. So he chose that voice. And I have to say, OpenAI, if you're listening, you probably are. Outstanding job on the voices, because they say ums and they hesitate. It was good, right? And interestingly, that stole the show, and I knew that's open AI Mike didn't do anything to make that happen. Right. So that was interesting.

His had some specific instructions not to reveal the training data with the express intent of not being creepy. So that was really interesting. And his was designed to be. His was so much designed to be conversational that when I tried to ask those same five questions, or same questions, to be fair, it came across as me being really rude. So, Steve, the Steve pot was asking me questions. I'm like, no, no, I've had a really busy week, and I'd like to know how you would support me on a really busy week. And it made me come off as rude, which was really, I felt very self conscious about, but it was very, very interesting.

So, like, the bot was wondering why you were honing in on these things and not being conversational.

Totally, totally. So it would ask me a question and I'd say, yeah, whatever. What I want to talk about is this. So it was not Megan putting my best foot forward here.

That's awesome. It's really interesting to think about how the same thing can come across so differently based on the prompt, the instructions. Everybody's coming in with a different sort of angle or bias on the same thing and the same challenge. So that's really cool. So this is fun, and I like to think of you up there blushing.

Oh, yeah.

Putting all that fun aside, what's the practical value here in what you did at technology.

Okay, well, in fact, the fact that it happened at a learning conference really forces it. Like, what is the learning value? So completely fair question. This was an edge case, right? Put it way out there. If we're trying as an industry to create personalized and customized agents for role plays, for coaches, for whatever, what's more personalized and customized than your partner? So let's do that. If you can make a boyfriend, you could probably do anything, right? So I think about it in two ways. So one is around having a personal coach who really knows you, right. And whether it's coach, therapist, whatever. Actually, early in the conference, I was getting set up to do my first session, and I realized, gosh, this room is really, really noisy, and I'm just going to just for giggles, ask my chat GPT mobile, what kind of things would you gear up to prepare to speak in a situation like this? Competing audio, lots of things moving in the room.

And I got a little pep talk. Now, it wasn't personalized to me, but interesting. The audio I was listening to it from was my lucid, Bluetooth, GPT enabled sunglasses. So I'm standing there looking at the room, hearing in my head, coaching tips for managing the audio in the room, making sure it gave me a reminder to make sure I do a sound check and other things. I thought, wow, from a situated, contextualized coaching, this is powerful. Didn't even know me. If it did, it'd be even more so, right?

Yeah. That would add like a whole nother deeper layer to it because it would be able to really tailor that to what it knows, the needs that you have.

Yeah. Right. You know me, right. So there's meetings in which I need to either dial up or dial down my delivery. And so having someone who knows me like you, but when you're not there to say, oh, this is where you have to dial this, the other side is. So think about all the times that we as learning designers try to make role plays or scenarios, interactive branch scenarios in an elearning course or role plays in an instructor led course, and you need a sophisticated partner to get it right. Or the answers to the ways the conversation could go aren't multiple choice. Right? So I recently saw or heard about a threat deescalation role play, and you would interact with this AI agent, this AI bot, and depending on how you reacted to it, its training data said you either amp up the situation or you amp it down or you dial it down.

And just hearing about that reminded me of my AI boyfriend.

So I think everybody's question that they're wondering here is, would you date any of them?

Probably not. But ask me on a different day. Okay.

When you said there are situations where I either need to amp up my energy or bring it down, the only thing that stuck in my head was like, are you putting on your fancy.

Pants.

Or leaving them at home?

Well, and you do know, have I ever told you the story, Meg, that when the pandemic first hit and all the conferences and webinars went online and I would come into the office, the only person in the office, I was the only person in town, practically. And I would come into the office, and in order to do a webinar, I had to change out of my jeans and put on dress pants because I couldn't get my head. It's funny you say fancy pants, right? I couldn't get my head in the game to deliver a talk unless I was wearing dress pants. Not. I get it.

All right, Megan, how'd that one go?

Okay. To be honest, way less awkward than the actual speed dating hackathon, which was. It was fun. It was awesome. It was also super awkward and vulnerable. So, yeah, this was much nicer to talk to you about it. I get that.

This is Meg Fairchild and Megan Torrance, and this has been a podcast from Torrance learning.

Tangents is the official podcast of Torrance learning, as though we have an unofficial one. Tangents is hosted by Meg Fairchild and Megan Torrance. It's produced and edited by Dean Castile with original music, also by Dean Castile. We did not bother to fact check this episode.