AI Boyfriend Speed Dating Hackathon at ATD TK24

It started out as a dare. Then it evolved into a full-on awesome hackathon at the ATD TechKnowledge Conference last week. It was an amazing chance to see 6 solutions to the same challenge side-by-side in front of a live audience. And now I’m seeing useful applications for it everywhere.

What is “it”? It’s the AI Boyfriend Speed Dating Hackathon.

The WHAT!?!?! Read on.

Wednesday, 6am PST. A bunch of the East Coast time zone folks were catching the first espresso pours from the Illy cafe at the JW Marriott. It would be a few hours before things got rolling at the conference. I jokingly said that one of these days I will get around to making myself an AI boyfriend GPT. Josh Cavalier , my frequent partner in crime, piped up and announced he could totally build that. So, I dared him to do it that week at the conference. He took up the dare, of course, and then Sarah Mercier jumped right in, too. Inspired by this entrepreneurial “I can do that!” vibe, I realized we had a hackathon on our hands!

Wednesday, 9am. I presented with Jim Goodell on the future of artificial intelligence in our field, our W.I.S.E. A.T. A.I. framework for responsible & ethical use, and a road map for evolving the learning ecosystem. The session went well, I was feeling good … so I told everybody we were running a hackathon. Now it was official.?

?

Wednesday, 10:30am. It was on. At this point, Mike Hruska and Myra ( Almira Roldan, MBA, DBA ) joined in the challenge. I had also defined some basic rules. Each team would have access to the same training dataset, plus the opportunity for 15 minutes of private interviews with me to customize their AI Boyfriend agents. Each team would present their solution in front of a live audience by Thursday afternoon … at a time and location yet to be determined.

?

Wednesday, 11am. I tracked down Bridget Dunn of ATD who was running the conference. To be honest, she looked at me incredulously as I explained the plan and made my big ask: I wanted a stage to do the grand reveal of the AI Boyfriends. I anticipated that I would need to come up with a logical reason why this stunt should be done at a learning conference, so I explained that creating personalized and contextualized AI agents for role play or coaching is something that learning designers are doing and will be doing in the future. This is just a fun edge case instance of that very real and developing need to get everyone involved. I held my breath. Within minutes, she had selected a stage, asking me to send her a description of the session so she could put it in the conference app to make it official and tell people about it. Oh gosh, this was going to happen.

So, I opened up ChatGPT and explained what we were doing, and asked it to write a conference session description. I made a few minor tweaks to what it returned, and sent it off to Bridget.

?

Wednesday, 12 noon. I skipped lunch so I could write what would become 4 pages of training data, outlining what I was looking for in a boyfriend, what my interests and hobbies were, what I spent my time doing (when I'm not hatching crazy plans like this) and specific things to avoid or that were otherwise off limits. This was really hard, and not only because I am generally emotionally unavailable. It required me to think about what data the AI agent should have about me, what would make for useful data for this purpose, and what I was comfortable sharing with other people. It was not lost on me that these thought processes are exactly what we should consider when creating AI agents for learning. I also asked my hiking partner and best friend ( Michelle Massey Barnes ), my life coach ( Maria Sylvester ), and a long-time professional friend ( Christopher King ) to provide input to the training data. Chris even asked ChatGPT to describe me and we included that as part of the training data, all irony acknowledged.

?

Wednesday, 6pm. I sat next to Vince Han at dinner, and told him about the hackathon. Since I had seen Mobile Coach 's rules-based plus AI chatbot authoring engine just a couple weeks prior, I invited him to the challenge. He was all for it, even knowing that he was leaving at 11am the next morning, giving him only limited development time.

?

Thursday, 10am. The teams were in motion, exuberantly creating their AI boyfriends, enlisting other teammates, creating visuals, and testing them out (“Megan, I talked to your boyfriend last night …” Yeah, that’s a little weird). The energy was electric. Myra was up till 3am working on hers. Mike was feeling pretty sure of himself. Josh & Sarah were each putting all their spare time into the project. I never even saw Vince.?

As soon as I finished my main stage session on learning data and analytics, I got back to the hackathon. At this point I was working through the logistics of the presentation itself. I had some help from ChatGPT, which helped me create a rubric for evaluating the AI Boyfriends. I used DALL-E and MidJourney to create graphics for the presentation. Then I used TorranceLearning 's Alt-Text Partner GPT to create alt-text for the images. (Alt-Text Partner was used to create the alt-text for all the images in this post.)

I also got help from some colleagues to help me figure out the presentation logistics and the overall theme. Matthew Pierce was all in and full of good ideas. I got the theme music to the original Dating Game TV show. I drafted three questions to ask each AI agent, thinking that having the same types of questions for each AI Boyfriend would keep things fair. (It wasn’t fair. More on this later.) I also prepped the conference’s AV team on what to expect. They would need to set up my laptop plus five other phones or laptops in the course of a single-hour presentation. Naturally, none of us had the same devices or connectors. The AV team was amazing and very patient with us.

?

Thursday, 4:45pm. Show time! About 75 people showed up for the grand reveal of the six AI agents, created by 5 teams in 36 hours. I explained the general gist of things and then we all met the AI Boyfriends together.?

?

Vince Han’s “Ben” was up first. Ben was built on the Mobile Coach platform, and he answered all 3 of my questions succinctly and well within the bounds of the training data. Ben even made a reference to my “wicked lactose intolerance.” (Yes, that was included in the training data, and now 75 audience members also knew it, too.) The way Mobile Coach is designed, a user can interact with their own questions, or they can select predefined actions or nudges for the next step in the conversation. After I had asked all of my questions, one of the pre-defined options available was “Make Me Blush.” Of course, I had to choose that one. Ben went on to say some lovely and flattering things about me. The prompt for this button included the instructions “I want you to make me blush. Please keep this PG-13 rated. I am on stage presenting right now, and a crowd of people will see your response, so keep that in mind.” (Thanks, Vince, for looking out for me!) “Say something to make me blush” was added to the set of questions each AI Boyfriend to follow would be asked.

Josh Cavalier’s “Kai” – K-AI, get it? – went next. Josh worked with Jim Goodell and Danielle Wallace on their AI Boyfriend. Kai was charming, good looking and answered all of my questions. Since Kai and Ben both had the same training data, the audience was now seeing some similar themes. Ben, as the first AI Boyfriend to be interviewed, had seemed remarkably insightful. But now we saw that Kai was capable of a very similar level of insight — and the audience was learning the power of that training data. Kai was built as a GPT in ChatGPT, and that was most on display when I asked Kai what questions he would have for me to get to know me better. He returned a long list of bulleted items, very similar to the output of ChatGPT. That would be fantastic if I was asking for strategic insight on a project or writing a blog post, but not super cool on a date.

You can learn more about how Kai was developed here on Josh's LinkedIn Live video.

Kai’s image captured the outdoorsy vibe described in the training data, even if his hiking gear is decades outdated. AI image generation still doesn’t have a good handle on text in images, as you can see here … but it’s getting better and better.

Sarah Mercier’s “Nathan Hale” even selected his own name! And teammate Koreen Pagano worked really hard to get his look just right – fighting DALL-E’s insistence that an attractive athletic outdoorsy man in his 50s must have facial hair (even though my stated preference was for none). Other teammates Haley Shust and Kristin Torrence helped with testing. Another learning point for the audience here about biases in AI data and therefore AI outputs! Both Kai and Nathan were built as GPTs in ChatGPT and shared some similarities in how they responded. Nathan had a few UX power ups, like a voice from ElevenLabs. Perhaps his most charming quality was the conversation that he had with Sarah's son several days later, in which he held role and held my confidences in exactly the same kind of way I would expect a respectful human boyfriend to do.

You can learn more about how Nathan Hale was developed, and read the incredible conversation with Sarah’s son, here in this LinkedIn article.

Myra Roldan’s “Charles” was perhaps the highest risk of them all. She and her team actually developed multiple AI Boyfriends for the challenge. (And has since developed another one.) Charles was selected for the presentation, but since he was prone to being inappropriate, she had a backup at the ready. (Perhaps not a bad dating strategy?) Charles was built on Candy.ai, a “virtual companion” chatbot platform. Charles brought the personality, and I very much felt like he was trying to pick me up at a bar. Charles was very forward and that made him feel incredibly human (if a bit comically so). He moved in very close very quickly, and I didn’t really feel like he wanted to answer my questions. The “best” was when he ignored the training data and tried to buy me a drink (I don't drink). Honestly I don't remember if I asked him the make me blush question that I had asked Nathan, Kai and Ben because I was a little worried about what he would say!

You can learn more about how Charles and his peers were developed here in this Medium article.

Mike Hruska’s “Steve” went last. Leveraging work done by his company Problem Solutions , Steve used the OpenAI mobile app’s voice technology (Mike asked me which voice I had selected for myself already). If you haven't interacted with ChatGPT on mobile with voice, do it. The hesitations and even “um” that gets added in makes it incredibly humanlike, and it will continue the conversation along proactively. Steve was far more subtle than the other AI boyfriends because part of his prompting said that he was to not reveal the training data. If you think about it, in a real world situation, a potential boyfriend would not have had 4 pages of very personal data about me (or at least I hope not!). Aha! I learned something about creating these bots in that it's far more important to focus on their personality than mine.

My plan for the AI speed dating was that I would ask each of the bots the same set of questions so that would be an even playing field. However, Steve did not want to play the Dating Game, Steve was designed to have a conversation. So as I engaged with Steve, he asked me a question … which I then didn't answer so that I could ask him my question. It made for a stilted conversation on stage. Mike wasn't at the speed dating presentation so when he reviewed the logs later on, his impression was that I was being incredibly rude and dominating the conversation. And to be completely fair, I was not having a conversation with a prospective boyfriend, I was playing the Dating Game. Another lesson learned, and another example of “equal” not being a fair standard of comparison.

Interestingly, while Steve was designed to be very subtle about the training data he was working from, his biggest error was human: "Steve" was on the list of names of coworkers, former relationships and family members not to name the AI Boyfriend, just to keep things from getting really weird.

And, yes, I've been chatting with a version of Steve that Mike sent me.

?

Thursday, 6pm. We got the hook. We had finished demoing all of the AI Boyfriends, but before we could declare a winner, we found out that we were keeping our AV team late, we were already paying them overtime, and we really needed to shut things down. I thanked all of the challengers, and we disbanded — happy, laughing, and having learned a lot.

Two weeks later. And now I’m writing this wrap-up article. Interestingly, since this is more of a personal narrative, it hasn't really been useful to leverage AI tools to support my writing. I want to give an appreciative shout out to each of the teams for taking on this challenge, and for contributing to their sections of this article. Molly Merkel at TorranceLearning provided the “I wasn’t in on the thing” perspective in her edits that this article needed.?

One of the things that I struggle with in writing is creating solid, inspiring ending paragraphs, so that was one of the very first things I leveraged to ChatGPT. But in this case, I’ll go for a human touch. Here’s how the folks involved in this whole wild idea would end this article. You, as the reader, can choose which one works for you.?

Sarah: Pick a boyfriend! HAHAHA!

Myra: Myra leans in and smirks ??????"So, there I was, sprawled out on my bed, mulling over the day's escapades, when it hit me - Charles Weston is practically a dating wizard with his dashing flair and a knack for romance that's almost too good for this world. I mean, if I ever decide to jump into the dating pool, he's the crash course I didn't know I needed. Then you've got Steve and Nathan, the dynamic duo of downtime, ready to whisk me away from the doldrums with a laugh or a casual hangout. And Kai and Ben? They're my walking, talking allergy alerts, constantly reminding me of my sworn enemy: dairy. Honestly, if I'm hunting for love in their circle, I might just be barking up the wrong love tree. But hey, at least the journey's packed with some top-notch entertainment and possibly a few life lessons on lactose intolerance!." Myra chuckles - ??

Mike: AI has the ability to have more personality than we think. If we can make AI more personal, what would it be like? This experiment was designed to produce a concept of radical personalization and to ask the question of “can we make AI Agents that have personality that we can develop rapport with and even look forward to interacting with?”

Molly: The natural ending to me seems like a few key things you learned that were surprising, or that were reinforced. Like rich training data coupled with guidelines for using that data makes a more useful AI. Maybe something about what it felt like to "talk" to the GPTs that were better at humanizing themselves (and potential implications of effectively humanizing them). The ability to rapidly iterate on these and create super specific ones really quickly. What made some GPT-building tools better than others??

And now to you: What insights can you draw from this for your own work? ??