Alessio Fanelli: belief

Transcript context

…I guess the first time we worked on a profile that was doing at base over $150K a month, we put the product on and produced nothing in earnings over the course of two days. We were producing a few hundred bucks when you expect $5,000 per day or more. And so we're like, okay, what went wrong? The profile had been run by an agency that had an offshore chat team before, and we were trying to figure out what they had done and why they were successful. And what we were seeing is just that the team was threatening fans, threatening to leave, harassing fans. Fans were not happy. It was complaining, demanding they tip, and we're like, what's going on? Is this sort of dark arts guilt? And so what it turned out was that this creator was this well-known inaccessible diva type. She was taking on this very expensive shopping trip. People knew this. And the moment we put a bot on the profile that said, oh, I'm excited to get to know you. What's your name? Whatever. We're puncturing the fantasy that the creator is inaccessible. And so we realized that we need to be able to provide a coherent experience to the fan based off what the brand of the creator is and what sort of interaction type they're expecting. And we don't want to violate that expectation. We want to be able to give them an experience, for example, for this creator of where you prove your masculinity to them and win them over in some way by how much you spend. And that's generally what the chat team was doing. And so the question is, what does that overall fan experience look like? And how can our product adjust to a variety of significantly different contexts, both serving significantly different creators and serving fans that are wanting one or multiple on different days of a relatively small set of things? That makes sense. And I think this is a technical question that kind of spans across industries, right? Which is how do you build personality into these bots? And what do you need to extract the personality of a person? You know, do you look at previous conversations? You look at content like how do you build that however much you can share? Of course. People are running the same thing when they're building sales agents, when they're building customer support agents, like it all comes down to how do you make the thing sound like how you want it to sound? And I think most folks out there do prompt engineering, but I feel like you figure out something that is much better than a good prompt. Yeah. So I guess I would say back to replicating tone. You have the option to handcraft your prompts. You have the option to fine tune. You can provide examples. You can automate stuff like this. I guess I'd like to inject the overall fan experience just to provide sort of a structure of it is that if you imagine sort of online girlfriend experience or girl next door, if you reach out to this creator and say, I'm horny and she just goes, great, here's a picture of me. I'm ready to play with you. That's not that interesting to a fan. What is interesting is if you say the same thing and she says, I don't even know who you are. Tell me about yourself. And they get to talking and the fan is talking about their interests and their projects. And she's like, oh, that's so cool. Your project is so interesting. You're so smart. And then the fan feels safe and gets to express themselves and they express their desires and what they want. And then at some point they're like, wow, you're really attractive. And the creator just goes from there. And so there's this structure of an escalation of explicitness. There's the relationship building phase. The play that you do has to not make the customer win the first time or even the second time. There has to be more that the customer is wanting in each successive interaction. And there's, of course, a natural end. You can't take these interactions on forever, although some you can take on for a very long time. I've played around with some other not safe for work chatbots. And I've seen fundamentally they're not leading the conversation. They don't seem to have objectives. They're just sort of giving you what you want. And then, of course, one way to do this would be to meticulously handcraft this business logic into the workflow, which is going to fail when you switch to a different archetype. So we've done the meticulous handcrafting, especially in our prototype phase. And we in our prototype phase have done a lot of prompt engineering, but we've needed to get away from that as we scale to a variety of different archetypes of creators and find a way to automate, you know, what can you glean from the sales motions that have been successful on the profile before? What can you glean from the tone that's been used on the profile before? What can you glean from similar profiles? And then what sort of pipeline can you use to optimize your prompts when you onboard or optimize things on the go or select examples? And so that goes into a discussion, perhaps, of moving from our prototype phase to doing something where we're either doing it ourself or using something like DSPy. DSPy.…

Stored transcript either side of the excerpt. The highlighted words are the published quote; the surrounding text is unedited source, never generated.