Jesse Silver: evaluation

Transcript context

…Talking about open source model inference, how do you think about latency? I think most people optimize for latency in a way, especially for like maybe the Diva archetype, you actually don't want to respond for a little bit. How do you handle that? Do you like as soon as a message comes in, you just run the pipeline and then you decide when to respond or how do you mimic the timing? Yeah, that's pretty much right. I think there's a few contexts. One context is that sometimes the product is sexting with a fan with content that's sold as if it's being recorded in the moment. And so latency, you have to be fast enough to be able to provide a response or outreach to people as they come online or as they send you a message because lots of fans are coming online per minute and the average session time seems like it's seven, eight minutes or so for reasons. And you need to be able to interact with people and reach out to them with sort of personalized message, get that generation to them before they engage with another creator or start engaging with a piece of media and you lose that customer for the day. So latency is very important for that. Latency is important for having many, many concurrent conversations. So you can have 50 concurrent conversations at once on large model profile. People do take a few minutes to respond. They will sometimes respond immediately, but a lot of the time people are at work or they are just jumping in a car at the gym or whatever and they have some time between the responses. But yes, mostly it's a paradigm. We don't care about latency that much. Wherever it's at right now is fine for us. If we have to be able to respond within two minutes, if we want the customer to stay engaged, that's the bar. And we do have logic that has nothing to do with the latency about who we ignore and when you come back and when you leave a conversation, there's a lot of how do you not build a sustainable non-paying relationship with a fan. And so if you're just continuously talking to them whenever they interact with you, and if you just have a chatbot that just responds forever, then they're sort of getting what they came for for free. And so there needs to be some at least like intermittent reward element or some ignoring of someone at the strategic ignoring or some houting when someone is not buying content and also some boundaries around if someone's been interacting with you and is rude, how to realistically respond to people who are rude, how to realistically respond to people who haven't been spending on content that they've been sent. Yep. And just to wrap up the product side and then we'll have a more human behavior discussion, any sign from the actual fan platforms that they want to build something like this for creators or I'm guessing it's maybe a little taboo where it's like, oh, we cannot really, you know, incentivize people to not be real to the people that sign up to the platform. Here's what the dynamics are there.…

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