High Signal Podcasts Evidence ledger
Method
Browse
← Back to evidence

Evidence receipt / evaluation

Published · transcript-backed

Jesse Silver: evaluation

16 May 2024 Latent Space Emulating Humans with NSFW Chatbots - with Jesse Silver

“And so as a result, don't generally do a good job of chatting. And there's also elements of the fan experience that if you're training from data from a chat team, they will do a lot of management of people that don't spend, that we don't need to do, because we don't have the same sort of cost per generation as a human team does.”

— Jesse Silver

Source trail

Everything needed to verify it.

Speaker
Jesse Silver
Attribution
Verified speaker
Claim type
evaluation
Recorded
16 May 2024
Publisher
Latent Space

Transcript context

…When you think about fine-tuned model, how do you build that data set, I guess? More like the data set itself, it's like, what are the product triggers that you use to say, okay, this is like we should optimize for this type of behavior. Is there any sort of analytics, so to speak, that you have in the product? And also like in terms of delivery, is the chat happening in the fan kind of like app? Is it happening on like an external chat system that the creator offers to the customer? And kind of like, how do you hook into that to get the data out? I guess it's like a broader question, but I think you get the sense. Yeah, so we have our backend, which needs to scale to potentially millions of conversations per month. And then we have the API, which will connect to the fan platforms that we work with. And then we have the workflow, which will create the generations and then send them to the fan on the fan platform. And gathering data to fine-tune, I think there's some amount of bootstrapping with more intelligent models. There's some amount of curating data from scraping the profiles and the successful history of interaction there. There's some amount of using model graded evaluation to figure out if the fan is unhappy and not paying, or if something has gone wrong. I think the data is very messy. And sometimes you'll onboard a profile where it's doing tons of money per month. It's doing 200k per month, but the creator has never talked to a fan ever. And it's only been a chat team based in the Philippines, which has not terribly great command of English and are not trained well or compensated well or generally respected by an agency. And so as a result, don't generally do a good job of chatting. And there's also elements of the fan experience that if you're training from data from a chat team, they will do a lot of management of people that don't spend, that we don't need to do, because we don't have the same sort of cost per generation as a human team does. And so if there's a case where they might say, I don't have any time for you, spend money on me. And we don't want to pick that up. And instead, we want to get to know the fan better. Yeah. Interesting. Do you have an estimate for cost per generation for the human teams? What do they charge actually?…

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

Search evidence