Thecryptodeviant

Building the AI of Thirst The first version of the Juicy feed was dead simple. Most recent posts. That's it. I think we updated it to be people you were following at some point. The second version got more clever with complicated trending queries. It was interesting in a way, but still mostly irrelevant to people. Then I had this idea to try to build something that felt more like TikTok. Well... we missed. What we built was cool, but it really underestimated the problem. We took every photo and video, ran it through AWS rekognition, got all of the descriptive and moderation labels (which is a lot of info). Then we piped that info through GPT 4 to generate detailed JSON categories based on an in depth content taxonomy we defined. Then as users engage with things, you can say, "oh this person is interested in XYZ... add a point to that category." Then the algorithm started showing people things based on the most prominent things in their own personal taxonomy scorecard. Not bad. But nowhere near TikTok. It would actually work *decently* well with a mountain of content, but it totally falls flat when you only have limited content. A couple months back, I started experimenting with building my own custom AI solutions. Idk when it will be production ready - hopefully during or right after the holidays - but the next iteration is going to get 80% within a TikTok quality of recommendation system.* We are creating a set of 3 vector embeddings for each video, scoring them on a proprietary set of "thirst axes" (custom subspace), and then suggesting videos in real time that are more and more in line with your interests. There's actually several more stages to the pipeline, but if you're interested, that's the general idea. The current version of juicy.box is just random ordering. Mostly to work on UX quirks while we wrap this up. Once this is done, we will open it up so any creator can post, and the entire thing will be regulated and managed by very refined AI systems. *Niching down to "thirst content" actually makes this probably an order of magnitude easier than a general purpose social network.

We need people building more things, not less. More experiments. But we need to focus energy outside of crypto. We’ve become too circular, and this is the heart of pvp.

Roughly 35% of TikTok traffic and about 45% of Instagram traffic is thirst traps or thirst adjacent. Yet only around 25% of TikTok posts and 35% of Instagram posts fall into that category. Thirst is over represented in distribution and under represented in product. Both platforms, especially TikTok, aggressively suppress and ban creators in these categories. Launching a new account here is now extremely difficult unless you deeply understand how to work with the algorithms. That contradiction creates demand for something that is not outright p**n, is thirst forward, not heavy handed with censorship, and has a TikTok level UX. That is the gap we are building into.