17 Sep 2021
Adam Jesionowski
Butterfly Algorithms

Consider the task of designing a discovery algorithm for video content. Our immediate point of reference is TikTok–cluster users by how they interact with videos and show videos that capture the attention of alike users.

On the surface this seems to give the user what they want. The user signals desire through attention, the algorithm returns what sates it. Yet–when we look at the effect TikTok has on its users, do we find them better? Far from it: TikTok has taught our populace to be more neurotic, more dopamine sick. Fundamentally what TikTok wants is for you to watch another video. The purpose of a video is not to enact some real change in your life, but to provide an emotional response that keeps you scrolling.

A virtuous video content platform would not want this capturing of a user. Instead, we want videos that enable real world action. The ideal outcome of a cooking video is that the user makes a meal, a repair video that the user fixes something in their life, a sports video that the user plays a game with friends.

In each of these examples we see that a video’s purpose is memetic action. This suggests a new metric for the quality of videos: the ones that inspire the most real world action are the best. From this we may define certain elements of the platform.

Proof of Work

We want to know which videos inspired real world action. Requiring users to film themselves performing the action is too much of a hassle. Similarly, a button saying “I did this” is too easy. A user uploading a picture showing the result marks the right balance.

Consider the following flow:

The more photos uploaded, the more the video will be recommended. Which videos a particular user completes will allow us to cluster users not by attention spent, but by physical results.

Back to index