About the podcast

This episode from A16z features Eugene Wei, who has written a series of deep dives on TikTok, and formerly led product at Hulu, Flipboard, and video at Oculus. In this episode, he talks about TikTok and the power of it's AI algorithm and algorithm-friendly design. A great listen for anyone thinking about building social / entertainment platforms.

Key Takeaways:

  1. TikTok AI algorithm bypassed the social graph to go straight to the interest graph. This was extremely powerful and allowed it to scale across cultures and geographies.
  2. TikTok has achieved strong creativity network effects i.e. every new creator on the platform makes the existing community more creative. This is very hard to achieve.
  3. The platform has been designed to 'see as an algorithm' (with incredibly algorithm-friendly design) where every user action / product feature makes the algorithm more effective.
  4. We consistently underrate how engaging video is as a medium. Video is a very high bandwidth medium and you can layer lots of more info on top of it. Plus, video is extremely international because a lot of images, body language is universal and cuts across cultures. We have only scratched the surface of what we can do with video.
  5. While it is true that the algorithm is not unique in itself (and can be replicated), purchasing TikTok and not having the algorithm as part of the deal would be a huge setback because a new algorithm would take a lot of time (likely years) and a lot of training data to get as good as it is now.

Notes:

  1. While TikTok's AI recommendation algorithm is indeed powerful, it is not unique and can be replicated. What makes TikTok so formidable is the combination of the algorithm and the set of training data it is able to constantly feed the algorithm by (successfully) encouraging its users to create short-form videos on the platform.
  2. TikTok's creation tools are highly underrated, they in fact in some ways are the key to TikTok's success because they make it so easy and enjoyable to create high-quality videos for the ordinary user. For example, licensing the music tracks was such a powerful move because it meant users could lip sync to actual tracks of their favorite artists, which was not possible on other platforms (there was always a risk that the video would be removed).
  3. TikTok has creativity network effects i.e. every new creator on the platform makes the existing community more creative. This is very hard to achieve. TikTok's solves the 'blank page' problem for creators (i.e. do not know where to start) by allowing you to remix other people's videos. In fact, most videos on TikTok are actually riffs off other users' videos. This is actually encouraged on the platform and is structurally built-in e.g. users can do duets together or can use someone else's audio in their own videos etc.
  4. TikTok's AI recommendation algorithm also promotes creativity network effects. The platform's discover page tells you exactly what challenges are trending and incentivizes users to participate in that challenge (because users know which challenges are being promoted and gives them the greatest odds of being viewed). So TikTok, in a sense, is a cross between a free market and managed economy.
  5. Further, TikTok's algorithm does not take into account any social graph and focuses on the interest graph instead i.e. users see more of the kind of videos that interest them. This means that every creator has the same probability (irrespective of who they are or their social 'following') of their content being seen. This is much more meritocratic than other platforms (for example on TikTok, a user could have one video with several thousand views even when if all their earlier videos have very poor views). This meritocratic distribution is extremely valuable for new creators and is the main reason creators flock to the platform.
  6. Bypassing the social graph and focusing on the interest graph also allows for abstraction of culture i.e. content can scale across geographies and cultures (if I know you are entertained by cat videos, I will keep showing you more cat videos). It is a much smarter way to recommend content and increases the odds of a user being engaged with the platform. Almost all the other social media platforms use a social graph to approximate the interest graph, which is sub-optimal (because I am not interested in everything that a friend of mine is interested in).
  7. Bytedance has built TikTok using an algorithm friendly design, even introducing friction if necessary e.g. the full-screen, one-video-at-a-time format is technically not optimal from a time perspective for a user (it would be much more time-efficient to show thumbnails of several videos on the screen) but TikTok forces you to watch a single video at a time so that it can understand exactly how you engage with that video (better feedback loop), which helps train the AI algorithm better (eventually benefiting the user by showing him / her more engaging content).
  8. Interests graphs are also more valuable from an advertising perspective because eventually people pay for things they are interested in. So it makes a ton of sense to design platforms using interest graphs.
  9. The AI algorithm also adjusts automatically for a change in the user's taste (which happens all the time). This keeps users engaged more than other platforms which depend on social graphs.