Notes from a16z's "TikTok & Beyond: The Algorithm Question" with Eugene Wei

22nd September 2020 - 14:23:53 - 3312 characters

Notes from TikTok & Beyond: The Algorithm Question, The Future of Product based upon Eugene Wei's TikTok and the Sorting Hat:

  • Creation: TikTok avoids the cold start problem by having creators have a set of known challenges that they build upon
  • Mutation: TikTok provides near professional grade tools (i.e. Adobe Creative Suite) with minimal learning curve for modifying the video and audio including cutting and splicing videos together (i.e. Duets)
  • Variation: TikTok strongly suggested what would trend and hence creators knew which channel of content creation would provide them the best spread
  • "Does every additional creator on TikTok make the rest of the community more creative?"
  • "TikTok is a mix of a free market and a managed economy"
  • TikTok is a powerful force for meme creation and propagation that has been sped up ("short circuited") to maximum, not relying on a slow to construct follower graph for information to flow
  • The app allowed the audience to create a new form of video that didn't exist on any other application
  • The recommendation algorithm encourages an interest graph over a follower graph where this prevents the cold start of new creators having to create a following before their content can be consumed
  • YouTube's algorithm that prioritizes those you've subscribed to means that an established creator putting out low quality content can pollute the ecosystem but still remain on top (higher noise to signal ratio in one's feed)
  • TikTok's algorithm uses a small audience test to decide whether your content will sink or float regardless of how many followers you have (i.e. a continuously running fitness function)
  • Hash tags provided numeric values such that you know exactly how popular a hash tag is, not just that it's trending
  • Single item of content per page allows for easier attribution of actions
  • The TikTok team provide explicit features such that they're not reliant on image classification that won't have the breadth to cover an evolving and edge case based social network
  • AI allowing for a social network to transfer and scale to an arbitrary country
  • The algorithmic "sorting hat" allows you to scale the sorting function infinitely such that a billion people can be sorted into a near infinite variety of buckets
  • Reddit does a passive version of this where you explicitly need to find a given subreddit to follow
  • TikTok's active version sorts you according to the way you interact with the app where annotation and usage are equivalent
  • Video is likely the highest bandwidth medium compared to text or audio and is a fundamentally unexplored domain
  • Video isn't inspectable as text is but this can be alleviated with transcripts and potentially AI analysis
  • If the underlying issues of video are solved we could find that video grows into other fields (education, recommendation, ...) and becomes a bed for other information to be layered on top of it
  • Video is an "international language" compared to audio or text
  • Social graphs are likely the most valuable graph / private dataset in the world and they're not impenetrable