Created by Rishabh Srivastava, Founder of Loki.ai
This summary was largely done for my own note-taking, sharing it just in case it adds more value to other people.
I have no affiliation whatsoever with anyone in this note. This is a summary largely taken for my own reference, and may contain errors :)
Context
Source URL:
Why is it important: This will help me pitch the “TikTok for X” idea in ways that non-technical people can understand
Keywords
Personalization, TikTok, Content
Summary
...
Highlights
The algorithm for the “For You Page” is the most important piece of tech that ByteDance introduced to TikTok. It helped a team to people who had never left China to crack the cultural barrier and grab massive market share in places they had never experienced first-hand
The ByteDance algorithm is based on very conventional research. It’s not like they have an algorithm that no one else has. What matters is the combination of the algorithm, the UI, and the training data ⇒ this makes it super powerful.
There isn’t a large corpus of publicly available training data for what is addictive to users, unlike other text or image-based AI
How did TikTok get this kind of training data?
- Started with the app Musical.ly. It was launched in both China and US. The clone of Musically by ByteDance – Doujin – ended up taking China by storm
- People creating this new type of video and ByteDance training their data on this helped them create an addictive experience for its users. The combination of easy-to-use creation tools + curation helped them kick ass
- There are also creativity network effects. Every additional creator on TikTok helps the rest of the community become more creative
- TikTok allows you to just remix anyone else’s idea and add your own stuff on top of it. There is shared inspiration in the community
- TikTok also makes it physically possible for you to rip off someone else’s idea with their tools (like the Duet feature)
What about TikTok helped push this remix culture?
- The algorithm helps for this ⇒ it amplifies trends that others are talking about
- TikTok is a mix of a free market and a managed economy in terms of what it promotes
- TikTok’s algorithm has also removed the Old Money effect, where the creators who have been there the longest have a huge advantage over new creators. The main idea is not to follow people. It’s just to see content that you’ll enjoy
- TikTok looks at the interest graph, and bypasses the social graph
TikTok Algorithm and UI ⇒ Algorithm friendly design
- Classify and structure content in units that make it easy to understand by a machine, almost like humans in a census
- TikTok epitomizes the idea of seeing like an algorithm. If the algorithm is going to be one of the key functions of your app, how do you design the app so that the algorithm is able to see what is sees
- ByteDance has a huge ops teams that tags videos with features and attributes (this video has a kitten in it, this video has people dancing in it etc)
- The fact that the only thing you see on the screen is the video (and nothing else) also gives TikTok far more signal salience about what you think about the video. For instance, if you flip past the video very quickly, it knows that you’re not interested. Feedback loop is closed and super quick here. Attribution of your interest to the video is almost perfect
Algorithmic classification of people
- Algorithmic sorting can allow you to classify individuals into one of many categories. This allows you to scale almost instantly
- Instead of asking you to profile yourself, TikTok just asks you to consume stuff and notices your behaviour over time
- TikTok also detects changes in your tastes very, very quickly and responds accordingly. If you’re suddenly interested in soccer videos, it will start showing you more soccer videos almost instantly
What are the implications for algorithmic discovery?
- Could lead to a new kind of social network that is interest-graph based instead of social-graph based
- Can you design a user experience that allows a Machine-Learning algorithm to get access to new and uniques kinds of training data?
How do you see this affecting the future of video?
- We haven’t really figured out video, there’s a lot to be done there
- We consistently under-rate the degree to which we respond to and consume video. Far more people enjoy watching video that reading text
- Before the internet, TV > Radio > Print
- For reaching a broader audience, we don’t have a medium that can challenge video
- In order for video to scale as a medium, we have to solve for a few things. Video is typically a little bit harder to scan for conceptual information quickly (metadata and transcripts can solve this)
- If you can solve this well, video can become more prominent for commerce, education, travel, and other things