Interview, Fireside Chat, Podcast
Cristos Goodrow: YouTube Algorithm | Lex Fridman Podcast #68
- Recommendation systems are projected to be a highly impactful AI area throughout the 21st century, with the goal of striking a balance between openness and responsibility over the next 20 to 30 years.
- The engineering team, having focused for eight years on balancing similar content with discovery, aims to evolve the algorithm from simple heuristics to fine-grained systems that react to individual situations without fixed rules.
- Future recommendation strategies include collaborative filtering to create a "related graph," clustering videos by language and topic, and automatically detecting multi-lingual users to provide relevant suggestions.
- Quality metrics are expected to shift beyond views and watch time to include user satisfaction surveys with four or five-star ratings, while success is defined by user retention and the ideal outcome of every video receiving a five-star rating.
- To address policy and free speech, the platform intends to evolve rules over time, demoting borderline content rather than removing it and raising authoritative sources for debatable ideological points.
- Efforts to reduce meanness and trolling will continue through comment ranking and blocking features, aiming to eliminate harassment for creators.
- Machine learning systems will utilize signals such as likes, dislikes, subscriptions, comments, shares, and "Don't Want to See This Video" inputs to predict satisfaction and suppress offending content like racy thumbnails or excessive all-caps.
- Human reviewers are instructed to prioritize scientific consensus and expertise while removing biases through diverse backgrounds, working in tandem with machine learning to extrapolate decisions to billions of videos.
- The company acknowledges that removing unfair biases involving protected classes requires hard work and ongoing improvement, though the understanding of video content for summarization and labeling is estimated to be less than 25% complete after eight years.
- YouTube search will leverage Google's technology for syntactic and semantic matching based on user behavior, while automatic clip features and video segment detection are under development to replace manual annotations.
- The algorithm is expected to expand a video's viewership circle based on positive reception until engagement drops, though viral success cannot be predicted ahead of time as it is often identifiable only after the fact.
- Platforms will test every change via A/B experiments lasting one week to several months to measure confidence intervals across hundreds of variables before implementation.
- The platform plans to maintain the connection between fans and creators, demonstrating that creators can take breaks without losing their audience, and aims to replace television with on-demand, tailored content.
- YouTube intends to ensure availability in the virtual reality space if it gains significant interest, though it is not currently a primary focus, while also working to eliminate negative societal consequences.
- The system expects to enable users to discover content they previously did not know existed, such as niche historical footage, and assist in learning for areas with low literacy.