Where Data Science Meets Shrek: How BuzzFeed uses AI

The Challenge of Content Classification

Understanding what content is truly about goes beyond simple named entity recognition. BuzzFeed publishes diverse content that often defies traditional categorization - like an article about an Indiana woman who brought her pet raccoon to a fire station fearing it had overdosed on marijuana. These unique stories require sophisticated understanding beyond standard topic taxonomies.

From Universal Sentence Encoders to Modern Embeddings

BuzzFeed transitioned from using Google's Universal Sentence Encoders and DistilBERT (which maintained 97% of BERT's performance while being 60% smaller) to modern embedding approaches. They now use Nomic embeddings with 124 million parameters to create dense vector representations of their content.

These embeddings serve as foundational infrastructure for:

• Content clustering and similarity detection
• Recommendation systems
• Trend analysis and business intelligence
• Understanding audience consumption patterns

The Viral Barbie Dreamhouse Experiment

BuzzFeed's editorial team began experimenting with Midjourney to create images relevant to trending topics. Sarah, a content creator, researched and crafted prompts representing each U.S. state to generate "Barbie's Dreamhouse in Every State." The post went viral organically on Instagram and TikTok, with one TikTok video alone garnering over 13 million views.

Building Custom Image Generators

Following early successes, BuzzFeed developed in-house capabilities using:

• Stable Diffusion XL (SDXL) models
• Low-Rank Adaptation (LoRA) for efficient fine-tuning
• APIs like Replicate for model hosting

This technology powers interactive generators including:

• Shrek Generator: Allows users to "Shrek-ify" anyone
• Moo Deng Generator: Creates images of the viral baby hippo in various scenarios
• Mormon Wives Generator: Face-swaps users into themed content
• Medieval Pet Generator: Transforms pet photos into medieval artwork
• AI Emoji Contest: Enables users to create custom emojis

Historical A/B Testing Data as Training Material

BuzzFeed has collected years of A/B testing data from their Bayesian-based testing system, which evaluates different headline and image combinations. This historical dataset includes:

• Multiple headline variants
• Performance metrics (click-through rates)
• Winning combinations

AI-Powered Headline Generation

The team trained a model using Hugging Face's Accelerate and Transformers library to predict winning headlines. The workflow involves:

1. Generating 16 candidate headlines using Claude
2. Running headlines through the trained model
3. Using a "Battle Royale" style competition where groups of four compete
4. Identifying the predicted best-performing headline

This approach combines BuzzFeed's unique historical data with large language model capabilities to create diverse, high-performing headlines while helping writers learn new approaches.