UNDER THE HOOD: HOW OPENAI'S SORA MODEL WORKS

Summary from the TLDR Web Dev newsletter

OpenAI's Sora model for AI-generated video builds on diffusion models instead of operating on raw pixels. It works in a compressed 'latent space' and uses the Transformer architecture. Like with large language models, throwing more compute at Sora yields better results. Sora is good enough for real-world use, but it is expensive as the inference cost could require hundreds of thousands of GPUs at its peak.