GPT-4o answered 40. Gemini 2.5 Flash answered 42 and stuck to it even after working through the reasoning incorrectly. The Gemini Diffusion model, considerably smaller than both, answered 60 on the first forward pass, then 49, then corrected itself to 39 once it finished reasoning. Bidirectional attention means it can see future tokens and go back to fix mistakes. Autoregressive models cannot do that.

Brendon Dillon covers why text diffusion is fast (24 denoising steps to generate 256 tokens means roughly 10x fewer memory transfers than autoregressive generation), what the tradeoff is (lower throughput at large batch sizes makes it expensive to serve at scale today), and what gets unlocked when latency drops to 2,000 tokens per second. The demos include a fake Wikipedia generated on the fly, a Reddit clone with AI generated comments and images, an operating system where every click generates the next screen, and a todo app built in 15 seconds by voice.

AI Engineer

A Google DeepMind research scientist presents text diffusion as an alternative to autoregressive language model generation. Unlike token-by-token autoregressive models, text diffusion initializes a full sequence as random noise and iteratively denoises it over multiple forward passes. Key advantages include significantly lower latency (up to 10x faster due to GPU/TPU memory bandwidth efficiency), bidirectional attention enabling self-correction during generation, adaptive computation where harder problems automatically use more denoising steps, and in-place editing capabilities. The main drawback is lower throughput for large batches, making it expensive to serve at scale. The Gemini Diffusion research demo achieved ~2,000 tokens/second. Demos showcase real-time Wikipedia generation, fake Reddit page creation, an AI-generated operating system, and live voice-driven coding — all enabled by the low-latency properties of diffusion models.

Text Diffusion — Brendon Dillon, Google DeepMind