Text Diffusion — Brendon Dillon, Google DeepMind

TL;DR

• 1

  Text diffusion trades token-by-token generation for block refinement: Instead of predicting one token at a time like GPT or Gemini, the model starts from random tokens and denoises an entire span over multiple passes, which lets it attend to future tokens and revise earlier text.

• 2

  The latency win comes from hardware, not magic: Dillon explains that GPUs and TPUs are memory-bound for autoregressive decoding, so generating 256 tokens in 24 denoising passes can mean roughly 10 times fewer memory transfers than emitting 256 tokens one by one.

• 3

  Gemini Diffusion reportedly matched Gemini 2.0 Flash quality with much better latency: In DeepMind's research preview from last year, the model served around 2,000 real tokens per second and stayed broadly competitive on quality, with some strengths in code.

• 4

  Bidirectional reasoning lets the model fix itself mid-answer: On a math prompt whose answer is 39, the diffusion model first wrote 60, then 49, then corrected the full solution and updated the opening answer, while GPT-4o and Gemini 2.5 Flash each made mistakes on the same problem at the time.

• 5

  Adaptive compute is built into the generation process: The model can stop early on easy tasks like outputting the first 100 digits of pi in 4 steps, take 18 steps for FizzBuzz, and spend 31 steps explaining quantum mechanics, with the stopping behavior learned by the model itself.

• 6

  Low latency enables weird, compelling products: Dillon shows prototypes where HTML, text, comments, and UI states are generated live, including fake Wikipedia and Reddit pages, a fully generated operating system, and a voice-driven coding demo that built and restyled a to-do app in about 15 seconds.