Europe 2026
Text Diffusion — Brendan O’Donoghue, Google DeepMind
Overview
This talk introduces text diffusion, a forward-looking research area in AI that offers an alternative to traditional autoregressive text generation. Unlike models that generate text one token at a time, text diffusion models start with a noisy sequence and iteratively refine it to produce clean output. This approach promises significant advantages in inference speed and enables unique capabilities such as bidirectional reasoning and in-place editing.
Who should watch
- AI engineers exploring new text generation paradigms.
- Product managers and builders seeking to leverage low-latency AI for novel applications.
- Researchers interested in the trade-offs between autoregressive and diffusion models.
- Developers looking for techniques to improve model self-correction and adaptive computation.
Key takeaways
- Text diffusion models iteratively refine a noisy sequence of tokens to generate text, contrasting with the one-token-at-a-time approach of autoregressive models.
- A primary advantage of text diffusion is significantly faster inference latency due to more efficient hardware utilization, enabling new real-time applications.
- These models can perform bidirectional reasoning, attending to future tokens to self-correct errors, a capability lacking in standard autoregressive models.
- Text diffusion allows for dynamic and adaptive computation, where models can spend more time on complex problems and less on simpler ones, optimizing resource use.
- The iterative nature of diffusion enables fast in-place editing of text, allowing for modifications that are consistent with the surrounding context.
- While offering lower latency, text diffusion currently faces challenges with lower throughput for large batches, making it more expensive to serve at scale compared to autoregressive models.
- Early research demos showcased capabilities like on-the-fly generation of entire web pages and operating systems, highlighting the potential of low-latency AI.
Notable quotes
*The main pro that people really like and probably is the biggest advantage that text-to-image models have is that it's just faster inference.*
*This ability to do bidirectional reasoning. So, to to not only see the past, but also see the future that it's going to utter.*
*Low latency models can really unlock some new experiences for users and new products.*
Unofficial community note. Prefer the recording for nuance.