Brendon Dillon from Google DeepMind presented text diffusion models, which generate text by iteratively refining random token sequences, enabling bidirectional reasoning, self-correction, and fast in-place editing with significantly lower latency compared to traditional autoregressive models. Despite current challenges in throughput and deployment costs, these models demonstrate promising capabilities for interactive applications, such as dynamic content generation and adaptive computation, pointing toward future advancements in AI text generation.
Brendon Dillon from Google DeepMind presented on text diffusion, a forward-looking research area that adapts diffusion models—commonly used in image and video generation—to text generation. Unlike traditional autoregressive models that generate text token-by-token, text diffusion starts with a sequence of random tokens and iteratively refines them over multiple denoising steps to produce coherent text. This approach allows the model to attend bidirectionally across the entire token sequence, enabling self-correction and reasoning over future tokens, which is not possible in standard autoregressive models.
One of the key advantages of text diffusion models is their significantly lower latency during inference. This is due to the way GPUs and TPUs handle computation: autoregressive models require many sequential memory transfers for each token generated, whereas diffusion models generate many tokens simultaneously over fewer passes, reducing memory bandwidth bottlenecks. Although diffusion models have lower latency per query, they currently suffer from lower throughput when serving large batches, making them more expensive to deploy at scale compared to autoregressive models.
Text diffusion models also exhibit unique capabilities such as bidirectional reasoning and self-correction. Dillon demonstrated how the model could iteratively refine its answers to complex problems, correcting initial mistakes by revisiting and updating earlier tokens based on later context. This contrasts with large autoregressive models, which often commit to early errors. Additionally, diffusion models support adaptive computation, dynamically allocating more denoising steps to harder problems and fewer to easier ones, improving efficiency and output quality.
Another notable feature of text diffusion is its ability to perform fast in-place editing. Because the model processes the entire token sequence simultaneously, it can make localized changes—such as fixing bugs in code or adding paragraphs to a story—while maintaining consistency with the surrounding context. This contrasts with autoregressive models that generate text strictly in sequence and cannot easily revise previous tokens without regenerating large portions of text.
Dillon showcased several innovative demos leveraging the low latency and flexible generation capabilities of text diffusion models, including dynamically generated Wikipedia pages, Reddit threads with AI-generated comments, and even a fully AI-generated operating system interface. These examples highlight the potential for new interactive applications and user experiences unlocked by diffusion-based text generation. While challenges remain, particularly around throughput and serving costs, the research points to promising directions for future AI systems that combine speed, reasoning, and flexible editing.
