Brendan O’Donoghue describes text diffusion as an iterative process that refines a noisy or masked sequenceA diffusion model generates or transforms data by learning to reverse a gradual noising process, often through a sequence of denoising steps., contrasting it with generating one token at a time. He explains why parallel token updates may reduce latencyLow-latency AI inference produces a model result quickly enough for an interactive or time-sensitive application. while repeated computation can affect serving throughput and cost.
Using demonstrations, O’Donoghue discusses self-correction, adaptive numbers of refinement steps and editing part of a response while preserving surrounding text. Reported quality and speed figures are presented as the speaker’s claims about particular models and hardware.
The talk considers interactive generation, bounded output windows and the applications where low latency might matter. It presents research and demonstrations rather than a guarantee that diffusion is universally faster or cheaper.
Watch on YouTube
