See full schedule

Over the past year, the dominant AI conversation centered on the advancements of large language models. But a parallel shift is unfolding in generative visual models with its own distinct set of challenges. Text is rich in semantics but capped there. Pixels carry far more raw information, but extracting useful signals at scale is far less straightforward. 
In this fireside, Tim Dockhorn will unpack what it actually takes to train and deploy frontier multimodal systems in 2026. He’ll discuss how Blackwell-class hardware is reshaping model architecture and scaling decisions; why image generation remains far from a solved problem; and how visual pre-training is evolving into a foundation for more advanced capabilities, from video understanding to robotics and world models.

What it actually takes to train frontier models

Speakers