The ML Technique Every Founder Should Know

Chaubard describes diffusion as a way to learn data distributions, including in high-dimensional settings with limited examples. The basic process adds noise to training data, then teaches a model to reverse that process and recover data.

Chaubard traces progress through changes to noise schedules and training objectives, including predicting the data, the added error, or velocity. He explains that flow matching uses a straight path between noise and data and trains a model to predict the velocity along it. In his example, the training loop is only a few lines, though scaling the model still takes substantial engineering.

The conversation covers diffusion for images, video, proteins, weather forecasting, and robotics. Chaubard says the core procedure can stay the same across domains while the model architecture changes; examples include protein design, GenCast, and diffusion policies for robots. He also notes that changing the number of inference steps from the number used in training can fail, while distillation can produce a model for fewer steps.

Chaubard contrasts diffusion with language models that emit one token at a time, suggesting diffusion can support randomness and generation across larger chunks. He connects this to iterative revision and concept-level thinking, while presenting it as a perspective on intelligence rather than a demonstrated account of how brains work.

For teams training models, Chaubard recommends evaluating diffusion as part of the training process, including as a way to learn a latent space. For teams using models, he advises updating expectations as image generation has improved and considering applications such as robotics, proteins, DNA, and self-driving. He frames time, money, and data as important constraints on scaling these applications.