This is a heavily interactive web application, and JavaScript is required. Simple HTML interfaces are possible, but that is not what this is.
Post
mr. TIM
timkellogg.me
did:plc:ckaz32jwl6t2cno6fmuw2nhn
pretraining scaling laws for diffusion models
- training data & parameter size scale roughly 1:1
- diffusion tolerates overtraining very well
- overtraining a small model produces same results as a larger model
so diffusion has real advantages over LLMs, theoretically
arxiv.org/abs/2608.17286
2026-08-19T10:32:08.966Z