How to build a diffusion language model

Let's be honest: the LLM world has become a monoculture. Everyone's building the same autoregressive transformer with a chat wrapper, and the "innovation" is just shoving more GPUs at the same predictive next-word game. The Kuleshov Group's blog post on diffusion language models is a refreshing middle finger to that status quo, offering a concrete path toward models that can denoise text in parallel rather than generating token-by-token like some kind of Victorian telegraph operator. The tutorial smartly breaks down the design space—continuous diffusion on embeddings, discrete diffusion over token categories, and the masked-diffusion hybrids that have been quietly outperforming expectations. What's striking is how much of this feels like unexplored territory compared to the autoregressive juggernaut. The post doesn't pretend diffusion is a solved problem; it acknowledges the trade-offs (slower sampling, trickier likelihoods) while showing that the barrier to entry is lower than most people assume. You don't need a fortress of proprietary infrastructure to experiment with these architectures. What makes this genuinely valuable is the pedagogical clarity. Too many papers are written by people who've forgotten what it's like to not know the field's secret handshake. This post assumes you've got a solid ML foundation but doesn't require you to have memorized the entire diffusion literature. It walks through the code-level decisions—how to handle discrete data in a continuous framework, how to structure the noise schedule, how to evaluate without relying on the same old perplexity metrics—in a way that lets you actually build something. The bigger picture here is the slow erosion of autoregressive dominance. Diffusion models already ate image generation's lunch, and the natural-language community is starting to realize that the sequential-generation assumption is a computational straitjacket. If you care about where the field is heading, this is worth your time. The source at https://kuleshov-group.github.io/blog/blog/2026/how-to-build-a-diffusion-language-model/ is a bookmark-worthy reference for anyone who wants to be ahead of the curve rather than chasing it.