According to research from Google DeepMind, a new text diffusion model called DiffusionGemma is about as transparent as traditional language models, despite generating text in a fundamentally different way. Researchers distinguished between variable transparency—understanding what a model is thinking at each step—and algorithmic transparency, or understanding how it reaches its conclusions. Since diffusion models generate all tokens simultaneously rather than left-to-right, their reasoning patterns can be surprising. One striking example: DiffusionGemma sometimes gets the math wrong initially, then retroactively corrects itself by revising its earlier output in later steps. The research matters for AI safety, since transparency is becoming crucial as models do more reasoning in latent spaces. The team published twenty-four open problems for the research community.

Listen to this story

Hear this and more stories in a personalized audio briefing.

Open The Chonkerton