AI

Post-training climbed maths, code and tool use on process data written by humans or synthesised under their direction. The research rung has no such corpus. To keep improving, models have to do science — which makes discovery AI's own bottleneck, not an application of it.

This article analyzes why Diffusion LLMs scale less efficiently than Autoregressive (AR) Models, focusing on two core issues: lack of KV caching and intractable sequence likelihood. 👉 Read the full post on Notion: Inherent Limitations of Diffusion LLMs — Notion Page