AI Daily — May 22, 2026
2026-05-22
www.technologyreview.com
Anthropic's Code with Claude showed off coding's future—whether you like it or not
Anthropic held a two-day developer event in London called Code with Claude, coinciding with Google I/O, where the focus was on AI-generated pull requests and autonomous coding workflows. The event highlighted a growing expectation that fully AI-written code contributions are becoming routine for working engineers. MIT Technology Review's on-the-ground coverage captures both enthusiasm and unease among attending developers.
rss.arxiv.org
Provably Learning Diffusion Models under the Manifold Hypothesis: Collapse and Refine
Researchers propose a theoretical explanation for why diffusion models efficiently learn score functions without suffering from the curse of dimensionality when data lies on low-dimensional manifolds. They identify a 'collapse-and-refine' mechanism: at small noise scales the score's singularity collapses the denoising map onto the data manifold, while moderate noise scales refine the intrinsic density. This is operationalized as Score-induced Latent Diffusion (SiLD), a two-stage framework that unifies manifold learning and density estimation.
rss.arxiv.org
Lens: Rethinking Training Efficiency for Foundational Text-to-Image Models
Lens is a 3.8B-parameter text-to-image model that matches or outperforms models exceeding 6B parameters on standard benchmarks while using only ~19% of the training compute of comparable systems like Z-Image. Key efficiency gains come from training on a curated 800M-image dataset with GPT-4.1-generated dense captions averaging ~109 words each, providing substantially richer semantic supervision per batch. The result is a notable data-centric argument for training efficiency over raw model scale in generative image models.