Periodic Labs aims to build a scientific super-intelligence (www.nytimes.com)

A new Silicon Valley start-up, Periodic Labs, has recruited more than 20 researchers who recently left leading AI labs (Meta, OpenAI, Google DeepMind) — in some cases giving up tens or even hundreds of millions of dollars — to pursue what it calls a “scientific super‑intelligence.” Co‑founders Ekin Doğuş Çubuk and Liam Fedus are assembling talent including people like Rishabh Agarwal, who declined a lucrative offer to join Meta’s new AI lab. Rather than chasing abstract AGI, Periodic’s pitch is concrete: build AI systems that accelerate discovery in hard‑science domains such as physics and chemistry. Technically, that means focusing on domain‑aware architectures and pipelines that combine large multimodal models with physics‑based simulation, experiment planning, and closed‑loop lab automation — essentially integrating learning, reasoning, and real‑world experimentation to speed hypothesis generation and validation. The move signals a strategic shift in the AI community from general‑purpose intelligence toward highly specialized, reproducible scientific automation. If successful, Periodic’s work could radically shorten discovery cycles for materials, drugs and climate tech, but also raises issues around model reliability, interpretability, experimental governance and dual‑use risks as automated systems gain more control over real‑world experiments.