Radical Numerics

Radical Numerics is a research-focused company developing next-generation artificial intelligence for scientific applications. Based in Menlo Park, California, and founded in 2025, the startup spun out of Stanford University with the mission to build AI that can design and improve itself. The company was co-founded by Michael Poli (CEO), Stefano Massaroli, Eric Nguyen, and a fourth unnamed founder. Poli, a Stanford PhD, and Massaroli knew each other from their hometown in Italy, while Nguyen was also a PhD candidate at Stanford. The founding team includes researchers from institutions like DeepMind, Meta, and Liquid AI.

Radical Numerics operates at the intersection of AI and life sciences, building multimodal models trained on biological data to accelerate discovery. The team is noted for its work on Evo and Evo 2, generative genomics models used to create gene-editing tools and the first AI-designed whole genomes. This technology demonstrated that AI can be used to create new biology, not just analyze it. Their work on a model named Evo, which was trained on DNA sequences to design a novel CRISPR-Cas system, was featured on the cover of Science magazine. Another creation from their models was a phage, a type of virus that only attacks bacteria, illustrating the technology's potential as an antibiotic.

The company develops foundational architectures and numerical computing frameworks to represent complex physical systems, such as in climate modeling, materials discovery, and robotics, as learnable world models. They focus on creating AI systems specifically for science, arguing that architectures designed for language are not suited for understanding DNA or physical systems. Radical Numerics has released open-source models, including RND1-Base, a 30-billion parameter diffusion language model, to encourage further research in the community. By concentrating on systems and architecture innovations, the firm aims to enable AI to autonomously design AI, thereby speeding up scientific and industrial advancements.

Keywords: scientific world models, generative genomics, AI for science, multimodal models, foundation architectures, numerical computing, biological data, AI-designed genomes, deep signal processing, diffusion language models, AR-to-Diffusion conversion, scientific discovery, computational biology, machine learning, Evo model, RND1 model, automated AI research, recursive self-improvement, complex physical systems, life sciences AI