Andrew White

Andrew D. White is a chemical-engineering researcher who co-founded FutureHouse in 2023 and Edison Scientific in 2025. He designed the world model in Kosmos, an AI scientist that combines scientific literature and data analysis to pursue research objectives. With Sam Rodriques, he also helped establish the commercial structure for deploying that work beyond FutureHouse’s nonprofit research program.[5][7][9]

Kosmos’s world model addresses a practical limit of scientific agents: a language model’s finite context can constrain how much information it synthesizes and how many steps it takes. The structured model incorporates information from many agent trajectories while keeping the work directed toward a research objective. In a later technical account, White and his collaborators described its scientific role as representing predictions about experimental outcomes, within a loop of generating hypotheses, proposing experiments and updating beliefs.[7][8]

White’s contribution was part of a divided team effort. The November 2025 Kosmos announcement credits Ludovico Mitchener with leading design and engineering, Benjamin Chang with discovery synthesis, report generation and evaluations, Angela Yiu with academic collaborations, and Jon Laurent with coordinating human evaluation. Michaela Hinks supervised and managed the technical-report work; Michael Skarlinski led platform engineering. Rodriques oversaw scientific work and characterization, and he and White jointly supervised work at Edison. Academic collaborators, principally Mathieu Bourdenx, also contributed.[7]

The developers paired their claims about research depth with a warning about scientific relevance: Kosmos could chase statistically significant but unimportant findings. Their widely publicized work-equivalence estimate came from beta users, with seven scientists estimating an average of 6.14 months of human effort for a 20-step run. It was an estimate of equivalent work, rather than a universal measured speedup, and the announcement warned that deeper runs could pursue spurious correlations.[7]

In How We Built Kosmos, published May 18, 2026, White, Skarlinski, Mitchener, Siddharth M. Narayanan and Tyler Nadolski dated their choice of code-writing agents to late 2025. Earlier specialist agents, including PaperQA and ChemCrow, used hand-crafted tool sets. The coauthors argued that improved LLM code writing now allowed agents to discover and use capabilities through code execution, with integrations maintained as documented software in a monorepo. Their formulation was: “Instead of AI engineering, we simply write quality code with clear documentation.”[8]

The team placed the world model inside a persistent agent that could discuss an approach, respond to feedback and show progress. Their example of the earlier interaction problem was concrete: asking whether one could take two Claritins in a day could trigger an elaborate pharmacokinetic model built from patents and clinical-trial outcomes. The newer design could first talk with the user about how to proceed.[8]

Scientific workloads and internal deployment at partner companies shaped the platform. The coauthors described Kubernetes-managed sandboxes running in containers on virtual machines, with substantial memory, CPU and GPU requirements. Their reported security choices included gVisor, rootless containers, single-tenant virtual machines and credentials injected outside the sandbox as requests left it. Self-hosted packages reduced supply-chain exposure and external traffic; a JuiceFS filesystem let multiple agents share large research files.[8]

White and Rodriques launched Edison Scientific on November 5, 2025 as a commercial spinout of FutureHouse. They explained that product development, payments, market entry and customer support called for commercial funding, while philanthropic funding should support basic research that could not otherwise be funded. They announced that they would initially run both organizations, with controls to manage conflicts of interest and avoid private benefit. Edison’s team page listed White as co-founder and CTO on September 18, 2026.[9][6]

Before founding the AI-science organizations, White worked on computational biomaterials and methods for combining simulations with experiments. He earned a BS in chemical engineering at Rose-Hulman Institute of Technology in 2008, then completed doctoral work at the University of Washington during 2008–13 with Shaoyi Jiang, applying computational modeling to biomaterial discovery and interactions with nature. As a postdoctoral researcher at the University of Chicago, he worked with Gregory Voth on combining simulation and experimental information for biochemistry and materials science.[10]

In July 2014, Rochester announced that White would join its chemical-engineering department as an assistant professor the following January. White cited the university’s research environment, including energy and materials expertise, data science, computing resources and opportunities for medical collaboration. These were his stated reasons at the time; his research then included thermodynamics, materials science and molecular simulation.[10]

By 2023, White’s publications included chemistry and scientific-research agents. He coauthored ChemCrow with Andrés M. Bran, Sam Cox, Oliver Schilter, Carlo Baldassari and Philippe Schwaller, and PaperQA with Jakub Lála, Odhran O’Donoghue, Arsha Shtedritski, Cox and Rodriques. He also published the perspective The future of chemistry is language. His former Rochester laboratory now states that all members have graduated and that he has moved on.[5]