Today, we’re announcing Edison Advances, a home for our research announcements, technical engineering blog, open-weights models and benchmarks. At Edison we do research, engineering, and application of our AI scientist Kosmos to solve foundational problems across scientific domains. We want to understand how to measure and assess AI scientists. We want to push their ability to work on long-horizon tasks. We want to understand what mistakes they make and how to make them secure and safe for broad use. We want to be the vanguard users and apply AI to make truly novel discoveries.
Our research culture at Edison Scientific is inspired by the pioneering research at FutureHouse to automate and accelerate scientific discovery. FutureHouse was founded in 2023 and became a leader in AI for Science research. FutureHouse built the first superhuman literature search agent, a public platform to run it at scale, and the first modern benchmarks for biology.
We’ve collected our existing research at Edison Scientific into the Advances website and will share more as it comes. This includes benchmarks, open weights models, technical papers describing what we’ve learned, and deep dives into implementation and architectural details of our products.
We are also announcing our first new result under Edison Advances: development of a vision language model for reading molecules. Reading molecules, especially in patents, is a great example of a difficult and meaningful problem for drug discovery that has received little attention. Research staff directed Kosmos to work in an “autoresearch” style, and together they produced a state-of-the-art vision language model able to read the complex molecule diagrams in patents. This result is a concrete example of an AI scientist autonomously speeding up one small part of drug discovery. You can read the full story in our announcement.
If you think Edison Scientific sounds like a place for you to be, consider applying! We’re hiring on engineering, AI research, biology, chemistry and more.