
We're a biotech building AI infrastructure for rare disease drug development and have been using @AnthropicAI models, among others. This week's announcement that Anthropic intends to develop preclinical programs in rare disease to learn and build better products for customers has made me think more deeply about how AI-native companies should architect their infrastructure to take advantage of both open and closed models. I also watched Karp's recent interview and haven't stopped thinking about this: "What the technical customers want is control over their compute, their models, their data stack, and their alpha. They want to know they own the means of production, and it's not being transferred to someone else." It also aligns with @satyanadella recent essays, where he argues that durable advantage won't come from the models themselves, but from a company's domain expertise, evals, and the systems it builds around them. I don't take that to mean companies shouldn't use frontier models. Quite the opposite. Foundation models are advancing at an incredible pace. But I do think it means companies need to be deliberate about what they own. Their data, workflows, institutional knowledge, and evaluations should for the most part remain under their control, while their infrastructure stays flexible enough to use the best model for each task. Recent examples like Fable's use in biology reinforce that it's unlikely a single model or provider will be the best choice for every task. The ability to combine both open and closed models will become increasingly important. At the same time, Anthropic's announcement feels like validation of something we've believed at @MendraInc from the beginning: rare disease is an important frontier for AI and one that deserves far more attention. There are more than 10,000 rare diseases, most still without effective treatments. More attention, more models, and more companies building in this space is ultimately a good thing.
















