Descriptions:
Andreas Stuhlmüller and Jungwon Byun, co-founders of Elicit, join the Cognitive Revolution podcast to discuss how their AI platform for scientific research is evolving from process supervision toward a new concept they call external world models. Elicit — which now works with seven of the top 20 life sciences companies on tasks ranging from drug target ranking to regulatory defense of launch pricing — originally bet that rewarding models for step-by-step reasoning quality would produce more reliable outputs than training on final answers alone. The challenge: powerful frontier reasoning models increasingly hide their chain of thought.
Their solution is a domain-specific language (DSL) that defines reasoning primitives as discrete microservices, allowing frontier models to dynamically compose structured workflows guaranteed to execute as defined. The founders explain why LLMs are still too easy to manipulate for high-stakes decision support, and introduce the concept of “certificates of reasoning” — verifiable proof that prescribed reasoning steps were actually carried out. They also discuss their internal automation system “the line,” which now delivers 30–50 code changes per week, their token spending as a company, and where Gemini fits in their stack.
The conversation ranges into epistemological territory — how to reduce hard-to-verify tasks to easy-to-verify ones, and why external structured representations of evidence may ultimately be more reliable than in-weights learning. Essential listening for anyone building AI into regulated, high-stakes workflows.
📺 Source: Cognitive Revolution “How AI Changes Everything” · Published June 17, 2026
🏷️ Format: Podcast
