Evaluation Engineer

ABOUT ELICIT Elicit is an AI research platform that uses language models to help researchers figure out what's true and make better decisions, starting with common research tasks like literature review. What we're aiming for: 1. Elicit radically increases the amount of good reasoning in the world. - For experts, Elicit pushes the frontier forward. - For non-experts, Elicit makes good reasoning more affordable. People who don't have the tools, expertise, time, or mental energy to make well-reasoned decisions on their own can do so with Elicit. 2. Elicit is a scalable ML system based on human-understandable task decompositions, with supervision of process, not outcomes. This expands our collective understanding of safe AGI architectures. Visit our Twitter https://twitter.com/elicitorg to learn more about how Elicit is helping researchers and making progress on our mission. THE MISSION OF ELICIT EVALS Some orgs build evals to warn us about dangerous capabilities. Some build evals to understand trends and predict where models are heading. Some build evals to hill-climb toward models that users will like more. At Elicit, we're after something different. We want to understand, and hill-climb toward, models that help us make better decisions. This is harder than "what will users like better." Decision support is difficult to evaluate, and users' knee-jerk reactions don't always track with what actually helps them decide. Because it's hard, and because the sales pitch is more complicated, few are doing it well. If we get this right, we have a real shot at pushing AI toward better decision-making, both inside Elicit and beyond. WHY WE'RE HIRING FOR THIS ROLE We need someone to own the technical foundation of our auto-evaluation systems. Our evals are much slower than they need to be, and our interfaces aren't built for the range of people who rely on them: ML engineers iterating on models, product managers monitoring quality, and customers assessing how much to trust a result. This role goes beyond building infrastructure. You'll work out what it actually means for Elicit to support decision-making in pharma, and encode that understanding into our evaluation systems. WHAT YOU'LL OWN The core auto-eval platform You'll build a comprehensive system that runs fast, is easy to use, and supports quickly building new evals: - Speed: You’ll build a lightning-fast basic evals infrastructure that schedules tasks to introduce practically no latency; and then you’ll figure out clever ways to solve the fundamental sources of latency (building a version of Elicit, running it on a query, and evaluating it using LMs) - Interfaces: ML engineers need evals to kick off automatically on relevant commits, with results they can see at a glance and drill into. Product managers need dashboards showing performance over time and what's going wrong in production. - Architecture: Your code must be well-architected so other team members and ML engineers can understand and build on it. An engineer starting on a new feature should be able to quickly add examples and run an eval. Ensuring evaluations are accurate and reliable - We need to evaluate how well Elicit actually helps with decision-making in pharma, not just measure what's easy to measure. This requires encoding real knowledge about how pharma customers make decisions (for example, choosing appropriate gold standards). - You'll provide appropriate statistical tests and confidence intervals so we can trust our results. A MONTH IN THE ROLE In a typical month, expect to spend: - 60% working on the core eval platform - 15% working closely with the evals team to build and improve specific evals (e.g., an eval of our paper search within our systematic review flow) - 10% mentoring our evals engineering intern - The rest on learning how people interact with the eval system so you can make it work better for them, and understanding what our users want from Elicit so evals measure what matters WHAT YOU BRING TO THE R