Anthropic STEM Fellow

Anthropic•San Francisco, CA

2d•$1 - $2•Onsite

About The Position

Anthropic’s mission is to create reliable, interpretable, and steerable AI systems. We want AI to be safe and beneficial for our users and for society as a whole. Our team is a quickly growing group of committed researchers, engineers, policy experts, and business leaders working together to build beneficial AI systems. Program Overview We're working toward an AI scientist — a system capable of the long-horizon reasoning and experimental judgment needed to push the scientific frontier. We believe frontier AI, developed safely, can meaningfully accelerate progress across science, engineering, and beyond. Getting there requires close collaboration with people who have spent years working in these domains. The Anthropic STEM Fellows Program brings students and experts from STEM domains into Anthropic for a few months to work alongside our research teams. Fellows pair deep domain expertise with our frontier models and internal tooling to evaluate, improve, and apply Claude's capabilities in their field. We're looking for Fellows across all different science and engineering domains. Prior machine learning experience is helpful but not required — scientific judgment and a willingness to learn quickly matter more. What to Expect: A concrete, scoped project matched to your background, with direct mentorship from Anthropic researchers Access to frontier Claude models and internal evaluation tooling A cohort of incoming Fellows to collaborate with and connection to Anthropic's broader research community What You'll Work On: Designing rigorous, hard-to-game evaluations that measure whether models can plan experiments, interpret data, and reason about mechanisms in your field, and systematically finding where they're confidently wrong. Identifying or creating data and techniques that target capability gaps. Applying Claude to open problems in your domain, by using various test-time compute strategies and integrating relevant scientific tools/software. Example projects might look like a materials scientist who finds Claude struggling to reason about phase stability, and builds a pipeline producing property prediction tasks to address this gap, or a climate scientist integrating atmospheric modeling tools and building environments that exercise them. Projects will be scoped to ship within the period of the fellowship. Application Process Candidates will need to progress through the following stages: Submit an application Complete an initial technical assessment Complete a take-home exercise Chat with a mentor about research and potential project directions We will accept applications until May 15 and plan to have final decisions by June 1. Logistics: Duration: The program will start June 15 and last until September 15; some fellowships may be considered for an extension. Location: In-person at our San Francisco office. Commitment: Full-time role for the duration of the program. Compensation: Weekly stipend of ~$3,800, access to our offices, and other benefits. Authorization to work in the US required for the duration of the program. The annual compensation range for this role is listed below. For sales roles, the range provided is the role’s On Target Earnings ("OTE") range, meaning that the range includes both the sales commissions/sales bonuses target and annual base salary for the role. Annual Salary: $1 - $2 USD

Requirements

Are excited about the potential for AI to accelerate scientific discovery
Are familiar with using Claude and Claude Code as part of their day-to-day workflow
Are comfortable working in a fast-paced, high-agency atmosphere for a few months and delivering concrete results
Hold (or are completing) a PhD in a STEM field, or have equivalent research experience.
Authorization to work in the US required for the duration of the program.

Nice To Haves

Prior machine learning experience is helpful but not required — scientific judgment and a willingness to learn quickly matter more.

Responsibilities

Designing rigorous, hard-to-game evaluations that measure whether models can plan experiments, interpret data, and reason about mechanisms in your field, and systematically finding where they're confidently wrong.
Identifying or creating data and techniques that target capability gaps.
Applying Claude to open problems in your domain, by using various test-time compute strategies and integrating relevant scientific tools/software.