Anthropic Launches Two Four-Month Fellowship Programs
Anthropic has opened two fellowship programs: STEM Fellows and Conceptual Reasoning Fellows, recruiting field experts to enhance Claude's reasoning capabilities on the most challenging problems in relevant disciplines.
The programs will run full-time for four months, from January 12 to May 8, 2027, with exceptional performers considered for extension; salaries range from $100,000 to $167,000 for the four months, equivalent to an annualized $300,000 to $500,000, along with other benefits; the default work location is San Francisco, with remote work possible within the U.S. under specific circumstances, but all participants must attend a one-week in-person training in San Francisco.
The STEM Fellows program targets researchers with a Ph.D. in mathematics, physics, life sciences, engineering, etc., who have several years of research experience (such as postdoctoral or faculty positions). They will collaborate in groups to define the hardest unsolved problems in their fields using Claude and internal tools, build datasets and simulation environments, and specifically identify model weaknesses; the first cohort of over 20 STEM Fellows will complete the program from June to September 2026, with their outputs directly used for Claude's training and internal scientific evaluation.
The Conceptual Reasoning Fellows program is aimed at experts in philosophy, cognitive science, economics, policy, and law who hold or are about to complete a Ph.D. (or equivalent research experience). They will assess and improve Claude's reasoning performance in these fields through professional judgment, design systematic evaluations and reinforcement learning environments, and address complex problems requiring expert judgment.
The application deadline is October 28, 2026, with offers expected to be made by November 20; subsequent processes typically include take-home assignments and discussions with research teams and potential mentors about research directions and specific projects; visa sponsorship is not provided, and applicants must already have or obtain a U.S. full-time work permit and sign through an employer record rather than being directly employed by Anthropic.
This program represents an event-driven investment in talent and capability building, with Anthropic attracting external field experts through high-paying short-term contracts, directing funds towards building more challenging evaluation data and training environments to strengthen the model's weaknesses in long-term experimental judgment and normative reasoning; the beneficiaries are Anthropic's own model capabilities and internal evaluation systems, while competitors face pressure in expert-level problem data and interdisciplinary judgment capabilities.
Source: Public Information
ABAB AI Insight
Anthropic has previously run a Fellows program, with the first cohort of over 20 STEM Fellows producing datasets and problem definitions from June to September 2026, which directly entered Claude's training and internal scientific evaluation processes; prior related directions can be traced back to the team's earlier work on GPT-3, interpretability, human preference learning, etc., demonstrating their ongoing path of converting external expert judgments into trainable signals.
Funding and resources are primarily directed towards short-term, high-density expert collaboration, signing fixed-term contracts through employer records, providing access to cutting-edge models and internal evaluation tools, and transforming domain knowledge into specific datasets, simulation environments, and reinforcement learning tests; the motivation is that after the model has strong coding capabilities, the bottleneck shifts to expert judgment on "where the problems are and what the mistakes are," thereby accelerating scientific and conceptual reasoning capabilities rather than merely expanding computational power.
Similar paths can be seen in OpenAI's early systematic alignment through RLHF and human feedback systems, as well as DeepMind's introduction of domain scientists to jointly define evaluations and data in projects like AlphaFold; currently, Anthropic is in a phase of expanding from general capabilities to vertical scientific and normative reasoning control, attempting to master evaluation assets that are harder to replicate publicly by embedding experts internally rather than relying solely on external collaboration.
This essentially represents a restructuring of the industrial chain and a transfer of pricing power: as model capabilities improve, scarce resources shift from computational power and data scale to "expert-level problem definition and failure mode annotation." Anthropic locks in this judgment capability through high salaries for short terms and internalizes it into training and evaluation infrastructure, transforming external expert knowledge into its own non-easily transferable capability barriers, thus reconstructing who controls evaluation standards and progress directions on the AI scientist path.
ABAB News · Cognitive Laws
The stronger the model, the more it lacks people who can pose difficult questions.
Expert judgment = non-scalable training signals.
What is bought with high salaries is not code, but the ability to know where the mistakes are.