Applied Research - Scientific AI

Prime Intellect•San Francisco, CA

5d•$150,000 - $300,000

About The Position

Prime Intellect is building an open superintelligence stack and a full-stack platform for post-training frontier AI models. In this externally facing Applied Research role, you will build and post-train agents for scientific research, including searching literature, writing and running code, using domain tools, and selecting experiments. The work spans autonomous research and computational science, using environments and evaluations shaped by research partners' tools, data, and experimental feedback. You will develop long-horizon scientific research environments, trainable tasks, evaluations, reward signals, and verifiers; connect automated checks and expert feedback to training; and develop and evaluate RL and multi-agent methods. You will maintain a substantive research agenda while working directly with customers, users, or research partners. The role also involves releasing reusable environments, evaluations, and research code, with a focus on reproducibility and careful experimental assessment.

Requirements

Depth in RL, agents, or post-training systems, and the ability to acquire relevant scientific context.
Hands-on experience building and debugging model-training, post-training, agent, evaluation, or closely adjacent experimental loops.
Strong experimental judgment, including meaningful baselines, careful trace inspection, clear failure attribution, and willingness to change direction when evidence changes.
Strong Python skills and the ability to build artifacts that other researchers and engineers can run, inspect, and extend.
Comfort owning ambiguous problems and keeping partners and teammates current while resolving them.

Nice To Haves

Prior work in autonomous research, ML for science, scientific computing, or reproducible computational research.
Publications, open-source systems, benchmarks, or shipped products demonstrating exceptional work.

Responsibilities

Build long-horizon environments where agents search literature, query scientific databases, run code and domain tools, and choose experiments within a fixed compute or experimental budget.
Turn research partners' workflows, datasets, and outcome criteria into trainable tasks, evaluations, and reward signals.
Design verifiers for correctness, novelty, and reproducibility, and test whether agents rely on underlying evidence rather than label leakage, lookup shortcuts, or unsupported conclusions.
Connect programmatic checks, simulations, domain models, and expert feedback to training; account for noisy or delayed results, calibrate confidence estimates, and define when human review is needed.
Develop RL and multi-agent methods for proposing, testing, and critiquing hypotheses; compare them with strong single-agent baselines at equal compute and evaluate on held-out problems.
Keep code, data, logs, and experiment configurations reproducible; rerun promising results, measure noise, and use ablations to distinguish model improvements from changes to data, harness, or inference setup.
Release reusable environments, evaluations, and research code.
Work directly with customers, users, or research partners while maintaining a substantive research agenda.