AI Researcher — New York, NY — LinkedIn — 4491465425
Зарплата не указана
ОткликнутьсяAbout 8X9UN
8X9UN is an AI quantitative finance lab advancing how machines understand markets.
Today’s models can explain finance and write research code, but they still struggle to form, test, and revise reliable views of markets. We are building agentic quantitative-research systems, expert-graded evaluations, and reinforcement-learning environments to close that gap.
8X9UN was founded by Jeremie Cohen, formerly a quantitative portfolio manager at WorldQuant and lead data scientist at Millennium. He started 8X9UN to address a problem he experienced directly: models can discuss markets, but they cannot yet conduct rigorous, point-in-time research that survives execution, backtesting, and out-of-sample scrutiny.
8X9UN is backed by Y Combinator, Frst, Robinhood Ventures, and leading operators across AI and finance.
The role
We are looking for an AI Researcher with direct experience post-training open-weight models.
You will own the research and engineering required to improve models on complex, multi-step quantitative-research tasks. This includes training-data design, supervised fine-tuning, preference optimization, reinforcement learning, verifier development, evaluation, and rigorous error analysis.
We already have deep quantitative-research expertise. You do not need to arrive as a finance specialist. We need someone who deeply understands how to change model behavior through training.
What you’ll work on
• Post-train open-weight models using SFT, DPO or related preference-learning methods, reinforcement learning, and verifier-based rewards.
• Build training datasets from research trajectories, expert demonstrations, executable tasks, critiques, rejection paths, and structured feedback.
• Train agents to generate hypotheses, express them mathematically, write and execute code, evaluate results, and revise unsuccessful approaches.
• Design reward functions and verifiers for point-in-time correctness, statistical validity, reproducibility, code execution, and research quality.
• Develop evaluations that distinguish genuine capability improvements from benchmark overfitting.
• Run ablations and detailed error analyses to determine which data, objectives, and training methods drive progress.
• Improve distributed training, inference, data-generation, and experiment-management infrastructure.
• Work closely with our quantitative researchers to translate institutional research workflows into effective training environments.
Your first 90 days
Within 30 days: You will understand akdMV’D research harness and establish baseline performance for selected open-weight models. You will build and benchmark a verifier for a mean-reversion research task, covering temporal leakage, execution correctness, statistical validity, and reproducibility.
Within 60 days: You will complete an initial post-training run, evaluate it against the baseline, and identify the highest-leverage failures across data, reasoning, tool use, and reward design.
Within 90 days: You will be running reinforcement-learning experiments to improve agentic hypothesis generation and research execution, with a clear account of what improved, what failed, and the next research agenda.
What we’re looking for
• Direct experience post-training open-weight language models—not only prompting models or building applications on closed APIs.
• Strong knowledge of SFT, preference optimization, reinforcement learning, reward modeling, distillation, and synthetic-data generation.
• Experience designing datasets and feedback signals that produce measurable changes in model behavior.
• Fluency with Python, PyTorch, and modern open-source model tooling.
• Experience running and debugging substantial training experiments.
• Strong experimental judgment across evaluation design, ablations, and failure analysis.
• The ability to move between research ideas and reliable implementation.
• Comfort owning ambiguous problems inside a small, fast-moving team.
Especially relevant experience
• RL with verifiable rewards or process-level feedback.
• Post-training reasoning, coding, mathematical, or tool-using models.
• Agentic systems operating over long, multi-step trajectories.
• Distributed model training and inference.
• Data curation, synthetic-data pipelines, or rejection sampling at scale.
• Evaluation harnesses, execution sandboxes, or model-based graders.
• Time-series modeling, backtesting, or point-in-time datasets.
• Published research, meaningful open-source contributions, or strong independent experiments.
Prior finance experience is welcome but not required. We care much more about evidence that you have trained models, diagnosed why they failed, and made them meaningfully better.
Why 8X9UN
• Apply frontier post-training methods to research problems with measurable outcomes.
• Work with quantitative experts who can provide high-quality demonstrations, feedback, and evaluation.
• Build from executable research trajectories rather than static financial text alone.
• Own akdMV’D post-training research agenda from an early stage.
• Run fast experiments with substantial technical autonomy.
• Help define how open models learn to conduct rigorous quantitative research.
How to apply
Apply through Bookface with your résumé or LinkedIn profile and links to relevant work.
In a short note, describe a post-training experiment you personally owned: the model, method, scale, evaluation, what failed, and what you learned. Papers, GitHub repositories, technical writing, and experiment reports are especially helpful.