Agent Post-Training, Personality

Responsibilities

• Develop a rigorous understanding of what makes an agent a great collaborator across professional, creative, technical, and everyday work.
• Turn qualitative judgments about model behavior into concrete hypotheses, evals, graders, and training interventions.
• Study explicit and implicit user signals to understand which behaviors create trust, satisfaction, continued use, and successful outcomes.
• Work with human experts and trainers to produce high-quality, tasteful rollouts and preference data that capture excellent collaborative behavior.
• Improve reward models and RL objectives for model behaviors.
• Work with pretraining and early-training teams on data mixtures, objectives, synthetic data, and other upstream choices that shape downstream personality.
• Build sustainable pipelines for updating older training data as our understanding of excellent model behavior evolves.
• Partner closely with ChatGPT, Codex, and other product teams to turn consumer insight into model improvements and validate them in real workflows.
• Own projects end to end, from observing a subtle behavioral failure through experimentation, training, evaluation, and launch.

Requirements

• Strong technical foundations in machine learning, software engineering, statistics, behavioral science, HCI, or a related field.
• Strong taste for model behavior: can look at user feedback and explain why one response feels thoughtful, natural, and useful while another does not.
• Experience with LLMs, post-training, RL/RLHF, reward modeling, evals, synthetic data, pretraining data, or production ML systems.
• Ability to work effectively with researchers, engineers, product teams, designers, domain experts, human-data teams and safety boundaries, and communicate clearly with each group.
• Ability to translate subjective-seeming product questions into falsifiable hypotheses and rigorous evaluations without losing the nuance that made the question important.
• Care about preserving individuality, adaptability, and behavioral diversity rather than optimizing every model toward one narrow style.
• Excited by ambiguous capability problems where the signal is noisy, the failures are qualitative, and the solution may involve data, training, evals, product changes, or all of the above.
• Like building load-bearing systems and processes when that is what the team needs, even if the work is not glamorous.

Nice-to-Haves

• Think instinctively from the user’s perspective and care deeply about how models feel to work with, not only how they perform on benchmarks.
• Want to shape how frontier agents communicate, collaborate, and build trust with millions of people.
• Want to train and ship the models that make agents genuinely useful for developers, enterprises, researchers, and everyday users.