Products
Everything we build begins with one principle: quality. Raising AI for the real world demands nothing less.
RL Environments and Agents
Practice tennis strokes all you want, but real mastery requires real gameplay. We create rich, complex RL environments that challenge agentic models in novel ways, and design verifiers that reward their behavior.
Rubrics and Verifiers
Some things can be scored: a biography about Shakespeare earns +5 points for tracing his path from Stratford to the Globe, and −5 for missing his mystery. This can be hard: how do you design a grading system that differentiates between Dicksinson’s poetry and your English teacher’s? Rubrics serve as scorecards; we design them to encompass the breadth of both brilliance and deficiency.
RLHF
Children (and adults and AI) learn from preferences and rewards: praise a ballet dancer when her movements flow, and she’ll become better and better. The RLHF data we generate embodies the richnesses and subtleties of the world — the notes that separate the ordinary from the inspired, Salieri from Mozart.
SFT
Before a model can learn from preferences or rewards, it needs a foundation of skills. We bootstrap model capabilities through demonstrations: teaching them to use computers, navigate the web, and reason for the first time.
Human Evaluation
Academic benchmarks are easily gamed; auto-evaluations are flawed proxies. Human evaluation provides the gold standard against which we measure all else: does this answer make sense, is it useful, is it safe? Humans can judge subtlety, wit, and emotions in ways no formula can.
Expert Professional Domains
There’s no substitute for expertise. We look for the most brilliant minds on the planet in every domain – doctors, lawyers, investment bankers, Fields Medalists, Harvard professors, and more across STEM and the humanities. They shape AI models through both theory and real-world judgment.
Internationalization
We operate in over 70 languages and growing — teaching AI not just language, but cultural values. Our linguists design data that reflect each language’s grammar, idiom, and worldview. A Korean legal summary, a Brazilian news headline, an Arabic dialogue.
Multimodal
Text alone can’t capture human experience. We teach AI to see, watch, and hear — teaching them to generate and understand images, audio, and video.