Traverse the pathto superintelligence

We give frontier AI labs training data that enables their models to acquire taste and handle ambiguous scenarios. If you'd like your models to be better at economically valuable work, get in touch.

Contact
Traverse
Backed ByY CombinatorWith Angels FromOpenAIGoogle DeepMindAnthropicMeta
01

Traverse is a research lab that partners with frontier AI labs to produce the training data required for models to develop taste and judgment. Our work focuses on subjective and long-horizon tasks where success depends on reasoning, context, and decision-making. Our long-term goal is to give frontier models the foundations needed to perform and eventually surpass human white-collar work, accelerating the path toward artificial superintelligence.

02

Reinforcement learning recently produced superhuman models in domains like math and coding because these environments are largely deterministic. Most economically valuable work is not as deterministic; the work done in fields such as law, healthcare, sales, writing, and strategic decision-making is inherently ambiguous, where many outputs can be valid and quality depends on judgment, taste, and context. Training models to operate reliably in these environments remains one of the central unsolved problems in AI.

03

We believe non-deterministic work is fundamentally a context problem. Tasks that appear ambiguous become verifiable when enough information is captured about the situation, the constraints, and the reasoning behind expert decisions. Synthetic data produced by experts answering contrived prompts can help models perform adequately, but it cannot produce superhuman capability. What matters is observing real experts operating inside real environments and preserving the reasoning process that leads to their decisions. At scale, this creates training signals that allow models to learn how judgment actually works, and that class of data does not yet exist. We are the pioneers.

04

The name Traverse reflects the nature of the problem we are working on. The path from today's models to systems capable of replacing complex human labor is not a straight line but a vast and largely unexplored landscape. The largest portion of that landscape is the non-deterministic work that current training methods struggle to capture. To traverse is to move through territory others assume cannot be crossed, and to continue forward until humanity reaches superintelligence. Traverse.