Noumad

Noumad is a research lab building reinforcement learning environments where the reward is a measurement of the world, not relative to the capabilities or opinions of humans. We specialize in long-horizon ML, computational science, and embodied AI applications. We do not operate a contractor-based contributor network. We are creating post-training gyms that guide models to outperform human experts rather than matching them. We believe that level of quality requires a dedicated team of researchers and engineers.

Companies like Anthropic, OpenAI, Google DeepMind, and Meta buy large quantities of training data from outside vendors. The major vendors operate contributor networks made up of gig workers. This strategy converts human expertise into training signal at volume. That worked when agents were unable to match human performance. But evals are saturated, and as models continue to improve, a new level of rigor is required - especially in the areas of ML engineering, computational science, and robotics. We're a research lab dedicated to building environments that teach agents to surpass human performance rather than meeting it.