Our thesis
Osseus is a research lab, building towards our north star of biomedical superintelligence. We believe that to do this we need to treat data as a frontier research problem, so we conduct our own research which we will share openly.
We believe improving human health and advancing science are the most important applications of AI. Making progress requires difficult, high-quality, and scientifically meaningful reinforcement learning environments, evaluations and data, which we sell to frontier labs and pharma. We also offer post-training for health AI systems and pharmaceutical applications, adapting models to clinical and scientific tasks.
Capability and alignment
As frontier model capabilities accelerate, alignment becomes increasingly urgent. More capable models need to remain honest, reliable, and responsive to human oversight in unfamiliar situations. This is especially important in medicine and biology, where model outputs can influence decisions that affect human health and safety.
We believe carefully designed medical, biological and other beneficial data, environments and reward signals can improve capabilities in medicine and biology while helping aligned behaviours generalise beyond these domains. Early research supports this, and we’re building towards this across a broad range of models.
More diverse data, better models
Frontier AI models have been trained on medical and biological data drawn from a narrow range of populations, conditions and settings, limiting what they can learn. In human genomics, for example, research has historically relied heavily on people of European ancestry. These gaps can reduce accuracy for underrepresented populations and leave important biological insights undiscovered, limiting opportunities to improve diagnosis, develop treatments and advance our understanding of disease.
We work with partners to build medical and biological datasets spanning different populations, conditions and settings. By broadening the data models learn from and crafting evaluations to assess their capabilities, we aim to improve both their performance and fairness.
A continuous data flywheel
Our partnerships support ongoing medical and biological data collection and let us gather feedback from real-world model deployments. This allows us to identify where models break down in practice.
We use these findings to curate datasets and build RL environments that reveal model limitations and help train models to overcome them. Continued deployment feedback and separately released benchmarks help us evaluate progress and identify the next capability and safety gaps to address.
Our team
We have assembled a team of engineers and researchers with backgrounds at Imperial College London, Apple’s ML and data teams and a joint Harvard–HKU lab. Our work spans AI safety, health AI, robot learning and data systems. Senior clinicians and scientists bring the expertise to turn important medical and biological problems into demanding training tasks.
We are committed to pushing the frontier forward and dropped out of an AI safety for health PhD programme at the University of Oxford to do so. Our goal is to build models that can help people understand disease, make better decisions and advance science.