An open Oumi platform for artificial intelligence (AI) researchers has been launched this week by researchers from 13 universities, including Carnegie Mellon University (CMU), Stanford University, and the Massachusetts Institute of Technology (MIT).
The platform is being administered by a namesake public benefit corporation (PBC) with $10 million in seed funding.
Researchers from other universities that participated in developing the platform include University of Illinois Urbana Champaign, Princeton University, California Institute of Technology, University of California, Berkeley, University of Washington, New York University, University of Waterloo, University of Cambridge and University of Oxford.
The overall goal is to provide AI researchers with access to the tools, code, weights, data, benchmarks and foundation models needed to collaboratively conduct AI research and reproduce results, says Oumi CEO Manos Koukoumidis. This should make the field more accessible to a wider range of academic institutions that today may not have the resources needed to research AI.
The approach will promote greater trust in AI systems in an era where much of the research today is being conducted by commercial companies that are training AI models without providing a lot of visibility, he adds. “They tend to be secretive,” he says.
Specifically, AI researchers will be able to train and fine-tune models from 10M to 405B parameters using multiple state-of-the-art techniques, synthesize and curate data, and access inference engines that can be used to deploy AI models in the cloud or on individual laptops.
Additionally, AI research will be provided with access to commercial application programming interfaces (APIs) through which they can access models from, for example, OpenAI, Anthropic or Google. Oumi will also provide access to prebuilt reusable workflows and recipes for post training and other common tasks that AI researchers need to routinely perform.
It’s not clear to what degree organizations might decide to one day mandate reliance on open AI models that are easier to validate and audit, but the one thing that is clear is the gap between open and commercial models in terms of capabilities and performance is rapidly closing. In fact, recent advances in China clearly suggest the amount of code and infrastructure needed to build a functioning AI model may not require billions of dollars in upfront investment.
Also unclear is how AI research may evolve, but the pace is certainly accelerating. New capabilities and techniques are now being adopted that in some cases instantly upend previous initiatives. A more collaborative approach to AI may accelerate breakthroughs in a way that might even prove to be less disruptive and, ultimately simpler to validate as AI regulations are more widely applied. For example, if the foundational model used to distill a smaller model has already been thoroughly tested that’s one less set of tests that may need to be run later.
One way or another, however, researchers typically find a way to collaborate. The Oumi platform simply provides a mechanism for streamlining those efforts in a way that benefits a much wider percentage of the global population sooner than later.
