What Research Could Empire AI Make Possible?

Even with an idea for a new learning method or scientific model, researchers sometimes lack the computing resources to test it adequately. An approach that looks promising in a small experiment may not work in the same way when the data or computation is scaled up. Access to facilities that allow researchers to test that next stage can influence how basic research develops.

New York State's Empire AI is a shared AI computing infrastructure for universities and research institutions. The Simons Foundation's current call supports research that uses these facilities to develop and apply foundational machine learning models to problems in physics, biology, neuroscience, mathematics and other fields. Researchers in Japan also have a route to participate as additional principal investigators, provided they meet the specified requirements, in joint proposals led by faculty members at eligible member institutions. Official call for applications ↗

How far might researchers now be able to pursue ideas they could not test because of computational limits? Research funding and computing allocations, however, involve separate procedures. Before starting a collaboration, I would want to establish what equipment and computing time would actually be available, whether access from overseas is possible and what technical support is provided. I am also interested in whether researchers outside universities and startups have any route to apply directly. Computing allocation guidance ↗

As research progresses, researchers may want to turn their results into a business. With computer science, it is relatively easy to imagine delivering a result to users as a model or software, and thoughts about commercial potential can arise even in discussions of the research itself. Yet the prospect of commercialization does not diminish the scientific value of the problem that has been solved. I see no reason why the desire to investigate an unresolved problem and the desire to put something useful into the world should not coexist within the same research project.

In practice, researchers need to reconcile publishing their work to encourage verification and further development with protecting intellectual property and know-how. Simons' general policies contain provisions that contemplate commercialization; they do not read as a blanket prohibition. At the same time, the policies require the sharing of research data, software and other outputs. How much must be made public, and who holds which rights, need to be clarified, including under the terms of an individual award. General grant policies ↗

Japan's Fugaku offers arrangements for industrial use and fee-based access that allows results to remain confidential. Comparing not only computing performance but also eligibility, the treatment of results, costs and support would make it easier to find a program suited to a particular research project. Fugaku fee-based access ↗

What research will be selected for this funding? If some of those projects involve ideas that could only be tested after gaining access to these computing resources, I would like to follow not just their results, but what the resources made it possible to attempt.