Universities Are Woefully Under-Resourced For AI Research. They’re Fighting To Change That. | Commentary

In front of a packed room inside the Dirksen Senate Office Building in Washington, D.C. on Tuesday, the leadership of Stanford’s Institute for Human-Centered AI made a plea.

Co-director Fei-Fei Li and executive director Russell Wald told the approximately 200 congressional staffers gathered that universities today simply don’t have the resources to do basic generative AI research. The chips, data centers, and energy costs are not in scope for university budgets. And they need serious help.

“All U.S. universities combined could not build a version of ChatGPT right now,” Wald told me in an interview this week. “That’s pretty problematic.”

The dire situation is reaching a boiling point. Microsoft is aiming to have 1.8 million GPUs by the end next month while Stanford University has approximately 300 of these chips, which are the critical components for testing and training generative AI.

Wald then recited some stunning statistics showing how far universities have fallen behind the private sector in an area where they previously led.

In 2022, Wald said, there were 22 significant AI breakthroughs that came from industry, compared to only three from academia. In 2011, AI PhDs would go into private industry and academia in about equal numbers — now 70% go into private industry, he said. Professors and grad students interested in doing cutting-edge research also struggle to choose universities because the resources available will not allow them to do the work.

“It’s deeply concerning about the future of the technology,” Wald said, “because it begs the question — Does the academy belong in frontier AI research? We at HAI would argue yes it does.”

University technology research, Wald said, is necessary because it’s freed from the intense product-focused work that is done within companies, helping it spark fundamental breakthroughs like GPS, MRIs, and the internet. “These are things that no investor would ever put any money into, because they wouldn’t see a return on investment,” he said. “They’d be dead by the time they saw any really true profitable margin out of this.”

Wald also noted university AI research can help train regulators to understand AI, so that knowledge extends beyond those within the companies themselves. And it can serve the public good by publishing research and open-sourcing models given that companies like OpenAI and Google have stopped sharing their cutting-edge research as the field has grown ultra-competitive.

Solutions

So, universities have a serious AI resource problem right now. Along with not having the money to invest in AI data centers (an NVIDIA GPU can cost up to $40,000 per chip) they are facing challenges in setting up facilities with enough space and cooling capacity to operate. And the energy costs add up as well.

But seeing the problem, some academic institutions such as Stanford and a new group in New York are fighting to fix it.

In New York, a first-of-its-kind coalition of universities, the state, and private philanthropy has just built an artificial intelligence computing center that started running experiments last week. The initiative, called Empire AI, is set up as a shared resource at The University of Buffalo and includes NYU, Cornell, Columbia, RPI, SUNY, and CUNY. The group received a donation of GPUs from the Simons Foundation, and support from Empire State Development.

“We’re just in the process of testing Empire AI, we’re starting to run the first jobs,” said Stacie Grossman Bloom NYU’s chief research officer, vice provost, and vice chancellor for global research and innovation.

This week’s gathering with Stanford HAI leadership and congressional staffers in Washington (Source: Stanford University)
This week’s gathering with Stanford HAI leadership and congressional staffers in Washington (Source: Stanford University)

NYU researchers are already using the machine to detect deep fakes. The university, Grossman Bloom said, has 19 work order requests in, with more to come. “The idea is for the machine to not be idle at all. But to really run it, run it hard and, run it at capacity,” she said.

Empire AI currently has 96 NVIDIA H100 GPUS, which, even at a multi-million dollar value, don’t come close to what companies like Amazon, xAI, Meta, Microsoft, OpenAI are using to build, interrogate, and run their models. But the coalition does plan to meaningfully expand the compute resources within Empire AI as it moves beyond its “alpha” phase.

Such initiatives ultimately won’t put universities in position to do foundational research without a massive increase in resources — and that’s where the federal government might come in.

Stanford’s Wald and Li were in Washington advocating for Congress to pass the “Create AI” act, a bill that would make permanent a shared resource that would enable large scale AI research. This “National Artificial Intelligence Research Resource,” which has a pilot established by the Biden administration through executive order, would make compute and data sets available to universities and students to run AI experiments.

The U.S. government and partner companies like NVIDIA and Microsoft have already put tens of millions of dollars into the project, but the bill would establish it in law and help it get funded through federal government appropriations, via the National Science Foundation. The current funding number discussed within Congress is about $400 to $500 million per year, Wald said, over the course of six years. It would be a paradigm-changing bill if passed.

Wald said that he and Li set up this week’s gathering with Congressional staffers last Friday. Initially, only 10 signups came in and they began to worry. But the event became so popular that a line wrapped around the hall as people scrambled to get in.

The Create AI bill has momentum within Congress today. There’s bipartisan support in both houses. Seventy-eight companies, universities, and institutions, from Princeton to Google, just wrote leadership this week supporting it. And it could pass during the lame duck session.

But Wald said the bill’s fate hinges on legislators agreeing to get it done in a very busy period, effectively giving it a six-week window before Congress turns over. “It is probably one of the top easy things to pass on any AI agenda,” Wald said. And now, the only question is whether the will is there to do it.

This article is from Big Technology, a newsletter by Alex Kantrowitz.

The post Universities Are Woefully Under-Resourced For AI Research. They’re Fighting To Change That. | Commentary appeared first on TheWrap.