Outspent by Big Tech, some academics are focusing on research that requires less computing power, even as they try to build more of it
Universities are in a race for relevance in the field of generative AI as private companies, loaded with talent and pricey chips, drive the conversation.
Outspent by Silicon Valley, some are turning their research focus to less computing-power-intensive areas of artificial intelligence, even as they seek to build additional computing resources capable of powering bigger models.
“Academic institutions are scrambling to get access to compute,” said Hod Lipson, chair of the mechanical engineering department at Columbia University.
While basic university research has been critical to waves of technological innovation, generative AI research has been dominated by private companies, thanks in part to their access to the data and dollars needed to build and train models like OpenAI’s GPT-4 and Google’s Gemini.
OpenAI Chief Executive Sam Altman said last year that training his company’s biggest models cost “much more than” $50 million to $100 million.
The stakes are high as universities compete against technology companies for the type of AI talent that can bring prestige to their computer science programs. Universities have a critical role in the talent pipeline for the tech industry, which has struggled in recent years to find qualified candidates for some jobs.
It is also important for universities to be part of the conversation around generative AI and to help inform how it is used, academic researchers said.
“I think it’s important that industry is involved in this. It’s important that government is involved in this. But to balance both of these forces, we also have to have other people, open source, academia, who have a say in where and how this technology is used,” Lipson said, adding that AI’s potential benefits are broad, ranging from designing better batteries to treating cancer.
The search for computing power
Lipson said universities like Columbia are spending big to build out their computing resources, but they are also looking to other arrangements like sharing resources among universities.
The University at Buffalo campus will soon host a state-of-the-art AI computing center as part of a New York state effort called Empire AI. The Empire AI Consortium includes Columbia, Cornell, Rensselaer Polytechnic Institute and other universities.
“These resources are increasingly concentrated in the hands of large technology companies, who maintain outsized control of the AI development ecosystem,” New York Gov. Kathy Hochul’s office said. “As a result, researchers, public interest organizations, and small companies are being left behind, which has enormous implications for AI safety and society at large.”
The University of Chicago has a relationship with Argonne National Laboratory that allows it to leverage some of the lab’s computing resources, said Hank Hoffmann, chair of the university’s computer science department.
But he added that the school is budgeting to expand its computing infrastructure.
“I definitely think it’s something we should be planning for and making sure we don’t get bitten by it,” he said.
Partnerships can also emerge between industry and academia, especially in tech hubs like Silicon Valley, Boston, the Pacific-Northwest, and Austin, Texas, that owe their existence in part to the cross-pollination of ideas between local companies and universities.
At the University of Washington, for example, some programs let academic researchers also work in industry, giving them access to better resources while the university benefits from keeping them on staff, according to Dan Grossman, vice director of the computer science and engineering school.
Martin Schmidt, president of Rensselaer Polytechnic Institute, said academics could approach companies about working on problems that might be mutually beneficial to solve.
But as more talent joins industry rather than academia, there is concern that companies won’t need universities for some of these partnerships, he said.
“I sort of don’t accept the ‘compete with big tech,’ premise, because I think we’ll never compete with industry in some respects,” Schmidt said. “But the question is, do we have the resources to work on the type of research problems that we think are important to advance the field? And that’s where industry has a lot more resources.”
Redefining university research
At the same time, university researchers are becoming more targeted in their areas of study so they aren’t outpaced by industry researchers.
Kavita Bala, dean of the college of computing and information science at Cornell University, said that means focusing less on building and training large language models and more on developing applications that could leverage LLMs for particular use cases.
“So it’s still the cutting edge, but it’s not building the big, extremely large models, which typically academics don’t have the resources for. It’s more applying to the different application domains where it can have a big impact,” she said.
Armando Solar-Lezama, a professor at the Massachusetts Institute of Technology whose work focuses on leveraging AI for code development, said building LLMs from scratch is simply not feasible in academia.
Instead, he said, students and researchers are focused on developing applications and even creating synthetic data that could be used to train LLMs.
Solar-Lezama is also associate director and chief operating officer of MIT’s Computer Science & Artificial Intelligence Laboratory, a research institute.
Faculty members are purchasing their own servers and chips, he said, but the challenge goes beyond financing.
“Even if you have the money, just getting your hands on top-of-the-line GPUs is actually quite, quite hard,” he said.
