Last week, dozens of academic AI researchers gathered in Mountain View, California, for a Schmidt Sciences AI2050 convening that included roundtable interviews and media training. The AI2050 program, funded by Eric and Wendy Schmidt, supports academics whose work involves artificial intelligence. The author notes receiving a Schmidt Sciences–funded science communication award in 2024.
Participants described a difficult moment for university-based AI research. Over the past four years the field has re-centered around large language models (LLMs), and the research frontier has shifted from academia to private companies. Universities often lack the funds and infrastructure — especially GPUs — needed to train and run state-of-the-art models. Even with compute access, firms such as Anthropic and OpenAI do not disclose the internal design and training details of systems like Claude or ChatGPT.
Nika Haghtalab, a computer science professor at the University of California, Berkeley, compared the situation to a biologist working in a world where private companies have exclusive control over the gene-editing tool CRISPR. Outside experts can study how models like ChatGPT and Claude behave, but they cannot perform detailed research on model design or training, nor directly steer those processes.
The AI2050 program provides some funding that fellows can use to buy GPUs; several attendees said that was an important benefit. Still, money remains a pressing concern, especially given reductions in federal scientific funding in the United States. For researchers who do not run local models, the repeated cost of querying OpenAI’s, Anthropic’s, or Google’s models to study them rigorously can be prohibitive.
Many fellows therefore focus on research questions they believe tech companies are unlikely to pursue. Anjalie Field, a computer science professor at Johns Hopkins University, said she tries not to work on problems she expects a tech company will solve. Companies need to make money, and questions with limited profit potential — or answers that might cast companies in an unfavorable light — are less likely to attract industry investment. Field recently found in a study that language models give less sophisticated responses to prompts phrased in ways more commonly used by women than by men, a result that would be hard to imagine coming from Anthropic or OpenAI.
A large group of academics work on non-LLM AI altogether. These researchers build specialized models to analyze data, make predictions, or simulate physical systems. They are not always in direct competition with frontier labs — for example, Google DeepMind’s AlphaFold team, known for its model that predicts protein structures, was disbanded last month — but they face their own challenges. Several attendees voiced concern that the broad lack of understanding about non-LLM AI hampers their work: researchers building AI tools for climate applications, for instance, sometimes struggle to argue for their projects in an environment where “AI” often conjures energy-intensive LLMs.
Those pressures are reshaping academia: several prominent academics have recently taken leave from universities to join frontier labs, and many AI2050 fellows hold industry positions alongside their academic posts. In the past six months another worry has surfaced: OpenAI’s models have solved a number of real research problems in mathematics, and some experts fear a diminished future role for humans in pure math. One fellow expressed concern about the mental health of mathematician peers.
It is not entirely bleak. Empirical science may be harder to automate than mathematics, since data collection is intrinsically slow. Some researchers view AI mathematicians and scientists as augmentative rather than replacement technologies. Tim Dettmers, a computer scientist at Carnegie Mellon University who works on making AI models faster and cheaper to run, argues that AI systems could make human scientists far more efficient, freeing them to pursue more ambitious and creative ideas.
Researchers are resilient: the very resource constraints that prevent them from training frontier models also drive efforts to build smaller, more efficient models or to explore new architectures. If the next major AI breakthrough comes from a scrappy academic lab rather than a large company, many would not be surprised.
Why it matters
The AI2050 convening highlighted a structural shift: the movement of cutting-edge AI development into closed, privately funded labs affects which questions get investigated, who has access to foundational tools, and how academic careers evolve. Those shifts will influence the trajectory of AI research and the kinds of innovations that emerge.
