AI Professors Adapt as Tech Giants Dominate Frontier Models

At a recent gathering of the Schmidt Sciences AI2050 program in Mountain View, California, academic researchers highlighted a stark shift in the artificial intelligence landscape. Because universities cannot afford the massive graphics processing units (GPUs) required to train and run frontier models, the cutting edge of AI development has migrated from academic institutions to private corporations. Companies like OpenAI and Anthropic keep the inner workings of models like ChatGPT and Claude strictly proprietary, leaving university professors locked out of foundational research.

UC Berkeley computer science professor Nika Haghtalab compared the current academic environment to a scenario where private entities hold "exclusive control over the gene-editing tool CRISPR." Without the resources to train their own large language models, and facing a reduction in federal scientific funding in the United States, researchers struggle even to pay for the API queries needed to study commercial systems. Programs like the one funded by Eric and Wendy Schmidt offer some financial relief for GPU access, but the resource gap remains vast.

To adapt, many academics are focusing on areas that commercial labs ignore. Johns Hopkins professor Anjalie Field explained that she tries "not to work on problems" that tech companies will solve. Field recently published a study showing that language models provide less sophisticated responses to prompts phrased in ways more commonly used by women than by men—a research angle unlikely to be prioritized by profit-driven firms. Other researchers are focusing on specialized, non-LLM scientific models, though they face public relations hurdles now that Google DeepMind recently disbanded its Nobel Prize-winning AlphaFold protein-prediction team.

This resource squeeze is driving some prominent academics to take leaves of absence for industry roles. However, others see opportunity in constraint. Carnegie Mellon computer scientist Tim Dettmers is working to make AI models cheaper and faster to run. He believes automated AI tools will ultimately make human scientists more efficient rather than replacing them. This pressure to innovate under tight constraints may ultimately force academic labs to pioneer highly efficient, alternative architectures that bypass the need for massive corporate supercomputers.