Why It Is a Hard Time to Be an AI Academic

University AI researchers are facing a systemic shift in the field as the cutting edge of development moves from academic institutions to private companies, according to reporting from MIT Technology Review. This transition is driven by the prohibitive cost of GPUs required to train frontier models and a lack of transparency from private labs regarding the inner workings of tools like ChatGPT and Claude.

The disparity in resources has created a divide where academics can observe the behavior of large language models (LLMs) but cannot access the design or training data. Nika Haghtalab, a computer science professor at UC Berkeley, compared the current state of AI academia to biologists in a world where private companies hold exclusive control over the gene-editing tool CRISPR, according to MIT Technology Review.

Funding constraints extend beyond hardware. Researchers noted that the cost of repeatedly querying models from OpenAI, Anthropic, and Google for rigorous study can be prohibitive, a situation exacerbated by a reduction in U.S. federal scientific funding.

To maintain relevance, some academics are pivoting away from capabilities that private firms are likely to prioritize. Anjalie Field, a computer science professor at Johns Hopkins, stated she avoids problems she believes will be solved by tech companies, as corporate research is often driven by profit motives.

Field recently conducted a study finding that language models provide less sophisticated responses to prompts phrased in ways more commonly used by women than by men, a type of research she suggests is unlikely to originate from companies like Anthropic or OpenAI, according to MIT Technology Review.

The Impact of LLM Dominance on Specialized AI

A significant portion of AI academics focus on specialized models designed for data analysis, predictions, or simulating physical systems rather than general-purpose LLMs. These researchers face a different set of challenges, primarily the public perception that AI is synonymous with energy-intensive LLMs.

Researchers building specialized tools for climate change, for example, reported difficulty advocating for their work due to this widespread misunderstanding of the AI field, according to MIT Technology Review.

The instability of these roles is evident in the industry’s movement of talent. Several prominent academics have taken leave from universities to join frontier labs, while others maintain dual positions in both industry and academia.

Automation Risks in Mathematics and Science

The ability of OpenAI’s models to solve complex research problems in mathematics has introduced new concerns regarding the future of the profession. One researcher expressed concern over the mental health of peers in pure mathematics who may feel threatened by automation, according to MIT Technology Review.

However, some experts argue that empirical science is more resistant to automation than mathematics because data collection is an inherently slow process. Tim Dettmers, a computer scientist at Carnegie Mellon who focuses on making AI models cheaper and faster to run, views AI scientists as a tool for efficiency.

Dettmers stated that AI scientists will not replace humans but will instead allow researchers to pursue inspired ideas that they otherwise would not have the time to explore.

The current resource gap is also driving innovation in efficiency. Because they cannot afford the massive compute power of frontier labs, academic researchers are pushed to develop smaller, more efficient models or entirely new architectures.

These developments occurred within the context of the Schmidt Sciences AI2050 program, an initiative funded by Eric and Wendy Schmidt that provides funding and GPU access to academics working with AI, according to MIT Technology Review.