AI Researchers Confront Challenges from Tech Giants

A Meeting at the Heart of AI Innovation

Recently, a gathering of artificial intelligence researchers took place in Mountain View, California, bringing together some of the brightest minds in the field. This meeting, organized by the Schmidt Sciences AI2050 program, aimed to support academics working on AI-related projects. Funded by Eric and Wendy Schmidt, the program serves as a meeting point for researchers. The list of participants is impressive, featuring influential figures in AI, although not everyone could attend. (Full disclosure: the author received a science communication award funded by Schmidt Sciences in 2024.)

Academic AI Research Facing Major Obstacles

Academic AI researchers are navigating a complex period. In recent years, the focus of research has shifted towards large language models, with a notable migration of talent from universities to private companies. Academic institutions struggle to compete with the resources of companies like Anthropic and OpenAI, which have the necessary infrastructure to develop and deploy advanced models such as Claude and ChatGPT. Universities lack the essential GPUs for these tasks, and even if they had them, access to the internal details of these models remains limited.

A Striking Comparison to the Field of Biology

Nika Haghtalab, a professor at UC Berkeley, compared the current situation of AI researchers to that of biologists in a world where genetic editing tools like CRISPR are exclusively controlled by private companies. Researchers outside cutting-edge labs can observe the behavior of models like ChatGPT, but they are excluded from the design and training processes of these tools, limiting their ability to influence the development of these technologies.

Financial Support from the AI2050 Program

The AI2050 program provides financial support to researchers, allowing them to acquire GPUs, a significant advantage according to several participants. However, funding remains a constant concern, especially as federal support for scientific research in the United States declines. For those not managing models locally, the cost of repeated queries to models from companies like OpenAI can be prohibitive.

A Shift Towards Neglected Issues by Companies

In light of these challenges, many researchers choose to focus on questions that technology companies do not address. Anjalie Field, a professor at Johns Hopkins, explains that she avoids working on problems likely to be solved by companies, preferring to explore topics that may not be profitable for them. For instance, she conducted a study revealing that language models respond differently based on the gender of the query formulation, research unlikely to be pursued by companies like Anthropic or OpenAI.

Specialized AI Researchers and Their Challenges

Not all AI researchers focus on large language models. Some develop specialized models to analyze data or simulate physical systems. Although they are not in direct competition with tech giants, these researchers face difficulties. During the meeting, concerns were raised about the lack of awareness of non-LLM AIs, complicating the valuation of their work, particularly in areas like climate change. Additionally, the AlphaFold team from Google DeepMind, which developed a model predicting protein structures and was awarded a Nobel Prize, was recently dissolved.

Transformations in the Academic Landscape

Current challenges are profoundly altering the academic landscape. Several researchers have taken leaves of absence to join cutting-edge labs, and many juggle academic and industrial positions. Furthermore, the advancement of OpenAI's models in solving mathematical problems raises concerns about the future of pure mathematics and the mental health of mathematicians.

An Optimistic Perspective Despite Challenges

Despite these obstacles, some researchers remain optimistic. Empirical science, for example, may withstand automation due to the complexity of data collection. Tim Dettmers from Carnegie Mellon views AI models as tools to enhance the efficiency of human scientists, enabling them to explore innovative ideas. Current constraints are pushing researchers to innovate, making models more efficient or exploring new architectures. Thus, the next major breakthrough in AI could very well emerge from a bold academic lab.