Last week, I headed to a hotel in Mountain View, California, to join top AI researchers at a Schmidt Sciences AI2050 convening. It’s a weird time for university AI researchers: the cutting edge has moved from academic institutions to private companies. Universities can’t afford the GPUs required to train frontier models, and even if they could, Anthropic and OpenAI aren’t letting anyone see the inner details of Claude or ChatGPT. Over lunch, Nika Haghtalab, a computer science professor at UC Berkeley, said being an AI academic is like being a biologist in a world where companies have exclusive control over CRISPR.

The AI2050 program, funded by Eric and Wendy Schmidt, offers fellows some funding for GPUs, but money remains a pressing concern given reduced federal funding. The cost of repeatedly querying models from OpenAI, Anthropic, and Google can be prohibitive. Rather than advancing capabilities, many fellows focus on questions unlikely to be addressed by companies. Anjalie Field, a computer science professor at Johns Hopkins, found that language models give less sophisticated responses to prompts phrased in ways more common among women—research unlikely to come from for-profit labs.

There’s also a huge group of AI academics who don’t work with LLMs. They build specialized models for data analysis, predictions, or simulating physical systems. These researchers face challenges from widespread ignorance of non-LLM AI. Several prominent academics have joined frontier labs, and some worry about AI solving math problems. But it’s not all doom and gloom: Tim Dettmers, a computer scientist at Carnegie Mellon, says AI scientists could make humans more efficient. The very resource constraints pushing academics away from training frontier models also push them to discover new ways to make models smaller and explore new architectures—so if the next big AI breakthrough comes from a scrappy academic lab, I won’t be shocked.