Join us for a talk by Stella Rose Biderman, Executive Director of EleutherAI. This talk is part of the Kempner Seminar Series, a research-level seminar series that covers topics related to the basis of intelligence in natural and artificial systems.
AI research increasingly feels like a game only industry labs can play. The most visible work trains ever-larger models from labs with compute budgets inaccessible to academic researchers. But the problems that dominate the field’s attention are not the only problems that matter, and in many cases they are not even the most important ones. Foundational questions about how models acquire capabilities during training, whether our evaluations measure anything meaningful, and how deployed systems affect real people remain wide open, with remarkably few researchers working on them. These are not niche concerns. They reflect a field that has prioritized engineering over scientific understanding and chasing hype over doing careful work. In this talk, I will discuss how we think about doing research that matters at EleutherAI and how we and our collaborators often find that the most important open problems in AI are hiding in plain sight, two steps off the beaten path.