2027 AI/ML Research Fellow

10 places are open for the 2027 term, with a duration of 6 or 12 months at the applicant's option.

Purpose

There is a crisis in academic AI/ML research. What is still worth working on?

Paper submissions have doubled on the arXiv and at ML conferences, and yet many submitters cannot even answer questions about the details of "their" work.

Many people across applied research and theory groups, whether math, physics, or computer science, have the distinct feeling that perhaps it is pointless to work on anything other than AI safety and continual learning.

If we simply draw a line continuing the capabilities progress of the past two years forward, it seems possible that human contributions will be a marginal input in many areas of research within two or three years.

At the same time, it is clear that fundamental questions in model architecture, sample efficiency, explainability, and creativity are still unanswered, and that making progress on these topics is very likely to benefit humanity over the next two years.

Designing or training for explainability and determinism may be much easier than achieving post hoc understanding of circuits forced out of scaled data by gradient descent and optimized through blind RL.

Our focus instead is on frontier research: order-of-magnitude improvements, rather than disposable papers barely better than baseline.

The intent of these fellowships is to give a