OpenAI just wrapped its third Scholars program with a virtual Demo Day, and the projects coming out of this cohort don’t feel like student work. These aren’t toy problems. Over five months, a group of relative machine learning newcomers—all from underrepresented groups—tackled everything from reverse-engineering how GPT-2 represents grammar to predicting epileptic seizures from brain recordings. One scholar even focused on measuring interpretability in models trained on Coinrun, a classic reinforcement learning environment. The range is genuinely impressive.
The program itself is built on a simple premise: stipends and mentorship for people who have the technical chops but lack the traditional pathways into deep learning. OpenAI is betting that self-motivation and domain expertise matter more than credentials. Looking at the output, it’s hard to argue with that bet. The organization stressed that each participant entered machine learning as a relative newcomer, a point meant to show the field’s growing accessibility. Microsoft provided the Azure compute credits to make the heavy lifting possible.
But the subtext here is bigger than a graduation announcement. OpenAI is making a direct link between diversity and AI safety—arguing that building systems that benefit everyone requires the people building them to actually reflect everyone. It’s a pragmatic argument, not just a moral one. If your training data and your engineering teams both skew homogeneous, the resulting models will encode that narrowness in ways that are difficult to audit after the fact.
Applications for the next class will be announced this fall. For anyone sitting on the fence, the message from this Demo Day is clear: the barrier to entry is lower than you think, and the problems you can work on are more serious than you’d expect. You don’t need a PhD. You need five months and a question that actually matters.
