Job Description
Brahma is assisting our client in this search.
\n
\n
About the company
\n
Frontier AI models are limited less by algorithms now than by the data available to train them. Our client builds the reinforcement learning environments and datasets that leading AI labs use to train their next generation of models.
\n
\n
The team is small, technical, and well backed, with a founding group out of top-tier ML and infrastructure organizations.
\n
\n
About the fellowship
\n
This is a fixed-term research seat, typically one or two semesters, with the option to extend to a full year.
\n
\n
It exists for one reason: a lot of the people best suited to this work are partway through a PhD and aren't ready to give it up. Walking away from four years of invested effort is a real decision, and it shouldn't be the price of admission for spending time at the frontier.
\n
So the fellowship is the same work, same problems, same ownership as the full-time research staff, on a defined term. Come do it for a semester or two, then go back and finish your degree. If you get here and decide you'd rather stay, converting to full time is straightforward. Either outcome is fine, and there's no pressure toward one.
\n
\n
The company is flexible on timing. A better question than how many months is how long you'd need to actually apply your research.
\n
\n
The role
\n
The core problem is designing tasks that a person can solve and can recognize as correct, but that a frontier model gets wrong. Then working out how to verify correctness programmatically, with no human in the loop. That verification piece is the hard part and it's where most of the research lives.
\n
\n
The company's premise is that this window is closing. Before long, models will be capable enough that a human can't reliably judge the output, so the systems that generate and verify this work need to run without us. That's the research agenda.
\n
\n
You'll own problems end to end. Track what's been published recently, decide what's worth pursuing, form the research question, build the software to test it, and understand the theory well enough to know what you're actually measuring. There's no one to hand the implementation to.
\n
\n
We're looking for evidence you can run research independently, start to finish. That usually shows up as one of the following, though none of them is a hard requirement:
\n
- \n
- Accepted, main-track papers at NeurIPS or ICML, ideally as first or second author
- Meaningful contribution to a well-regarded public benchmark, as author or task contributor
- A record of building and shipping research systems yourself
\n
\n
\n
\n
\n
Track record matters more than pedigree. Several of the strongest people on this team came from programs that don't get recruited and got there through serious open-source work. Degree level is not a hard requirement.
\n
\n
Current focus areas are long-horizon agentic tasks and security, though the underlying skill transfers across domains.