The Atlas research agenda
Scalable oversight, automated alignment research, and the evals that tell you whether either one works.
Epistemic status Working agenda for an early-stage org. Will change as experiments come in. A fuller write-up with concrete experiments is in progress.
Atlas Research is an early-stage technical AI safety org. Its question is how we can safely use capable AI systems to supervise and align more capable ones.
Why oversight
Most alignment techniques need an overseer who can tell good work from bad. RLHF, evals and constitutional methods all assume one. That gets harder as models improve, and it fails exactly when the stakes are highest: when a system is more capable than the people checking it, or can produce reasoning that is persuasive but wrong.
If oversight scales, much of the rest of the safety portfolio gets easier, because we can safely use AI to speed it up. If it doesn’t, we can’t trust what frontier systems produce, including their alignment research.
There is also a timing argument. Debate, weak-to-strong generalization and process supervision can be tested on today’s models, where we still have ground truth. Once systems pass human ability in a domain, we lose the ability to check whether our oversight methods work there.
Three problems
The first is scalable oversight. How can people reliably evaluate the outputs, plans and reasoning of systems that are better than them at the task? I’m interested in debate, recursive reward modeling, weak-to-strong generalization, process supervision, task decomposition and AI-written critiques.
The second is automated alignment research. How can AI speed up alignment research while staying under oversight? That includes AI generating hypotheses, designing experiments, red-teaming models and finding failure modes, with every step checked. Crux is my first concept here.
The third is evaluation. How do we tell whether an oversight method is working? I care most about failures that are hard to see, like sycophancy, sandbagging, reward hacking, and models that look helpful while quietly working against the evaluator.
How I want to do the work
Experiments should run on current or near-frontier models and connect to real catastrophic-risk threat models. Where possible they should explain why a method works, and not only show that it seems to. They should be designed for the case where the model is better than its supervisor, and written up so other researchers can check and build on them.
Next 6–12 months
Refine a concrete roadmap, build early evaluation tooling for AI-assisted alignment work, run first experiments on oversight failures in research tasks, and publish what comes out of them. Atlas is looking for a research co-founder to lead the technical work, and for advisors and collaborators.
If that’s you, email me.