The posting
ABOUT RESOLUTION
Resolution does research on how to align artificial superintelligence (ASI). ASI may be developed in the next few years, but it is unclear whether alignment is on track to be ready in the same timeframe. We aim at higher a priori confidence in aligned outcomes by pursuing a portfolio of theory and empirics bets, any one of which, if it succeeds, would meaningfully advance the field. We invest heavily in research automation to accelerate progress, and we believe that stronger alignment theory unlocks higher automation: more principled approaches give us better filters for which directions of automated research are promising.
Resolution was founded in 2026 by researchers from UK AISI's Alignment Team, who ran the £30m Alignment Project https://alignmentproject.aisi.gov.uk/, and Timaeus https://timaeus.co/, who pioneered applying singular learning theory to alignment.
For more information, see our announcement https://resolution.org/launch.
ABOUT THE TEAM
Many teams at Resolution make use of Lean to accelerate their research, including Learning Theory (see our recent autoformalization of Hironaka’s Resolution of Singularities http://resolution.org/post/formalizing-resolution-of-singularities), Scalable Oversight (which is aiming for things like formal guarantees and impossibility results over debate protocols), and Agent Foundations (which is formalizing results on feedback, fixed points, and previsions in Lean).
Although these projects operate independently, they share a need for knowledge, infrastructure, tools, and wisdom for doing theory research with Lean. You would be the first person to drive this initiative. This means working closely with researchers and engineers in our research divisions, as well as with engineers in central engineering and automation roles.
ABOUT THE ROLE
You would be responsible for making Resolution good at Lean at scale.
At the highest level, this means providing taste and expertise in Lean and autoformalization. Resolution already houses several Lean users, but none of us are experts. So far we’ve gotten by on using AI agents to write Lean for us and on our general mathematics and software-engineering judgment. But this is not enough as our Lean efforts continue to grow. We are looking for someone who can guide us in our Lean usage and think strategically about what structures to build to make the various org autoformalization efforts go better. This will let us accelerate our theoretical research efforts, especially as we hand off more and more research to AI agents.
Importantly, this role does not mean you will be writing Lean. Agents will continue to write almost all of our Lean and will be the primary consumers of the Lean we produce. The volume of Lean will be larger than any human can individually interpret. Your job is to make this reliable and effective at that scale: set the conventions, architecture, and review structures that let researchers set the right definitions and trust agent output without reading every proof.
RESPONSIBILITIES
- Nurture Lean wisdom throughout the organization and be a champion of Lean and autoformalization.
- Build and maintain the structures so we can trust agent autoformalization outputs: proof harnesses, code review workflows, etc.
- Build and maintain shared Lean libraries with clear definitions, useful abstractions, and a structure that supports review and further (automated) research.
- Consolidate overlapping formalizations, resolve inconsistent definitions, and build shared foundations for researchers and agents to reuse.
- Improve the reliability, maintainability, and performance of shared Lean projects.
- Support researchers in developing custom tools, tactics, and workflows to help them make progress.
Implicit in all of these responsibilities is the understanding that this will be heavily AI-assisted.
YOU MAY BE A GOOD FIT IF YOU
- Have several years of experience writing Lean and mathlib.
- Have expertise in Lean community practices (so you can, e.g., explain why one definition, abstraction, or library design works better than another).
- Are an AI power user and have experience with autoformalization, and plan to (eventually) never write another line of Lean by hand.
- Communicate clearly and enjoy working closely with researchers and engineers.
- Have a strong mathematics or theoretical computer science background (Bachelor’s at minimum).
- Want to contribute to the alignment of artificial superintelligence.
STRONG CANDIDATES MAY ALSO HAVE
- Contributions to mathlib or other substantial Lean libraries.
- Experience designing or maintaining large formalization projects.
- Experience with Lean tooling, tactic development, metaprogramming, or proof automation.
- Experience with AI-assisted mathematical research or automated evaluation.
APPLICATION PROCESS
After application review, our process includes:
- Screening Call (30 minutes): discuss your background, motivation, and Lean experience.
- Experience Interview (1 hour): examine past projects, your contributions, as well as mathematical and library design choices.
- [Optional] Live Technical/Work test: may end up being skipped; details to be determined.
- Individual conversations with Geoffrey and other team members (30 minutes each): discuss the work, your experience, and how we could work together.
We will share the format before each interview. Each conversation includes time for your questions.
LOGISTICS
This is a full-time role based in Berkeley, California. Remote work may be considered in exceptional cases. We can sponsor visas for relocation to Berkeley.
Current Research Engineer salary levels are:
- L4: $208,000 remote, $346,000 in person.
- L5: $270,000 remote, $451,000 in person.
- L6: $402,000 remote, $670,000 in person.
- L7: $558,000 remote, $930,000 in person.
Benefits include five weeks of paid vacation plus public holidays, comprehensive medical, dental, and vision insurance, unlimited sick leave, and an unconditional 401(k) contribution equal to 4% of salary.
Minimum education: a relevant bachelor's degree.
Applications are reviewed on a rolling basis. Start date: as soon as practical.
