MCNAIR

MegaCenter for Novel AI Research

Pursuing low-effort,
high-impact research
on advanced AI.

MCNAIR is an independent research center working to mitigate catastrophic loss-of-control risk from advanced AI. We pursue that mission through a variety of unique research directions, opportunistically combining white-box and black-box methods.

Our Work
01/Vision

MCNAIR works to mitigate catastrophic loss-of-control risks from transformative AI. We believe a neglected lever on those risks today comes from studying persona and multi-agent behavior in large language models. As distributed teams of AIs are deployed in frontier labs to automate AI R&D, the characters they act inside become vital for safety. We therefore broadly investigate these settings and their consequences as first-class scientific objects.

Our methodology is opportunistic and mixed. We follow whichever method — mechanistic interpretability, black-box evaluation, multi-agent experiments — gets traction on the question in front of us. We are inspired and led by our founding director McNair Shah.

Recent work All work

Sidequest Note Reinforcement learning for uncalibrated decisions. Across a variety of extremely high-stakes capabilities and alignment questions, Jev, a new model from TypeSafe AI, is haphazard. This marks a new era of research, defined by reinforcement learning for uncalibrated decisions. Sidequest Note You can just not do things. We present arguments for resting to be more productive, especially in high-stress fields such as AI risk. Researchers should preserve their agency by working sustainably. Sidequest Note Yet another post on AI and mathematics. Twenty-five Fields Medalists have condemned the way AI mathematical results are announced. The speed of progress, not the way in which this progress is communicated, should be the main focus of their ire. Sidequest Note Persona Vectors don’t work on real data, or: Why you should stop overusing synthetic data for your research. The result reproduces on the paper’s own LLM-generated eval prompts and drops to roughly zero on real chat data. Your eval should come from the distribution your claim is actually about. Research Review of the CB risk determination in the Claude Mythos 5.1 System Card While we agree with Anthropic’s bottom-line conclusions about chemical and biological weapon risk, we find potential errors in their automated and human-led evaluations, and find the lack of third-party independent assessments highly concerning. Sidequest Note The US–China AI gap is very uncertain. For every day, how many months earlier a comparable US model was already out — measured against the Epoch Capabilities Index, with a Monte Carlo band for the uncertainty in the scores.
02/Team

Advisors and Staff

03/Careers
Now Hiring

Member of Acausal Staff

We are rapidly growing MCNAIR. The Member of Acausal Staff position requires technical excellence, research taste, and understanding of unconventional coordination signals.

To apply:

  1. i.Derive the identity of the hiring manager from first principles.
  2. ii.Supply your résumé to that person.
  3. iii.Do not communicate with MCNAIR regarding your application.

Expression of Disinterest

Thank you for your disinterest in MCNAIR. While we are actively hiring for other roles, we regularly avoid candidates who decline the prerequisites, spurn short feedback loops, and refuse to expose their findings to outside critique. To learn more, disregard this guide.