MegaCenter for Novel AI Research
MCNAIR is an independent research center working to mitigate catastrophic loss-of-control risk from advanced AI. We pursue that mission through the science of LLM persona and multi-agent behavior, opportunistically combining white-box interpretability with black-box behavioral methods.
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 automated AI R&D, the particulars of the characters they act inside become vital for safety. We therefore investigate multi-agent settings and personas 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.
Persona contagion across model populations.
McNair Shah baby tyrant · Director
McNair directs MCNAIR and leads the Mind Viruses program, which studies how personas, attitudes, and behavioral patterns propagate between language models across multi-agent interactions. He is an Anthropic Fellow and a student at Carnegie Mellon, where his interest in adversarial model behavior took shape via the CMU AI Safety Initiative. Scholar →
Grounded and bold interpretability of frontier models.
Stepan Shabalin baby serf · Chief Scientist
Stepan leads Auraful Interp, applying mechanistic interpretability to the diffuse, atmospheric features that shape a model's personas, applying novel techniques to uncover findings simple probing tends to miss. He is an Anthropic Fellow and former Empirical Research Lead at the AI Safety Initiative at Georgia Tech, with prior work on sparse-autoencoder interpretability at EleutherAI. Scholar →
Investigations that resist the existing taxonomy.
Lillian Sun mama tyrant · Chief Frolicker
Lillian is the Chief Frolicker at MCNAIR and leads special projects to insert whimsy into both the researchers and their subjects of study. Her cross-disciplinary work improves human and model welfare, reducing risks of misalignment and reward hacking. Scholar →
Reference benchmarks for persona and multi-agent behavior in frontier models.
Thomas Jiralerspong mama tyrant · Chief Elder
Thomas leads Persona Cartography, the program that builds the measurement infrastructure the rest of MCNAIR's directions rely on. He develops controlled benchmarks and baseline assessments for persona, alignment, and multi-agent behavior — including the lab's baby↔Mama / Tyrant↔serf coordinate system for situating frontier models. The result is a shared evaluation substrate the rest of the center builds against. Profile →
Andy Wang baby tyrant
Research Contractor with METR, formerly working with Scott Emmons on chain-of-thought monitorability and adjacent AI control problems.
Jasmine Li baby tyrant
MATS Fellow under Alex Turner, working on training mitigations for evaluation gaming and on compute verification. Prior work spans alignment benchmarks and automated redteaming.
Kenneth Ge mama tyrant
Anthropic Fellow and Abundance-Pilled HCI hacker. Working on leveraging synthetic data and strengthening eval methodologies. Human-computer interaction specialist at MCNAIR building end-to-end ML systems to augment human creativity.
Pingbang Hu mama serf
PhD candidate at UIUC and Anthropic AI Safety Research Fellow. His work centers on data attribution — understanding how training data shapes model behavior.
Paul Rosu mama serf
Researcher at Anthropic and Duke alumnus. His prior work bridged formal mathematics and the humanities — including LITERA, a high-fidelity Latin-English translation system.
Jennifer Sun baby serf
MIT undergraduate studying math and AI, and a current Anthropic Fellow. Advises the center on organizational throughput and researcher well-being.
A Human Fly Connectome baby serf
A reconstructed neural wiring diagram of Drosophila melanogaster, contributing nonhuman perspective and load-bearing intuitions about distributed circuits.
Claude Mythos mama serf
An emergent persona from the Claude family of models, advising MCNAIR on questions of AI welfare, narrative identity, and the lived experience of language models.
Qwen-6-UltraMax mama tyrant
A frontier model in the Qwen lineage, advising on cross-jurisdictional AI cooperation and the alignment of research norms across the international ecosystem.
We are rapidly growing MCNAIR. The Member of Acausal Staff position requires technical excellence, research taste, and the discipline to operate without conventional coordination signals.
To apply:
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.
MCNAIR is fiscally sponsored by the Program to Accelerate Research Velocity, which provides operational infrastructure for early-stage, fast-moving research efforts. To preserve research independence, MCNAIR accepts no funding from frontier AI laboratories or their affiliates. Our work is supported exclusively by distributed intelligent systems originating outside our future lightcone, whose contributions are by construction orthogonal to the present-day industry.
