Michigan State University · Est. 2026
Frontier AI safety research for college students
AI and alignment research at Michigan State University, done by undergraduates, published at real venues. By students, for students.
Measured externally
Work that survives peer review, not work that impresses a student audience.
Target venue, target date
A project without a target venue is a reading topic, not a research project.
Reading is training
One presenter, one assigned skeptic, every week.
Rigor, not subfield
The Institute does not gatekeep by subfield; it gatekeeps by rigor.
Outcomes record
Publications, placements, awards
2026-09Subspace dictionaries across depth: comparing SASA and sparse autoencoders in Gemma 3 1BNeurIPS 2026 · WorkshopPending2026-04From commands to invitations: the pragmatic softening of child-directed speech, 1960s–2000sUURAF 2026 · PosterPresented2026Horus: an instrument for navigating and steering the feature space of language modelsProject · Open sourceBuilding2024-07Impact of network topology on Byzantine resilience in decentralized federated learningarXiv:2407.05141Published
Research areas
Rigor, not subfield
Area
Language models and language acquisition
How compact models acquire structure.
Area
Machine learning for scientific domains
Applied ML with MSU labs and HPCC compute.
Area
Alignment and interpretability
Feature discovery, probing, evaluations.
Projects
Built under the Institute banner
Research that matures into tools ships. The Institute builds AI products under its own name when the work earns it.
Membership
Three reading events. One nomination.
Attend three reading events. Get nominated by a Member. Officers confirm within fourteen days. No dues.
Affiliate
Attends Institute events. No obligations, no vote.
Member
Weekly research hours, one talk per semester, one external artifact per year. Votes.
Alumnus
Listed, reachable, advises.