MIT AI Alignment (MAIA)

MIT AI Alignment (MAIA) is an MIT student group whose central work is a structured cycle of AI-safety fellowship programmes, research mentorship, and semi-annual retreats connecting MIT students with the wider AI-safety practitioner community. Founded in fall 2022 as a deliberate expansion of a smaller weekly reading group shared between Harvard and MIT, MAIA grew within three months from roughly 15 students meeting informally to an organisation running 13 reading group sections and attracting over 230 fellowship applicants. It is fiscally sponsored by the Cambridge Boston Alignment Initiative (CBAI), a Cambridge-based 501(c)(3), and is rated "Very Active" by the global AISafety.com community directory. Its faculty advisor is Professor Max Tegmark. It is the corpus's first entry for an AI-safety insider community group at a US Northeast research university.

Founding and early growth

MAIA and its Harvard counterpart AISST trace their common origin to a single informal group of roughly 15 MIT and Harvard students meeting weekly to discuss AI-alignment papers. In fall 2022, a founding team including Xander Davies, Sam Marks, Trevor Levin, Naomi Bashkansky, and others expanded the informal gathering into two distinct organisations: MAIA at MIT and AISST at Harvard. The expansion was deliberately rapid — over approximately three months, MAIA launched 13 introductory reading group sections, two AI governance fellowships, a Science of Deep Learning reading group (started by Eric Michaud with around 8 active participants), and a Research Fellows programme pairing 20 undergraduates and graduate students with mentors. The December 2022 founders' post recorded 230+ fellowship applications, 130 enrolments, approximately 60 continuing, and two retreats with 85 unique attendees.

The founding post was candid about early friction: MAIA had higher attrition than AISST, attributed partly to a late semester launch and the absence of a dedicated office space equivalent to Harvard's. The founding team named two near-term priorities — securing MAIA office space (planned for February 2023) and building out advanced programming for technically experienced members beyond the introductory fellowship — both of which the current organisation has addressed.

AI Safety Fundamentals Fellowship

MAIA's flagship programme is the AI Safety Fundamentals Fellowship (AISF) — an 8-week reading group meeting roughly two hours weekly, covering AI's trajectory, evidence of misalignment, threat models, technical approaches, policy, and career paths. Cohorts run each fall and spring semester in the MAIA office with dinner provided; a virtual summer cohort of ~6 fellows per section runs in parallel. The curriculum adapts Richard Ngo's AGI Safety Fundamentals programme, extending the original 7-week format to 8 weeks with longer sessions and in-person reading to improve comprehension and discussion quality. MAIA prioritises MIT undergraduates and graduates but accepts all applicants; no work is assigned outside weekly meetings. All but one reading group facilitator has conducted AI-safety research, with most being PhD students or full-time researchers.

Research Fellows and membership

For students continuing past the introductory fellowship, MAIA runs a Research Fellows programme pairing students with mentors in the field — early cohorts placed 20 undergraduates and graduate students with researchers at safety-focused organisations. A membership track provides 24/7 MAIA office access, weekly technical meetings covering current research, the ARENA upskilling programme with teaching assistants, and direct access to field researchers; past guests include speakers from Redwood Research, OpenAI, MIRI, and Anthropic, and the group's early period featured Q&A sessions with Chris Olah and Richard Ngo. MAIA documents research contributions across adversarial attacks, RLHF, and scaling laws, with partner organisations listed as CHAI, MIT CSAIL, EleutherAI, Redwood Research, Anthropic, OpenAI, Epoch AI, and METR. Alumni have moved into technical, policy, and government roles.

Retreats

Every semester, MAIA and AISST co-host a weekend retreat gathering students, professors, and professionals to collaborate on AI-safety research and policy. The first two retreats in fall 2022 drew 85 unique attendees with guests from Redwood Research, OpenAI, Anthropic, Lightcone, Global Challenges Project, and Open Philanthropy. Programme content has included hands-on technical sessions — building transformers, replicating papers — alongside discussions with industry and research figures.

CAMBRIA bootcamp

CBAI, MAIA's parent organisation, runs CAMBRIA (Cambridge Bootcamp for Research in Interpretability and Alignment) — a 3-week in-person ML upskilling programme using the ARENA curriculum covering transformers, deep RL, and mechanistic interpretability. CAMBRIA is open beyond MIT students, runs primarily in Harvard Square, and provides housing, meals, 24/7 office access, and travel support. MAIA's website lists CAMBRIA as an affiliated programme. It is the successor to the MLAB-inspired ML bootcamp the MAIA founding team had planned for January 2023 — a lineage that places CAMBRIA as part of MAIA's original agenda to serve technically experienced members who had outgrown the introductory fellowship.

Place in the movement

MAIA is the corpus's first entry for an AI-safety insider community group at a US Northeast research university — the institutional heartland of American AI alignment research. Its position at MIT places it alongside the researchers, labs, and policy convenings where much of the technical safety literature originates: the group's partner list (CHAI, MIT CSAIL, EleutherAI, Redwood Research, Epoch AI, METR) maps closely to the organisations whose work it reads and whose staff speak at its retreats. In this function MAIA parallels the community-substrate groups the corpus covers in Berlin, Tokyo, and Singapore — each serves as the node through which residents of a major AI research ecosystem are recruited into and sustained within the AI-safety research tradition.

What distinguishes MAIA is the formality and depth of its research pipeline. Where Berlin AI Safety centres on monthly expert talks for an already-employed practitioner community, MAIA runs structured fellowship programmes, a mentored research track, and a full curriculum designed to convert technically capable MIT students into AI-safety researchers. The CBAI scaffolding — fiscal sponsorship, the CAMBRIA bootcamp, a shared research infrastructure — reinforces this pipeline function, providing semi-institutional continuity above a student-run organisation. The AISST partnership at Harvard extends the pipeline across the Charles River, giving the Boston AI-safety insider community a dual-institution footprint that has no direct parallel in the other community groups the corpus covers.