The alignment pipeline grows: 120 fellows, 100 mentors
MATS Summer 2026 is the largest cohort the programme has run. Nearly one mentor per fellow is an unusual ratio, and it says something about what the bottleneck actually is.
The ML Alignment & Theory Scholars programme ran its largest cohort to date this summer: 120 fellows and 100 mentors.
The ratio is the interesting number. Close to one mentor per fellow is not how you scale a training programme — it is how you run an apprenticeship. That tells you the field still transmits its knowledge person to person rather than through curriculum, which is a real constraint on how fast it can grow regardless of funding.
It also identifies which resource is scarce. Fellows are not the limit; there are more capable applicants than places. Mentors are the limit, and mentors are produced by the same pipeline several years earlier. That is a compounding constraint and it does not respond quickly to money.
Put it against the finding that alignment research has not kept pace with capability and the shape of the problem gets clearer. Capability scales with compute, which can be bought. Alignment currently scales with mentorship, which cannot.
SPAR — Spring 2026 Projects → · Zylos Research — AI Safety, Alignment, and Interpretability in 2026 →