As AI has integrated itself into society, we have observed more cases of them performing surprisingly sophisticated tasks, such as (dis)proving long-standing mathematical conjectures\u2014but also failing in big ways, such as always agreeing with users, even when the user intends to do something harmful. Can we understand how and why AI systems do things like this? Can we then use this understanding to improve these systems? Mechanistic interpretability is a subfield of AI research that aims to answer these very questions.<\/p>\n
On August 14, 2026, Boston University hosted the 3rd Annual New England Mechanistic Interpretability (NEMI) Workshop. Led by students and faculty in the Department of Computer Science and Faculty of Computing and Data Sciences, the organizing team put together a large event in the George Sherman Union. This year, NEMI received over 100 submissions, and had nearly 300 attendees\u2014a significant growth over previous years. After the main workshop, a social event was hosted at the CDS building.<\/p>\n
While many attendees were from the New England area (including attendees from BU, Northeastern, Harvard, MIT, and Brown), a significant number arrived from New York and Pennsylvania, and even far-away states like California. Some also arrived from abroad, including researchers from the UK and the Netherlands. Presenters and attendees at the workshop included scientists from academia, industry, and at various career stages.<\/p>\n
The presenters at the workshop spoke about various topics, including the geometry of language model representations, how to make AI safer, how to figure out where AI systems are \u201clooking\u201d in complex images, and how to automatically watch, or \u201cmonitor\u201d, the thinking of language models.<\/p>\n
A noticeable theme throughout the day was whether interpretability can be thought of as a science, and how it might influence other scientific disciplines. A panel discussion featuring professors from the New England area focused on what interpretability can do for the natural sciences, and how those sciences can inspire AI interpretability as well.<\/p>\n
The event was sponsored by Goodfire<\/a>, an AI interpretability startup based in San Francisco. It was also sponsored by the Enigma Project<\/a>, a computational neuroscience and AI initiative based at Stanford University.<\/p>\n","protected":false},"excerpt":{"rendered":" As AI has integrated itself into society, we have observed more cases of them performing surprisingly sophisticated tasks, such as (dis)proving long-standing mathematical conjectures\u2014but also failing in big ways, such as always agreeing with users, even when the user intends to do something harmful. Can we understand how and why AI systems do things like […]<\/p>\n","protected":false},"author":23854,"featured_media":23752,"comment_status":"closed","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[18],"tags":[],"_links":{"self":[{"href":"https:\/\/www.bu.edu\/cs\/wp-json\/wp\/v2\/posts\/23749"}],"collection":[{"href":"https:\/\/www.bu.edu\/cs\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.bu.edu\/cs\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.bu.edu\/cs\/wp-json\/wp\/v2\/users\/23854"}],"replies":[{"embeddable":true,"href":"https:\/\/www.bu.edu\/cs\/wp-json\/wp\/v2\/comments?post=23749"}],"version-history":[{"count":3,"href":"https:\/\/www.bu.edu\/cs\/wp-json\/wp\/v2\/posts\/23749\/revisions"}],"predecessor-version":[{"id":23754,"href":"https:\/\/www.bu.edu\/cs\/wp-json\/wp\/v2\/posts\/23749\/revisions\/23754"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.bu.edu\/cs\/wp-json\/wp\/v2\/media\/23752"}],"wp:attachment":[{"href":"https:\/\/www.bu.edu\/cs\/wp-json\/wp\/v2\/media?parent=23749"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.bu.edu\/cs\/wp-json\/wp\/v2\/categories?post=23749"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.bu.edu\/cs\/wp-json\/wp\/v2\/tags?post=23749"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}