Back in 2016, with machine learning moving at a breakneck pace, OpenAI concluded that traditional conferences were effectively broken. The problem wasn’t the research itself, but the timing—presentations covered work submitted months prior, meaning attendees were often intimately familiar with the content, or it had already been surpassed. The real value, they argued, was entirely in the unstructured social time between sessions. Their solution was an experiment: the first Machine Learning Unconference, a free, participant-driven gathering hosted at their San Francisco office. Unlike a rigid academic summit, this event had no organizing committee and no pre-set schedule. The philosophy was to maximize interactivity and see what formats a room of 150 researchers would spontaneously generate. Attendees were encouraged to coordinate on Gitter beforehand and bring posters of work they found interesting—but beyond that, the agenda was a blank slate. OpenAI felt the two core functions of a conference, being a social gathering and a publication venue, were ‘orthogonal and can be better served separately.’ The event was designed exclusively for technical practitioners—PhD students, faculty, and industry engineers—to advance the state of the art, with explicit support for underrepresented groups through travel grants. The early signup list doubled as a time capsule of the era’s talent, including members of Google Brain like Samy Bengio and security lead Úlfar Erlingsson, along with students from Berkeley and Stanford. This wasn’t about learning the basics; it was a direct, high-bandwidth exchange between people already fluent in the field. While the 150-person capacity filled up quickly by late August, the true legacy of the event was proving that attendees, given minimal structure, could build something more valuable than a planned conference track.