OpenAI Safety Committee: Members, Mandate and Model Training

Four directors, five department heads, two outside advisors

OpenAI announced that its board had formed a Safety and Security Committee chaired by director Bret Taylor, with fellow directors Adam D'Angelo, Nicole Seligman and CEO Sam Altman as the other members. The committee's job is to make recommendations to the full board on "critical safety and security decisions for OpenAI projects and operations." [1]

Five of OpenAI's own department heads were named to sit on the committee as well: Aleksander Madry (Head of Preparedness), Lilian Weng (Head of Safety Systems), John Schulman (Head of Alignment Science), Matt Knight (Head of Security) and Jakub Pachocki, who had become chief scientist two weeks earlier when Ilya Sutskever left. OpenAI said it would also retain outside advisors, naming former NSA cybersecurity official Rob Joyce and former assistant attorney general John Carlin. [1]

Atlas interpretation: Every voting member of the committee already sat on OpenAI's board or drew a paycheck from the company. Business Insider's contemporaneous reporting made the same point: the roster held no outside AI safety researcher and no academic with no other tie to OpenAI, so the committee's independence rests on the two non-employee directors, D'Angelo and Seligman, rather than on any structural separation from management. [2]

The oversight announcement and the training announcement were one post

The blog post that announced the committee also disclosed that OpenAI had "recently begun training its next frontier model," the model that would succeed GPT-4, and said the company anticipated the resulting systems would bring it "to the next level of capabilities on our path to AGI." [1]

The committee's first assignment, the post said, is to spend 90 days evaluating and further developing OpenAI's existing processes and safeguards, then deliver recommendations to the full board. OpenAI committed to publicly sharing an update on whichever recommendations the board adopts, worded as consistent with safety and security rather than as a full disclosure. [1]

Announced eleven days after the team built for this stopped existing

CNBC's reporting the same day framed the new committee against what it replaced: OpenAI had dissolved its superalignment team, the group built specifically to research long-term AI risk, after both of its leaders left within days of each other, Sutskever with a short public post and Leike with a longer thread accusing OpenAI of letting "safety culture and processes" take "a backseat to shiny products." [2]

Atlas interpretation: The summary above calls the departures' subtext legible in the new committee, and the sequence supports that reading without requiring OpenAI to have said so: a standing team built for exactly this kind of evaluation stopped existing, its two most senior members resigned citing safety concerns, and eleven days later the board announced a replacement structure staffed mostly by people already inside the company, in the same post that revealed training of a more capable successor model had already started. [2]

A retired general joined three weeks later

OpenAI's post carries an editor's note, dated June 18, 2024, stating that retired U.S. Army General Paul M. Nakasone joined the Safety and Security Committee upon his appointment to OpenAI's board that same day. The note is appended to the original announcement rather than issued as a separate one. [1]