Loading...
On Friday, November 17, 2023, OpenAI's six-member nonprofit board fired CEO Sam Altman via a Google Meet call with less than thirty minutes notice. The board — comprising Adam D'Angelo, Tasha McCauley, Helen Toner, and Ilya Sutskever — cited a loss of confidence in Altman's candor. They had the legal authority to do it. OpenAI's unusual structure placed a nonprofit board, with no equity stakes, in control of a for-profit subsidiary valued at $86 billion. What followed was a five-day crisis that exposed every fault line in that design. President Greg Brockman resigned within hours. By Sunday, 95% of OpenAI's 770 employees had signed a letter threatening to leave unless the board resigned. Microsoft CEO Satya Nadella, whose company had invested $13 billion, publicly offered to hire Altman...
Popular framing: A well-intentioned safety board tried to remove a CEO who had become too powerful, and lost because Silicon Valley's commercial interests are too strong to be checked by nonprofit governance. It wasn't a 'Fight over Safety'; it was a 'Fight over Control' between the 'Academic' worldview and the 'Entrepreneurial' one.
Structural analysis: The board never had real oversight capacity — it had a legal veto with no enforcement infrastructure, no information advantage, and no coordination mechanism to survive a counter-mobilization. The crisis wasn't caused by external commercial pressure overwhelming internal safety governance; it revealed that the governance structure was always a mechanism design failure: an agent-monitoring system with no ability to verify agent behavior, no aligned incentives, and no credible threat beyond a nuclear option that destroyed the principal's own interests when used. The 'Information Asymmetry' of the 'Loss of Candor'—the board had 'Data' (or a feeling) that the 'Market' didn't have, but they didn't have the 'Signaling' power to make the market believe them.
The popular framing treats this as a values conflict (safety vs. commercialization) when it was primarily an information and incentive design problem. This matters because it leads to prescriptions focused on 'stronger safety boards' rather than governance architectures that actually align incentives, provide information flow, and build in reversible escalation mechanisms. Solving the values problem without solving the mechanism design problem produces the same outcome.
