At The Heart of Altman’s OpenAI Feud Was A Research Paper – Here’s What It Said

This is not investment advice. The author has no position in any of the stocks mentioned. has a disclosure and ethics policy.

After a remarkable weekend in Silicon Valley that saw its latest innovation, A.I., sitting dead center between typical corporate boardroom politics, OpenAI’s former C.E.O Sam Altman returned to his company triumphantly. Altman’s friction with OpenAI’s board has caused a lot of discourse in social and traditional media, with the disagreements believed to stem from the divide between OpenAI and its holding company’s profit/non-profit nature. However, a for-profit approach might not have been the only reason for the divide, as a fresh report from the New York Times shares that a key source of friction was a research paper written by board member Helen Toner.

Research Paper At Heart Of Altman’s Ouster Characterized OpenAI’s ‘Signalling’ Approach As Inadequate

Ms. Toner is a director at Georgetown University’s Center for Security and Emerging Technology, and in October, she wrote a case study that covered how governments and companies could structure their communication to avoid misinterpretation by others. The paper, co-authored with others associated with Georgetown, defined communications tools called ‘signals’ that national security and A.I. space actors could rely on to clarify their intentions.

The four signals in the paper are tying hands, sunk costs, installment costs and reducible costs. These approaches vary from tying hands, which limits a firm by policy or other announcements that would be difficult to walk back from, to installment costs with higher initial costs (such as costly compliance commitments) that reduce over time as benefits accrue.

On this front, Ms. Toner’s paper specifically focused on OpenAI’s actions surrounding the launch of the GPT-4 model. OpenAI announced GPT-4 in March 2023, and according to the paper’s authors, while the model itself was a remarkable technological achievement, its relevance to signaling came through its supplemental documents.

Costly AI signals as defined in the paper. Image: Examples of Costly AI Signals/Decoding Intentions
Artificial Intelligence and Costly Signals – Andrew Imbrie Owen Daniels Helen Toner (Center for Security and Emerging Technology Georgetown University)

These documents included a system card that explained the risks with GPT-4 and how OpenAI had mitigated them as it rushed to launch the first GPT-4 model in the market. The authors shared that they believed that it was relatively unsuccessful as a communication tool for OpenAI’s commitment to safety.

The paper explains:

While the system card itself has been well received among researchers interested in understanding GPT-4’s risk profile, it appears to have been less successful as a broader signal of OpenAI’s commitment to safety. The reason for this unintended outcome is that the company took other actions that overshadowed the import of the system card: most notably, the blockbuster release of ChatGPT four months earlier.

The researchers then compared OpenAI’s approach to one used by its competitor, Anthropic, and its Claude platform. They lauded Anthropic’s decision to delay Claude’s launch to stop “advanc[ing] the rate of AI capabilities progress” and defined this decision as a “costly signal of restraint.”

As they put it:

By delaying the release of Claude until another company put out a similarly capable product, Anthropic was showing its willingness to avoid exactly the kind of frantic corner-cutting that the release of ChatGPT appeared to spur. Anthropic achieved this goal by leveraging installment costs, or fixed costs that cannot be offset over time. In the framework of this study, Anthropic enhanced the credibility of its commitments to AI safety by holding its model back from early release and absorbing potential future revenue losses. The motivation in this case was not to recoup those losses by gaining a wider market share, but rather to promote industry norms and contribute to shared expectations around responsible AI development and deployment.

The NYT’s report shares that Altman was not pleased with Holt after the paper was published, and he shared his concerns with the OpenAI employees through an email. Additionally, a rather interesting development took place soon after, with OpenAI’s head of research, Ilya Sutsekver, initially debating whether to oust Toner from the board. However, surprisingly, he then chose to move against Altman instead – a decision he would regret within days.

Share this story