Inside the Deposition That Showed How OpenAI Nearly Destroyed Itself

Ilya Sutskever sat for practically 10 hours of videotaped testimony within the Musk v. Altman lawsuit, again on October 1 of this yr.

The co-founder who helped construct ChatGPT and have become notorious for voting to fireside Sam Altman in November 2023 was lastly beneath oath and compelled to reply. The 365-page transcript was launched this week.

What it reveals is a portrait of sensible scientists making catastrophic governance choices, unverified allegations handled as details, and ideological divides so deep that some board members most well-liked destroying OpenAI somewhat than letting it proceed beneath Altman’s management.



The Musk v. Altman lawsuit facilities on Elon Musk’s declare that OpenAI and its CEO, Altman, betrayed the corporate’s unique nonprofit mission by turning its analysis right into a for-profit enterprise aligned with Microsoft—elevating high-stakes questions on who controls superior AI fashions and whether or not they are often developed safely within the public curiosity.

For these following the OpenAI drama, the doc is an eye-opening and damning learn. It’s a case research in how issues go unsuitable when technical genius meets organizational incompetence.

Listed below are the 5 most important revelations.

1. The 52-page file the general public hasn’t seen

Sutskever wrote an in depth case for eradicating Altman, full with screenshots, and arranged right into a 52-page temporary.

Sutskever testified that he explicitly mentioned within the memo: “Sam displays a constant sample of mendacity, undermining his execs, and pitting his execs towards each other.”

He despatched the memo to unbiased administrators utilizing disappearing e mail know-how “as a result of I used to be frightened that these memos will someway leak.” The complete temporary has not been produced by way of discovery.

“The context for this doc is that the unbiased board members requested me to arrange it. And I did. And I used to be fairly cautious,” Sutskever testified, saying that parts of the memo exist in screenshots made by OpenAI CTO Mira Murati.

2. A year-long recreation of board chess

When requested how lengthy he’d been contemplating firing Altman, Sutskever answered: “Not less than a yr.”

Requested what dynamics he was ready for, he mentioned: “That almost all of the board shouldn’t be clearly pleasant with Sam.”

A CEO who controls board composition is functionally untouchable. Sutskever’s testimony reveals he understood this completely and adjusted his technique accordingly.

When board member departures created that opening, he moved. He was taking part in long-term board politics, regardless of how shut Altman and Sutskever appeared publicly.

3. The weekend OpenAI nearly disappeared

On Saturday, November 18, 2023—inside 48 hours of Altman’s firing—there have been energetic discussions about merging OpenAI with Anthropic.

Helen Toner, a former OpenAI board member, was “essentially the most supportive” of this course, in accordance with Sutskever.

If the merger had occurred, OpenAI would have ceased to exist as an unbiased entity.

“I do not know whether or not it was Helen who reached out to Anthropic or whether or not Anthropic reached out to Helen,” Sutskever testified. “However they reached out with a proposal to be merged with OpenAI and take over its management.”

Sutskever mentioned he was “very sad about it,” including later that he “actually didn’t need OpenAI to merge with Anthropic.”

4. “Destroying OpenAI might be per the mission”

When OpenAI executives warned that the corporate would collapse with out Altman, Toner responded that destroying OpenAI might be per its security mission.

That is the ideological coronary heart of the disaster. Toner represented a strand of AI security pondering that views fast AI improvement as existentially harmful—doubtlessly extra harmful than no AI improvement in any respect.

“The executives—it was a gathering with the board members and the manager staff—the executives informed the board that, if Sam doesn’t return, then OpenAI will probably be destroyed, and that is inconsistent with OpenAI’s mission,” Sutskever testified. “And Helen Toner mentioned one thing to the impact that it’s constant, however I believe she mentioned it much more instantly than that.”

In case you genuinely believed that OpenAI posed dangers that outweighed its advantages, then a pending worker revolt was irrelevant. The assertion helps clarify why the board held agency at the same time as 700+ workers threatened to depart.

5. Miscalculations: One supply for all the pieces, an inexperienced board and cult-like workforce loyalty

Almost all the pieces in Sutskever’s 52-page memo got here from one individual: Mira Murati.

He did not confirm claims with Brad Lightcap, Greg Brockman, or different executives talked about within the complaints. He trusted Murati fully, and verification “did not happen to (him).”

“I absolutely believed the knowledge that Mira was giving me,” Sutskever mentioned. “In hindsight, I notice that I did not comprehend it. However again then, I assumed I knew it. However I knew it by secondhand data.”

When requested in regards to the board’s course of, Sutskever was blunt about what went unsuitable.

“One factor I can say is that the method was rushed,” he testified. “I believe it was rushed as a result of the board was inexperienced.”

Sutskever additionally anticipated OpenAI workers to be detached to Altman’s elimination.

When 700 of 770 workers signed a letter demanding Altman’s return and threatening to depart for Microsoft, he was genuinely stunned. He’d basically miscalculated workforce loyalty and the board’s isolation from organizational actuality.

“I had not anticipated them to cheer, however I had not anticipated them to really feel strongly both approach,” Sutskever mentioned.