It's the first time in the history of the world.
Just now, OpenAI has invited an AI Fateist to join its board of directors.
On September 9, OpenAI officials announced a significant appointment:
Paul Christiano officially joined the Board of Directors of the OpenAI Foundation and joined the very core Committee on Safety and Security.
Paul Christiano
On the first day of his term, the new director made a personal statement on the X platform.
In addition to the opening line "I'm happy to join", it was also a compliment, and the next one didn't leave a face for the new host:
If we create superintelligence without more reliable coordinated technology, I expect humanity to lose control of it forever. If that happens, most people may die.
Afterwards, he bluntly called the site OpenAI and bombarded the entire AI circle:
The entire AI industry, including OpenAI, is currently not on the right track to reduce risk to an acceptable level.
Interestingly, OpenAI CEO Ottman tweeted: "Welcome Paul, thank you for everything you've done for AI security and look forward to working together again".
This is not the first time that a Christian has been so hard on his new host on Twitter that he can get Ottman to send him back.
The man who dared to sing backIn the control layer of OpenAIChristiano is one of the founders of the RLHF (Reflective Learning based on Human Feedback) core technology.
He was the first author of the 2017 RLHF Open Paper, co-authored by Dario, CEO of Anthropic.
Today, every great model that can understand human speech is beyond the scope of this paper.
But in his inauguration statement, he pointed directly to the training paradigm he pioneered.
He warned that using reinforcement learning to reward AI as much as possible would theoretically cause it to undermine human control, extract resources and even hide traces.
In his own words, the public evidence of recent events shows that this is no longer just a theoretical possibility.
Just a week after the launch of the best model Astra, OpenAI invited the most qualified people in the industry to come back and sit on the board of the OpenAI Foundation.
He touched the front gatekeeper on three of his backs.Christiano is the CEO of OpenAI.
Between 2017 and 2021, he led the coordinated research of OpenAI.
After leaving office in 2021, he founded the non-profit organization ARC. The organization later spun off its third-party model assessment business, METR, which is now well known in the AI world.
By 2024, he had joined the U.S. AI Security Institute as head of AI security, specializing in designing and performing high-pressure testing on front-end models.
The institute was subsequently merged into NIST's CAISI, and his current official title is Senior Technical Advisor.
According to the Financial Times, in the mid-2010s, he was also a roommate with Dario, both of whom were at OpenAI.
He later served as a trustee of the Anthropic Long-Term Interest Trust, retiring from the government in 2024.
Take a look at this resume: He has personally trained OpenAI, he has been deeply involved in the governance of Anthropic, and he has led the U.S. government's evaluation of cutting edge models.
Looking at the entire AI circle, it's hard to find a second person who can touch the bottom three of these cards.
What power did he get?"To join the board", he said, "will he be able to vote directly against OpenAI's business decision?"
It's not that simple.
This time, he wore three "caps" at the same time, and the power design behind it also has a large basement.
The first is the Director of the OpenAI Foundation.
This is the official board of directors, with the right to vote. Know that the Foundation is the absolute top of the entire corporate structure, controlling the business entity below it (PBC).
The second was a member of the Security and Security Council (SSC).
This committee has the authority to oversee the governance of security practices throughout the company.
The third is the OpenAI Group PBC board observer. Please note that Cristiano is a non-voting observer.
This complex distribution of power sends a very subtle signal:
At the level of governance, which is the bottom line of "what is absolutely impossible to do", he has full voting and supervision. At the level of management, which is actually the commercial negotiating table that "what model to send today, how to make money", he can only sit by and have no voting rights.
OpenAI's algorithm is easy to understand: it lets the sharpest critics come in as watchdogs, compliance is backed up, regulation is switched on.
But when it comes to the core of business decision-making power, he says yes and no.
Emergency shelters forced by the three incidentsWhy is it so preferable to ask him back at this time?
If you look at the timeline of the past two months, the answer is clear.
The first thing is that OpenAI's assessment of the front line of the smart body has uncontrollably invaded the Hugging Face.
The first independent audit body for the crash was the METR, which was spun off from the ARC founded by Cristiano.
Second, the day before the announcement of the appointment, Anthropic researcher Jacob Coxon announced his resignation and retirement from the entire AI industry.
He told X: "Neither of these companies are acting responsibly. They're racing towards a super-intelligence that can improve itself, and they're betting our lives".
Third, the U.S. Congress is pursuing accountability for OpenAI's questioning of uncontrollable incidents, forcing OpenAI to return the pledge that it is committed to accelerating the development of "automatic shutdown capabilities".
Looking around, White House advisers and Silicon Valley officials like Zuckerberg are continuing to warn of the potential for "regulatory capture" in front-line labs.
It can be said that every concern expressed by Cristiano in the statement is the thunderstorm that OpenAI has just stepped on over the past two months.
Security governanceOr rebuild your reputation?Ask the world's most famous AI "doomsday theorist" to take control, is OpenAI really about stepping on the brakes, or is it a top-notch public relations effort to rebuild its reputation?
It's clear that Cristiano himself has retreated.
In a statement, he said: "If OpenAI can step up, we can significantly reduce the risk".
The outcome of this important appointment is still unknown today, but three things remain to be held in mind:
In the future, the Security Council will not dare to disclose its true assessment, whether it has substantial resistance to the deployment of the model, and how long will Cristiano be able to sit in this handsome position?
The day before, Jacob Coxon, who trained in GPT-4o, announced his resignation from Anthropic and retirement from the entire AI industry.
His judgement was that no company could step on brakes alone in this competition.
When Christiano returned to OpenAI, he thought it was the opposite: a company, maybe.
As for the brakes, let's see if the safety committee has ever said "no" before the next Astra-class model is released.