Reported by New Intelligence Yuan
Editor: KingHZ
[New Intelligence Yuan Digest] From intercepting Peter Thiel and warning Elon Musk to now publicly stating that "one must be adaptable," Demis Hassabis has made an epic pivot: the window for AI safety is permanently closing. He no longer fantasizes about institutions; instead, he is betting his entire fortune—betting on influence, on conscience, and on himself.
The person most afraid of AI destroying humanity is frantically building AI!
The person who believed most in governance now admits: governance is completely unreliable!
DeepMind, which emphasized safety the most, has also been dragged into a wartime arms race!
Just hours ago, Demis Hassabis, the founder of DeepMind, personally shattered the illusion of AI safety, transforming completely from an idealist to a realist.
He admitted: Superintelligence could extinguish humanity, yet the only option left is to "grab a seat at the table" to escort it through.
Hassabis, the man who once vowed to lock AI in a cage with strict institutions, laws, and ethics committees, sat under an old chandelier and helplessly admitted:
"Safety no longer depends on governance structures. Even with a board of directors, at critical moments, they may not do the right thing."
He then proposed a new solution that is extremely realistic, even carrying a tinge of tragedy: "I must sit at the table and fight for influence."
After Repeated Illusions Shattered, Hassabis Turns Around
Hassabis is not your average accelerationist.
Generally, AI leaders are divided into two camps: one represented by Sam Altman of OpenAI, the "accelerationists," who believe technology can always solve the problems it creates; the other consists of the worried "doomsday prophets," who believe AI could lead to human extinction just like nuclear weapons.
But Hassabis is an extremely complex contradiction: he decided to build AI personally in order to prevent the apocalypse.
This nearly paradoxical motivation stems from his profoundly deep anxiety.
Hassabis knows the "risk of AI extinction" inside out, familiar with it even more than the most radical protesters.
He met Shane Legg, the other co-founder of DeepMind, at a lecture on AI safety.
Currently, Shane Legg serves as Google DeepMind's Chief AGI Scientist, co-chairs the AGI Safety Committee, and also leads research into what the world might look like in the post-AGI era.
At the "Singularity Summit," he once intercepted Peter Thiel to sell him a vision of how machines could surpass humans.
He once told Elon Musk to his face that colonizing Mars was meaningless, because if superintelligence is malicious, it could also build rockets spanning the stars; Mars would never be a sanctuary for humanity.
This anxiety gave birth to his early, highly heroic concept of the "Singleton Scenario":
He seriously planned to take the world's top scientists and hide in a secret "bunker"—perhaps in the Moroccan desert or on some isolated island.
There, like the Manhattan Project, free from capital pressure, geopolitical competition, or secular interference, they would develop the first "absolutely safe" superintelligence on behalf of all humanity.
Was Hassabis Betrayed by Musk?
"As long as only one hand in this world holds the torch, we can still control the fire." This was Hassabis's original logic.
For the sake of this logic, when he sold DeepMind to Google in 2014, he set the most peculiar acquisition terms in history:
Even after Google paid, Hassabis insisted on maintaining absolute independence in AI safety matters;
An external committee composed of independent experts would be established to supervise the entire process;
Military applications of AI would be strictly prohibited.
Since his youth, Hassabis has been determined to build strong AI. Only by ensuring its safety could he find a justifiable reason for his life's work.
Shortly before the Google acquisition, Hassabis reminded DeepMind researcher candidates that if they signed on, they must be psychologically prepared—when the final battle arrives, he would enter an underground base to achieve the feat of singleton governance.
In 2015, to implement the AI oversight committee Google promised to establish, DeepMind organized a secret gathering, inviting philosophers and technical experts.
Hassabis invited Musk to host the gathering, which was held at Musk's SpaceX headquarters in California.
But this move backfired. This gathering marked the beginning of the disintegration of Hassabis's safety concept.
Musk listened to the presentation by Hassabis and his co-founders, then went in the opposite direction. He joined forces with Altman to found OpenAI.
At that moment, Hassabis's safety utopia began to die.
Lost the Core, and Lost Face Too
Considering human nature, this situation was inevitable. Faced with the epic prospect of AI technology akin to drilling wood for fire, people do not coalesce into a single collective. They argue endlessly, envy each other, and form their own camps.
At this point, Hassabis could have stopped to rethink.
If the "single entity" assumption was too naive, how could AI developers avoid a race toward the cliff?
But Hassabis did not stop; instead, he accelerated. In 2016, AlphaGo emerged out of nowhere. The timeline for machine intelligence surpassing human intelligence was suddenly brought forward.
To keep up with this AI progress, Hassabis proposed new ideas, hoping to make AI serve humanity. Together with co-founder Mustafa Suleyman, he began negotiations with Google, seeking to establish a new set of governance safeguards.
To advance this secret "Project Mario," he hired a top-tier legal team, secured a $1 billion funding commitment from external sources, and considered spinning DeepMind out of Google if control over the technology could not be obtained.
Meanwhile, Suleyman led the DeepMind team to participate in a project helping the UK's national acute kidney injury management.
If independent AI governance could be achieved while also improving the health of ordinary British citizens, Hassabis might have been convinced that his pursuit was indeed for the greater good.
However, both efforts ended in failure.
The tug-of-war with Google over governance rights lasted three years and was fraught with difficulties.
The project helping the UK's National Health Service (NHS) faced strong backlash from privacy advocates—they were angry that a subsidiary of an American tech giant might access patient data.
By 2019, Hassabis and DeepMind had retreated on both fronts. Suleyman eventually left DeepMind as well.
ChatGPT's Fatal Blow, Google AI Goes on a Rampage
In 2022, the explosion of ChatGPT was like a nuclear bomb dropped into Hassabis's office.
Before this, Hassabis was still trying to maintain "scientific elegance."
He led his team to develop AlphaFold, solving the protein folding problem and winning a Nobel Prize.
He believed this was the highest realm of AI serving humanity: pure, beneficial to medicine, and controllable science.
But the market thought otherwise. When ChatGPT became the fastest-growing application in history, Google panicked, and the entire Silicon Valley went crazy.
On the day ChatGPT exploded in 2022, Hassabis set the tone internally: We are in a state of war!
In this state, all previous safety commitments became fragile:
Google began actively contacting defense departments, attempting to sell AI to the Pentagon (which was once Hassabis's red line).
In the power struggle, the original "independent ethics committee" became almost invisible.
To catch up with GPT-4, DeepMind forcibly merged with Google Brain; speed replaced rigor.
Hassabis saw a cruel reality clearly: In the AI race, no non-profit institution can parasitize within a for-profit giant.
When survival is threatened, all ethics become a luxury.
Hassabis watched with his own eyes as all the walls he had built with his own hands fell down, one by one.
He finally admitted: "Even with a governance committee, wrong decisions may be made at critical moments."
What is truly out of control is not the model, but the human race.
The Lonely Human Watchman
Now, Hassabis has completely turned to realism.
His goal is to secure "a seat at the decision-making table so that when safety issues arise," he can participate in deciding the solutions.
"Things are not black and white, especially when you are facing a technology with unknown consequences," he told reporters. "So you must be adaptable. You have to transform from an idealist to a realist, but hopefully, you can still stick to your values."
Hassabis has completely disillusioned himself with "governance," abandoning institutional safety to rely on personal influence to "grab a seat at the table."
While frantically pushing AlphaFold to win the Nobel Prize and charging ahead with Gemini, he tightly grips hope in the hands of "conscientious people."
An even harsher reversal has come: even Musk, once the most radical safety advocate, has started slamming on the gas pedal.
Optimists are still shouting that "alignment technology can solve everything," but Hassabis himself no longer believes it.
Institutions are completely dead.
The OpenAI board fired Altman, only to be instantly counter-killed by capital forces.
He chose the most helpless path: "To make oneself part of the power structure."
His current safety logic is: Since I cannot stop this race, I must win this race. I must stay at the table; I must possess the highest voice.
This way, when the true "Singularity" arrives, when that decision which humans cannot understand needs to be made, at least the person sitting in that critical position will be Hassabis, someone who has revered AI since day one.
Now he can only gamble—betting that "good people in power" can save the world.
But the question remains: If AI safety can only rely on "good people in power," is this redemption, or the most dangerous gamble of all?
The window for AI safety is permanently closing.
If superintelligence fails to align, the rockets built and the space colonized will not escape total destruction.
Human nature has already turned unified governance into an eternal joke.
The only hope left is this most fragile and dangerous solution: "conscientious people holding the power of speech."
Hassabis is not the only one who has changed; he is just the first to speak the cruel reality thoroughly.
Now, the 37th move has already landed on each of us—
What you and I are witnessing is a civilization-sized gamble.
References: