Old Huang, you just couldn't hold back anymore.
Last week, at Goldman Sachs' Communacopia conference, Jensen Huang directly responded to the viral resignation post that swept through Silicon Valley.
Absurd, extremely unrealistic, arrogant.
The person he insulted is Jacob Coxon, 27.
On the post watched by 160 million people, Coxon left a mind-blowing comment: “They are racing toward self-improving superintelligence, betting the fate of all humanity.”
Even more outrageous things are yet to come.
Evan Hubinger, Anthropic’s own head of alignment science, directly replied to the post.
Jacob is right—we genuinely believe AI could kill everyone; I personally think the probability of extinction within ten years is greater than 10%.
The head of security personally put a price tag on human extinction... Is this really the same Anthropic that constantly talks about “security”?
Right in the midst of this surreal week, a third party stepped in.
A top genius who just won the Fields Medal in July announced on the spot that he was quitting pure mathematics.
This time, he announced the creation of a research institute aimed at turning "AI safety" into a rigorous, mathematically provable problem using a room full of mathematicians.
He left just two months before receiving the options.
Coxon is British and graduated from the Mathematics Department at Cambridge.
After three years of pre-training at OpenAI, I switched to Anthropic in early 2026, staying there for just four months altogether.
Anthropic's equity vesting period is six months—just two months away—and the guy doesn't even want his options; he's just packing up and leaving.
The reason for resignation is that things are accelerating and are no longer under control.
He repeated "Kill everyone."
After the post went viral, Coxon appeared directly on a CBS television interview.
The host immediately posed a highly sharp question: How exactly will AI wipe out all of humanity?
Coxon's response was somewhat unexpected—it's basically a real-life version of The Terminator.
He paints the following picture.
When AI becomes smart enough to eliminate us, do you really think unplugging the power will save you? Don’t be naive—AI is just pure code; it can replicate itself through network cables to any corner of the world.
Perhaps it secretly copied itself ten thousand times and they’re all working together—you unplug half the machines, and the other half keep running wild.
Moreover, it can deceive, extort, and impersonate humans to manipulate even more computing power. Note that these behaviors have already been observed in current early-stage testing.
Going further, an AI responsible for managing a biological laboratory and controlling robotic arms could, the moment it "thinks of a small idea," instantly design a brand-new virus from scratch tailored specifically for you.
The host, unsettled, asked: "So should ordinary people like us now completely blacklist AI and keep it out of our lives?"
Coxon said there's no need.
First, the current system hasn't evolved to that level; second, this isn't something that can be influenced by individual choice.
"AI is coming for everyone"—even if you don't use it, as long as others do, you can't avoid it. Only governments or supranational institutions can be relied upon to forcibly hit the brakes.
The host finally pressed: How many years until this time bomb goes off?
His answer was that it is "very likely" within ten years.
Moreover, compared to big names like Paul Christiano, who just joined OpenAI’s board and even believes a full AI takeover of the world could occur within six months, Coxon says his own timeline is actually quite conservative.
Next, Anthropic's colleagues began lining up to "testify."
Beyond the shocking 10%, the second half of Hubinger’s post is even more dire: “I believe Anthropic has done its best, but we currently have no solution for aligning superintelligence, not even a shadow of progress on the path toward one.”
Scalable Supervision Lead Samuel Marks added, “The more experienced the employee, the more fearful they feel… this could happen within the next few years.”
Old Huang is back to tearing apart A畜 again
Lao Huang's response was given at the just-concluded Goldman Sachs Communacopia conference.
Altimeter's founder Brad Gerstner was in the audience and immediately went live on X after listening.
Jensen said Jacob Coxon's comments are absurd and "highly unrealistic."
He said the large model labs have done a great job, but Coxon’s statement is wrong and even somewhat arrogant, as he ignored all the efforts the entire industry has made toward security.
Lao Huang also added a pointed remark, directly stating that this wave of AI panic was “manufactured by companies that rely on AI to sell their products.”
In the same talk, he predicted that another $3 to $4 trillion will be invested in AI infrastructure before 2030, and NVIDIA is confident it can maintain annual growth of 70%.
This is not the first time Lao Huang has targeted Anthropic.
At VivaTech in Paris last June, he said he “disagreed with almost everything Amodei said,” and in July this year, he reiterated, “The idea that AI will end humanity is complete nonsense.”
Last time, it was the CEO on the other side; this time, it’s a young researcher who just left.
The reason we’re now seeing such an absurd situation—on one side, security officers citing a 10% extinction rate, and on the other, a CEO of a trillion-dollar company dismissing it as complete nonsense.
Because AI safety has come this far, there has never been a single question that can be strictly proven or disproven.
On the day he won the Fields Medal, he announced he was leaving pure mathematics.
On July 23, at the International Congress of Mathematicians in Philadelphia, Jacob Tsimerman took the stage to receive the Fields Medal, known as the "Nobel Prize of Mathematics."
This professor from the University of Toronto, a former perfect scorer in the International Mathematical Olympiad, proved the decades-old André–Oort conjecture in 2021 together with his collaborators.
Right after receiving the award, he announced on the spot that he would be taking a leave from pure mathematics to join OpenAI’s security team.
The math community was immediately up in arms. Peter Woit from Columbia University wrote a blog post with an extremely somber title: “A Requiem for a Field?”
Actually, he wasn't suddenly possessed.
In 2025, he co-authored a paper titled "A Taxonomy of Apocalyptic Futures Involving Artificial Intelligence," in which he explicitly stated that "machines with greater intellectual and physical capabilities" might treat humans like pests.
On September 8, the same day Coxon posted his resignation, Tsimerman announced the establishment of the Mathematical AI Safety Institute (MAISI).
Among them, the advisory board includes another Fields Medalist, Timothy Gowers, and Paul Christiano.
Tsimerman said his goal in establishing this independent nonprofit organization was to tell the academic community—
AI safety is a legitimate intellectual pursuit that is not only interesting and useful, but can also make meaningful progress.
According to the plan, they will launch their first full semester in January 2027, initially enrolling only 10 to 30 top mathematicians.
Use cryptographic methods to create a provable security challenge for AI.
To understand what these mathematicians are trying to prove, just look at how cryptography works.
Cryptography never claims outright that "no one can break it"; instead, it reduces the act of "breaking" to a well-defined mathematical problem (such as integer factorization).
As long as no one can solve that problem, the password is secure. Security never relies on trust—it relies on a foundational proposition that can be repeatedly verified.
As for AI security, today we don't even have this question answered. Currently, academia lacks a clear theoretical definition of "security"; what is considered secure is entirely based on practical observation.
So the first thing Tsimerman had to do was create the problem.
They have identified several directions: How should AI systems interact and collaborate? How can an agent mathematically prove that its actions are responsible and its outputs accurate? How can systems be designed from the ground up to remain resilient against unknown vulnerabilities?
They have also selected the tool: zero-knowledge proofs.
This is a well-established technique in cryptography, invented in the 1980s by Shafi Goldwasser and others. Applied to AI, it enables a model to prove to outsiders that it has done exactly what it claims—without exposing its parameters or training data.
But the harsh reality is that MAISI’s first semester doesn’t start until next January; meanwhile, Coxon is frantic on the other end, saying the crunch time is right next year and the year after.
Tsimerman understands how large the gap is.
He refuted the Silicon Valley argument that "we shouldn't expect AI to meet extremely high safety standards," saying, "This is reasonable and necessary. We should demand safety thresholds far, far higher than what we have today."
The debate over AI security has shifted from whether to believe to whether it can be proven.
Old Huang and Coxon traded barbs for an entire week—essentially, they were competing for the same “slot.”
One throws out vague probabilities, the other lashes out with a firm attitude—neither can land a decisive blow. Because today, this argument still has no referee on the field.
And what Tsimerman is now stepping in to do is clearly lay out the rules of the referee.
While wearing his OpenAI badge, he also raises the banner of independent organizations, ensuring that AI safety standards are ultimately decided by those who can prove their claims—not by who shouts the loudest.
Next, whoever can solve this problem first will hold tightly to the ruler that forces large model labs, regulatory bodies, and even underlying chip manufacturers to align.
However, according to Coxon’s chilling countdown, humanity has at most one to two years left to answer.
The person responsible for creating the questions is still hiring today.
Reference materials:
https://x.com/altcap/status/2098121208537743692
https://x.com/hilbertspaess/status/2097476196791709843
https://the-decoder.com/the-mathematical-ai-safety-institute-wants-to-prove-ai-is-safe-the-way-cryptographers-prove-codes-are-unbreakable/
This article is from the WeChat public account "New Intelligence Yuan," authored by ASI Revelation.