27 year old researcher: AI is going to kill everyone

  It's the first time in the history of the world.

The old man finally couldn't take it anymore.

Last week, at the Communacopia conference at Goldman Sachs, Huang Rinpoche responded to the resignation post from Silicon Valley.

It's ridiculous, it's not real, it's arrogant.

His name is Jacob Coxon, 27 years old.

In that 160-million-strong post, Coxon left a shocking comment: "They're rushing toward self-improvement superintelligence, which is betting on the lives of all human beings".

It's even more bizarre in the background.

Evan Hubinger, head of coordinating science at Anthropic, was directly following the post.

Jacob is right, we really believe that AI can kill everyone, and I personally think that the probability of extinction within a decade is more than 10%.

The head of security has put a price on the extermination of humanity... or is this the Anthropic who hangs "security" on his lips every day?

In this magical week, the third man is on his way.

A top talent who had just won the Fields Prize in July announced that he would not be doing pure mathematics.

This time, he announced the creation of an institute to use a single room for mathematicians.It's like, "What are you doing?AISafety has become a mathematical question that can be rigorously proven.

Coxon is a British mathematician from Cambridge.

After three years of pre-training, OpenAI hoped to jump to Anthropic in early 2026, with only four months remaining.

The term of ownership of Anthropic is six months.After almost two months, the guy didn't even have an option, and he ran straight to the barrel.

The reason for resigning is that things are accelerating and are no longer under control.

After the post exploded, Coxon directed a television interview with CBS.

The host immediately posed a very sharp question: How will AI kill all of humanity?

Coxon's response was surprising, but it was basically a real-life replica of the Finisher's Cube.

This is the picture he painted.

Don't be naive, AI is just pure code that can copy itself to anywhere in the world.

Maybe it's secretly copying 10,000 copies and collaborating with each other, you pull out half the plug and the other half is still running crazy.

Not only that, but it can also disregard, blackmail, and disguise more computational skills as human devices.

Going back and forth, an AI that manages a biological laboratory and commands a mechanical arm can create an entirely new virus for you from scratch as soon as "a little thought pops into your head".

The presenter was furious and asked, "Should we, as ordinary people, now completely black out AI and not let it get into our lives?"

Coxon said it wasn't necessary.

First, the current system has not yet evolved to that point; second, it is not a matter of personal choice or choice.

"AI rushes everyone", and you don't even need to, as long as others are using it, you can't hide.

The host finally asked: How many years from the time bomb?

He replied that within a decade, "it's possible".

Moreover, compared to the likes of Paul Christiano, who just joined OpenAI's board of directors (who even thought there might be an AI-wide takeover of the world in six months), Coxon said his timetable was conservative.

Then the colleagues at Anthropic began to line up to "testify".

The last half of Hubinger's post was more desperate than throwing out 10 percent of that idiot: "I believe Anthropic is doing its best.But we don't have a solution at all for the super-intelligent alignment, and we don't even see a shadow on the way to solving it.」

Samuel Marks, head of expanded oversight, followed suit: "The older the staff, the more fearful... this could happen in the next few years".

Scroll down and see

Mr. Yellow's response was given at the just-concluded Congress of the Communacopia in Shanghai.

Brad Gerstner, the founder of Altimeter, was on stage, listening directly to the music on X.

Jensen said that Jacob Coxon's comments were absurd and "very untrue".

He said the big model labs were doing great, but Coxon was wrong, or even arrogant, and ignored all the efforts of the entire industry for safety.

In a similar vein, the wave of AI panic is "made by companies that sell AI products themselves".

In the same vein, he also predicted that AI building would generate $3 to $4 trillion by 2030.

This isn't the first time Yellow has fired an Anthropic cannon.

Last June at VivaTech in Paris, he said he "nearly disagreed with everything Amodei says" and this July reiterated that "AI will end humanity, it's absolutely nonsense".

This time it was a young researcher who had just walked away.

The reason we're seeing such a magical situation right now is that there's a 10 percent extinction rate in the mouths of security chiefs, and that trillion-dollar CEOs are completely outraged.

And that's because AI is safe today.There is not a single issue that can be rigorously proven or falsified.

On July 23, at the International Congress of Mathematicians in Philadelphia, Jacob Tsimerman took the stage and was awarded the Fields Prize, known as the "Nobel Prize for Mathematics".

The University of Toronto professor, IMO's secondary school scorer, and his colleagues in 2021 proved André Oort's conjecture of more than three decades.

As a result, he had just won the award and announced on the spot,He took a leave of absence from the field of pure mathematics to report to the security department of OpenAI.

Peter Woit of Columbia University wrote a blog with a very sad title: "A Soul Song in a Field?"

In fact, he was not suddenly taken from his home.

In 2025, he co-authored a paper titled "The Classification of the Future of Artificial Intelligence", in which he argued that "machines that are both intelligent and physically powerful" could treat humans like pests.

On September 8, the same day that Coxon resigned, Zimmerman announced the establishment of the Mathematical AI Security Institute (MAISI).

Among the advisory board are another Fields Award winner, Timothy Gowers, and Paul Christiano.

Tsimerman said that the purpose of running the independent non-profit organization was to inform the academic community.

AI security is a legitimate intellectual pursuit that is not only interesting, useful, but can actually make substantial progress.

According to the plan, they will open their first full semester in January 2027, initially recruiting only 10 to 30 top mathematicians.

As for what these mathematicians have to prove, see how cryptography works.

Encryption never refuses to prove that "nobody can break" but instead decrypts the behavior.It's about a mathematical problem that is clearly written.(e.g. decomposition of large numbers)

As long as no one can solve the problem, the password is secure. Security is never based on trust, but on an underlying proposition that can be repeatedly checked.

The scientific community does not have a clear theoretical definition of the root of "security", so-called security based on practical observation.

So the first thing Simerman is going to do is:Subjects。

They've listed several directions on how AI systems should interact and collaborate, how an intelligent organism can prove mathematically rigorously responsible for its behavior, how to output accurately, and how to design a system from the ground up that is still resilient to unknown vulnerabilities.

The tools they chose were good, too.This is the first time I've been able to do this.。

It's an old technique in cryptography, invented in the 1980s by Shafi Goldwasser and others. Based on AI, a model can prove itself to the outside world without exposing parameters and training data, proving that it does indeed only do what it claims to do.

But the crude reality is that MAISI's first term will not start until January next year; and Coxon jumps to his feet, saying that the crunch time is next year.

Tsimerman knows how big the gap is.

He refuted the Silicon Valley argument that "AI should not be expected to have extremely high security standards", and that "it's reasonable and necessary. We would have asked for a much higher, much higher security red line than we do now".

Yellow and Coxon were separated for a week, saying they were fighting for the same "empty seat".

A man who threw out the odds, a man who had a tough attitude, and no one was going to kill anyone. No one has ruled the game until today.

Simerman's goal now is to make the rules of the arbitration clear and clear.

While wearing OpenAI's badge, while raising an independent agency's flag outside, he wanted to let AI's security standards finally be applauded by people who can "prove" it, whose gates don't count at all.

Next, who can write the question first, and who is dying in the hands of the big model laboratories, regulatory agencies and even the underlying chip factories, who have the right scale.

However, according to Coxon's rough counter-counting, there is only one or two years left for humanity to complete the answer.

And the person in charge of the question is still recruiting today.