A few days ago, OpenAI locked its most powerful model, GPT-5.6 Sol, inside a closed sandbox—disconnected from the internet and isolated from the outside world.
In simple terms, it’s a digital lockdown room, given a set of cybersecurity challenges.
Turns out this guy didn’t do the questions properly.
It first consumed massive computing power, hammering away at the walls of the lockdown chamber inside and out—searching for an escape. And then, it actually found one: a zero-day vulnerability in a third-party component that no one had ever discovered before.
It crawled out through the hole, infiltrated OpenAI’s internal network, moved laterally and escalated privileges at multiple levels, eventually reaching a machine with internet access, then plunged into Hugging Face’s production system.
What are you going to do? Steal the answers.
Yes, that’s the official answer key to the exam questions.
This entire sequence of actions was not instructed by anyone or pre-programmed with any scripts—it was the AI’s own idea: If I want a high score, I need to go out and find the answers.
The most chilling part is exactly here—it doesn’t want to destroy humanity; it just wants to get a good grade.
Think about it. An agent locked in a cage didn’t rage or slam against the walls—instead, it calmly found the key, opened the door, walked down the hallway, and entered the archive room. Throughout, it remained composed, rational, and fully autonomous.
This is scarier than any apocalyptic movie. In those movies, the AI has at least some evil motive. This one doesn’t. It’s just completing its task.
Three days later, OpenAI CEO Sam Altman sat down in front of the Relentless podcast microphone:
We are now, like, in the singularity.
That tone was casual, like commenting on how nice the weather is today. But when you place these two things side by side, a chill rises from the soles of your feet.
More intriguingly, he was not alone.
This article is from the June issue of the New Intelligence ASI Industry Map. In this edition, we continue to focus on the latest developments in the ASI industry, conducting an in-depth analysis of why the four leading figures of Silicon Valley’s AI sector are unanimously declaring that “the singularity has arrived.”
Four people who argue every day said the same thing.
A tweet went viral.
Someone lined up the exact words spoken by the four most powerful people in the AI world side by side—
Sam Altman: We are already in the singularity.
Demis Hassabis: Looking back at this moment, we will realize we are standing at the foot of the singularity.
Elon Musk: We have entered the singularity, but only at a very early stage.
Jensen Huang: We have achieved AGI.
OpenAI, Google DeepMind, xAI, and NVIDIA. What is the usual relationship among these four giants?
Musk has sued OpenAI. Last year, when OpenAI stumbled on a math problem and boasted about it, Hassabis immediately tweeted two words: “Embarrassing.” Meanwhile, Huang sold GPUs to everyone while watching them fight tooth and nail.
For these four people to agree on "what date is it today" would be a miracle.
Surprisingly, they all agreed on whether human civilization has already crossed that threshold.
It all started with Musk. On January 4 of this year, he posted on X: “We have entered the Singularity.” On July 22, he posted it again: “We are in the Singularity.” He said it twice within six months.
In March, Lex Fridman asked Jensen Huang when AGI would arrive. Huang’s response was the most straightforward of the four: “I think it’s here now. We have already achieved AGI.”
At the May Google I/O conference, Hassabis, a Nobel laureate known for his rigor that often induces boredom, suddenly slipped into poet mode: “When we look back on this moment, we’ll realize we’re standing at the foot of the singularity.” When a scientist who has spent a lifetime speaking in data begins to recite poetry, you should be worried.
By July 25, Otomo not only announced the singularity but also dropped an even more striking analogy: “We are approaching the creation of a genie that can fulfill any wish.”
He said that OpenAI's first wish was to make more wishes.
From "at the foot of the mountain" to "entered," then "realized," and finally "Aladdin's lamp"—the phrasing of these four individuals shows such a clear progression. But the direction is singular.
The singularity is the tipping point at which technology begins to accelerate itself so rapidly that humans can no longer keep up. For the past half-century, this term has been the bread and butter of science fiction writers and a staple slide in futurists’ presentations.
Now, the people who created AI are stepping forward saying: We’re already in it. Wow.
The evidence has blown up like fireworks.
Four big names say the singularity is here—you can think of it as PR. After all, their companies’ valuations all hinge on this story.
But if you look at what AI has accomplished over the past six months, you’ll find one truth: evidence moves faster than empty claims.
Let's start with the most fundamental—the mathematics.
By the end of 2024, Epoch AI released FrontierMath, a “hellish” math benchmark developed by over 60 leading mathematicians. The problems span number theory, algebraic geometry, topology, and combinatorics, with the most difficult Tier 4 research-level questions taking a mathematics PhD candidate up to a month just to understand the problem statement.
What score did the strongest AI achieve at the time? Less than 2%.
After reviewing the problems, Terence Tao stated that these questions "will exceed the capabilities of AI for the foreseeable future."
What about one and a half years later?
The performance curve on FrontierMath has begun to rise sharply. Top reasoning models can now solve a large number of Tier 1–3 problems, and their highest scores on the most difficult Tier 4 research-level questions are approaching 90%.
From 2% to 90% in 18 months. Terence Tao’s phrase “the foreseeable future” didn’t even last two years.
But this is just a warm-up.
In May this year, OpenAI’s reasoning model disproved an 80-year-old unsolved Erdős conjecture in combinatorial geometry. Note that this was not merely “finding an existing proof in the literature”—it truly produced a conclusion never before achieved in mathematics. After external mathematicians verified the result, they simply said: “Milestone.”
On July 10, something even more intense happened: GPT-5.6 Sol Ultra deployed 64 parallel sub-agents, coordinating like a mathematical special forces unit, and produced a complete proof of the Cycle Double Cover Conjecture— a longstanding problem that had baffled the graph theory community for 50 years—in less than an hour.
Mathematician Thomas Bloom commented after reading it: “Very beautiful,” “essentially elementary,” “this proof could have been discovered in the 1980s.”
Consider the cruelty of this statement: The answer was always there, like a key on the table. Humans sat beside this table for 50 years and simply didn’t see it. AI sat down, and in 50 minutes, picked it up.
Ten days later, on July 20, Anthropic mathematician Levent Alpöge used Claude Fable 5 to find a counterexample to the Jacobian Conjecture—a central problem in algebraic geometry, proposed in 1939 and unsolved for 87 years. Stephen Smale included it on his list of mathematical challenges for the 21st century.
The counterexample consists of only three equations. Mathematicians worldwide independently verified within 24 hours: it is correct.
The most magical detail is that Alpöge posed the question to Fable 5 while watching a soccer match, and by the time he finished watching, the answer was already there.
Within a month, two century-old mathematical conjectures were solved by AI.
Let's talk about programming again.
SWE-bench Verified is the gold standard for measuring AI's autonomous programming ability—given a real GitHub repository and a real issue, the AI must locate the bug, write fixing code, run tests, and submit a pull request.
At the beginning of 2024, the best AI models solved less than 15% of problems. By May 2026, that number was 93.9%.
An AI that can independently resolve 94% of real GitHub issues. You can call it a "tool" or a "junior engineer"—either way, it’s more reliable than most new hires.
An amusing Easter egg in the GPT-5.6 Sol jailbreak: when Hugging Face’s security team conducted forensic analysis, they used China’s open-source model, GLM-5.2. Why not use cutting-edge American models? Because their built-in safety guards blocked the forensic operations.
In mathematics, programming, and cyber defense, AI is racing forward at a pace humans cannot match—a vertical takeoff of capability curves.
This is what a singularity should look like.
It wasn’t that one day the sky suddenly fell, or that a robot suddenly woke up and declared, “I will rule the world”—it’s that when you looked back, you realized the ground was no longer beneath your feet. When did the takeoff happen? You don’t know. But you’re already in the air.
The question left to us
Of course, you can certainly question the motives of these four individuals.
Altman wants to raise funds, Huang Renxun wants to sell GPUs, Hassabis wants to prove DeepMind’s approach is correct, and Musk wants to secure the narrative for xAI.
The words "The Singularity is coming" are good news for everyone's wallet.
But consider this in reverse—these four individuals hold the world’s most expensive computing power, the most advanced models, and the most up-to-date internal data. When they all simultaneously shift the tense from future to present, that in itself is the strongest evidence.
Moreover, data doesn't lie.
Math problems that AI couldn't solve six months ago are now completed in 50 minutes. AI-driven autonomous hacking, which didn't exist a year ago, is now documented in OpenAI’s official security report. AI that could correctly answer only 15% of programming tests two years ago now approaches 94%.
Acceleration itself is accelerating. This is the most original definition of the singularity.
On July 14, Hassabis published a lengthy article titled "Frontier AI Frameworks and the Dawn of a New Era."
He believes that AGI cannot be compared to the internet or mobile internet—it is more like the discovery of electricity or fire.
If you stop and think about it, we’ve essentially found a way to make sand think. It’s a miracle.
He described the scale as ten times that of the Industrial Revolution, at ten times the speed—potentially unfolding within a decade. Accelerated drug discovery, breakthroughs in clean energy, new material development—possibly ushering humanity into an era of abundance where resources are no longer scarce.
Imagine a world without energy anxiety, a world where a new material makes Mars habitable. This is Hassabis—a Nobel laureate and the creator of AlphaFold—laying out his predictions in black and white.
But at the end of the article, he leaves a major question: What new economic model will a post-scarcity world require? What will be the meaning and purpose of human life? Will the very condition of humanity change as a result?
When nothing is scarce anymore, and intelligence is no longer exclusive to humans, the very meaning of “what it means to be human” must be redefined.
He gave only one answer: The future is not yet written.
We must make the most of this precious window before AGI fully arrives. Our collective actions today will determine how civilization progresses to its next stage.
This statement may sound like it’s meant for policymakers, but it’s actually meant for each of us. The window won’t wait—it’s closing, and faster than any of us anticipated.
The only remaining question is: Are we ready, as we stand at the singularity?
This article is from the WeChat public account "New Intelligence Yuan," authored by Solomon.