May 11, 2026 | Xiangming Technology
This story is news that happened in the real world today—May 11, 2026.
And on the very same day, Anthropic's valuation approached 1 trillion USD, 75 OpenAI employees became billionaires overnight, and Claude's latest model directly "blew past" the evaluation ceiling of the METR evaluation organization, to the point that evaluators began asking an incredible question:
"When AI's capabilities exceed the tools we use to evaluate it, what should be downgraded—the evaluation standards, or our understanding of ourselves?"
In today's article, I want to use three "logical skeletons" to help you sort out the chaotic signals of this day clearly.
"It's not that AI has become stronger; it's that thespeedat which it becomes stronger has increased."
In 2018, Google's BERT model had 340 million parameters. Training it once took a full 4 days.
In 2023, when GPT-4 was released, everyone said: "Wow, it can win exams." But for a 16-hour long-horizon task, its success rate was almost zero.
By today, May 2026, everything is completely different.
Today, METR—the world's most authoritative AI capability evaluation organization—published data that made even them feel a chill down their spines. In their "50% success rate time horizon" test, Claude Mythos Preview has already reached a 50% success rate on complex long-horizon tasks that humans need16 hoursto complete.
16 hours.
What kind of concept is that? Last year's best model could only complete tasks on the scale of 2-4 hours. The model from the year before that could mostly only do tasks of a few minutes.
This is an "S-shaped" curve. Except the "elbow" of this curve bends more steeply than anyone imagined.
Former member of OpenAI's superalignment team Leopold Aschenbrenner once predicted that 2027 would be the singularity for AGI. He drew a trend line and, based on the pace of development at the time, estimated that around 2027 AI would reach the capability threshold sufficient for "self-improvement."
What does today's data tell us?
Mythos's performance is already slightly above that 2027 predicted trend line. A full year faster than predicted.
If you think "a year faster doesn't sound like much," then let me put it another way:
When you're driving, a car next to you suddenly overtakes you attwice your speed. Then it keeps getting faster and faster, while you continue driving at a constant speed.
This is not called "it's driving fast." This is called "in your rearview mirror, its image is rapidly enlarging."
Anthropic co-founder Jack Clark recently stated publicly:"AI will achieve self-evolution before the end of 2028."
What does "self-evolution" mean? It means no longer needing human engineers to write code or change parameter counts. AI designs AI itself. Tests AI itself. Iterates AI itself.
Once this "recursive self-improvement" loop starts, the growth curve is no longer "exponential growth"—it is "hyperexponential growth."
In Anthropic's internal words: this is not a stone rolling down a hill faster and faster; this is an avalanche.
🔑 Run-style summary:In the past, competition was a question of "who is stronger." Today's competition is a question of "who evolves faster." And the challenger you face—its speed of evolution is itself accelerating. This is a competition with a positive "second derivative." If you only stare at your own annual growth rate, then you have not understood the situation before you at all.
"Money itself does not create value, but the direction of money's flow shows the direction in which value is being recognized."
Let's look at some numbers.
Anthropic, this AI safety company founded only in 2021, had full-year revenue of 10 million USD in 2022. Not much, right?
In 2023, revenue was 150 million USD.
In 2024, revenue grew to about 6 billion USD.
By May 2026—this very month—its annual recurring revenue (ARR) had soared to45 billion USD。
From 6 billion to 45 billion, in just 5 months. 7.5 times.
This revenue curve, in business history, I cannot find a precedent for.
Data from the prediction market Kalshi: the probability that Anthropic announces an IPO before November 1, 2026, is as high as 72%. Market rumors say its valuation is heading toward 1 trillion USD. If it successfully goes public, it will surpass SpaceX and become one of the largest IPOs in history.
But you might ask: a company that isn't even profitable yet—how can it be worth a trillion?
My answer is:The capital market is no longer pricing AI companies using "today's financial statements." It is pricing them using "what the world will be like in 2028."
In other words, at this stage, capital is not betting on "how much money Anthropic can make next year." It is betting that "if AI truly achieves self-evolution in 2028, the foundations of this world will be reconstructed—then everyone who has seized a piece of land on that foundation today will be a shareholder of the new world."
This sounds crazy.
But to put it more bluntly, let's look at the other side:
Nvidia today announced its total investment in the AI industry for 2026. That number is 45.3 billion USD—about 308 billion RMB. China's most closely watched AI company, DeepSeek, is valued at only around 350 billion RMB. In other words, the money Jensen Huang is pouring into AI this year is "almost enough to buy a DeepSeek."
And on OpenAI's side, a recent employee share tender offer allowed 75 employees to cash out up to 30 million USD each. These 75 "workers" became billionaires (in RMB) overnight.
The last time Silicon Valley saw a wealth-creation wave of this magnitude was in the early era of Google and Facebook.
But what's different this time is that—the wealth explosion speed of AI companies is faster than in the internet era. You don't need to endure a ten-year "long entrepreneurial road." ChatGPT was released at the end of 2022, and less than four years later, the outline of a trillion-level business empire has already begun to appear.
🔑 Run-style summary:When capital begins to price "a future that has not yet happened but whose probability is rising sharply," the market has already entered the "pre-singularity period." The one yuan in your pocket today is being valued at the same time as one hundred million tomorrow. This huge tension of temporal dislocation is an opportunity for all AI entrepreneurs, and also an abyss for all practitioners in traditional industries.
"The day AI passes all exams may not be so terrifying. What's terrifying is the day AI starts to 'teach' people."
Professor Gowers—a mathematical giant at Cambridge University and a Fields Medal winner—after obtaining "priority access" to ChatGPT 5.5 Pro, casually tossed several open problems in additive number theory into the dialog box.
Then three things happened:
The first,In less than two hours, the AI independently completed a mathematical result. Gowers's exact words were: "Fully qualified to be written into a doctoral dissertation."
The second,The AI received no mathematical "guidance" throughout the process. The most the professor did was at most "Hmm, this direction is good, try expanding on it" and "Can you help me format it in LaTeX?"
The third,After the experiment ended, Gowers sounded a red alert for mathematics students. He said: "If AI's mathematical level continues to develop at the current speed, mathematics researchers will soon face a crisis."
Another Fields Medal winner, Terence Tao, also followed up on this matter. His attitude was relatively milder, but the conclusion was similar: "But digestion belongs to humans."
This sentence is very interesting. What Terence Tao means is: AI can produce in 17 minutes a paper that would take you a year to write. But once the paper is written, you have to be able to understand it, you have to be able to provide review comments, and you have to further deduce on the basis of what it proposes.
What if you can no longer even "understand what it is doing"?
This is not a distant problem. This is a real dilemma that is about to happen, right now, to mathematics doctoral students.
And what is even more worth pondering is: this dilemma is spreading from the field of mathematics to every field of knowledge.
Eight changes in new AI professions. Tencent Research Institute just released a set of data:
On this year's "May Day" holiday, they conducted a full count of recruitment positions at 7 AI companies (OpenAI, Anthropic, DeepMind, Zhipu, Moonshot AI, DeepSeek, Tongyi Qianwen).
The results found:
—The number of open positions at the 7 companies rose from 718 in September last year to 1,570 now, more than doubling.
—But more importantly, the "structural change": the proportion of positions "doing models" clearly declined, while the proportion of positions "doing products and commercialization" clearly rose.
What does this mean?
The AI industry itself is shifting from "R&D-intensive" to "application-intensive."
Technology is no longer the biggest bottleneck. The biggest bottleneck has become: how to put AI into real-world business scenarios so that others are willing to pay.
This signal is a serious reminder for those who think, "I'm not in the AI industry, so AI has nothing to do with me."
Because in the stage where the "technology dividend" spills over—when AI's capabilities exceed evaluation standards, capital begins to price the future, and the industry's center of gravity shifts from R&D to commercialization—AI is no longer a "player on the track," but has become the "renovator of the entire stadium."
What it changes is the cost structure of your industry.
What it changes is the yardstick of "how much value a human worker can produce in how much time."
What it changes is even the pricing logic of "how much knowledge itself is worth."
🔑 Run-style summary:In the past, what we worried about was "will AI replace my job." Today, what we should truly worry about is—"what AI replaces is not my job, but the underlying logic of my industry." When AI can write a mathematics paper in 17 minutes, the value of your "ten years of accumulation" in this field is being repriced. This is not anxiety; this is economics. When supply and demand change, prices must change.
In today's article, I used three skeletons to help everyone understand the AI news that happened on May 11, 2026:
I often talk about a concept:"Friends of time" and "enemies of time."
For some things, as time passes, your accumulation grows more and more valuable. This is called a "friend of time."
But for some things, as time passes, the "system" you rely on for survival suddenly changes, and your past accumulation is reset to zero. This is called an "enemy of time."
The AI singularity acceleration we face today is essentially creating a huge time fault zone.
At the two ends of this fault zone, the definitions of the three most basic things—"knowledge," "labor," and "wealth"—are being rewritten.
But please note: a fault zone is not a disaster. A fault zone isopportunity。
Because at the moment of rupture, everyone is on the same starting line.The advantages accumulated in the past have been erased, but the burdens accumulated in the past have also been removed.
Those who quickly find new anchors on this fault zone—whether they are people building AI products, people devoting themselves to AI commercialization, or people using AI to redefine their own industries—will all become the first settlers of the new world.
And those who are still using "last year's logic" to look at "this year's AI," when the 2027 singularity truly arrives, will face not "falling behind"—but "losing their voice."
Just like today, when the METR evaluation organization faced Claude Mythos's data, it found that it had no mature standard it could use.
This is not the limit of AI.
This is the limit of our understanding of AI.
And the first step to breaking through this limit is to admit:The "ceiling" we thought existed is merely AI's "floor."
Recognize this, and then, set out again.
Room 1409, Building C, Shenzhen Century Industrial Center, No. 743 Zhoushi Road, Hezhou Community, Hangcheng Sub-district, Bao'an District, Shenzhen
