In today's AI world, if you had to describe it with just one word, it would be—Disruption。
Just a few hours ago, I came across three pieces of news that made me unable to sit still:
· Ramp released its latest AI index: Anthropic's enterprise market share is 34.4%, surpassing OpenAI's 32.3% for the first time—a throne that had not been shaken for three years was kicked down today.
· Tian Yuandong and eight other top AI researchers founded Recursive Superintelligence with $650 million (about 4.4 billion RMB). The company is valued at $4.65 billion. Their goal: let AI train itself, and then make the profession of AI researcher disappear entirely.
· Zhipu's stock price soared 36.9%, closing at HK$1,150, a record high. Tang Jie posted late at night, proposing a core viewpoint: AI's step from tool to labor is just missing "long-horizon tasks."
· Google announced the "Magic Pointer" at the Android Show—equipping the mouse cursor with an AI brain. The form of AI interaction is changing from "typing typing typing" to "pointing and clicking."
Four things happened on the same day; this is absolutely not a coincidence. They are telling us:The AI industry is moving from one era into another。
In today's article, I want to talk with you about the underlying logic behind these four things that is reshaping the AI world.
Let's start with the biggest news.
The AI index released by Ramp for May 2026 shows: Anthropic's enterprise adoption rate reached 34.4%, surpassing OpenAI's 32.3% for the first time.
Many people may not know much about Ramp. It is a fintech company, and its data comes from billing data of more than 50,000 enterprises and billions of dollars in real spending. The value of this data lies in the fact that it is not a survey questionnaire, not sample statistics, butreal money card transaction records. When CEOs actually take out their credit cards to pay, whom they choose is the most real market vote.
Look at two sets of data and you will know how dramatic this reversal is:
Time| One year ago: Anthropic 9% vs OpenAI 32%
Time| Today: Anthropic 34.4% vs OpenAI 32.3%
Anthropic quadrupled. OpenAI rose 0.3 percentage points.
Isn't this comparison shocking? Actually, the logic is simple—
In May last year, if you asked me: Who sits on the throne of AI companies? I would say OpenAI without hesitation. Today, the data says: the throne has already changed hands.
OpenAI's response is also interesting. An official spokesperson said: "Ramp's data only counts credit card payments; our tens-of-millions-level enterprise transformation contracts do not go through credit cards." Translated, this sentence means—"We don't think the data is real". But at the same time, Sam Altman urgently launched a promotion for "two months of free Codex access."
Saying it's fine verbally, but the body is very honest.
So the question is: How did Anthropic turn things around?
Two killer moves.
First, the brutal aesthetics of the billing model. They made a decision that could be called "against the ancestral rules of SaaS"—abandoning fixed subscription fees and fully shifting to token-based billing.
How powerful is this change? In the past, the logic of traditional SaaS was "whether you use it or not, $30 per person per month." In the AI Agent era, this has become a huge mismatch. Because when AI becomes a "digital laborer" running automatically 24/7, token consumption explodes exponentially.
For example: if an enterprise uses ChatGPT for customer service, and the back-end agents ask 100 questions per day, it consumes a fixed quota per month. But if it uses Claude Agent to run autonomously, it might handle 10,000 conversations a day and automatically call 100 APIs—consumption is on a completely different order of magnitude.
Usage-based billing is equivalent to completely opening up the upper limit of enterprise spending. As a result, Anthropic's annual recurring revenue soared to the $45 billion level—a figure that was only one-third of that at the end of last year.
Second, the "money-printing effect" of AI Agents.
Microsoft has OpenAI in-house, yet this year its spending on Claude reached as high as $500 million. More than 1,000 major clients each spend more than $1 million annually on Anthropic.
What trend does this show?Enterprises are shifting from "buy an AI tool and try it" to "let AI directly do the work". And Anthropic's Claude is indeed one step ahead in making AI "really do the work."
Ramp chief economist Ara Kharazian said something very much to the point: "Anthropic has long been ahead in finance, technology, and professional services, and OpenAI's advantage in other industries is rapidly shrinking."
So, having many ChatGPT users does not mean enterprises are willing to pay. When enterprises vote with real money, they look at ROI, not brand awareness.
If Anthropic surpassing OpenAI is today's most shocking news, then the public debut of Recursive Superintelligence is the most insane news.
Let's look at this company's parameters—
· Co-founders: 8 people
· Total employees: 25 people
· Time since founding: less than half a year
· Valuation: $4.65 billion
· Funding: $650 million
· Investors: GV, Greycroft, AMD Ventures, Nvidia
· Core business: no product yet.
No product, 25 people, $4.65 billion valuation. Isn't this insane?
Let's see who these 8 people are—
1. Richard Socher: Founder of NLP word embeddings, former Chief Scientist at Salesforce
2. Caiming Xiong: Former SVP of Salesforce AI Research
3. Tian Yuandong: Former Research Director at Meta FAIR
4. Alexey Dosovitskiy: First author of Vision Transformer (ViT)
5. Jeff Clune: Pioneer of evolutionary algorithms
6. Tim Rocktäschel: Core researcher on DeepMind Genie world model
7. Josh Tobin: Builder of OpenAI's robotics team
8. Tim Shi: Early OpenAI member, co-founder of an AI customer service unicorn
This roster is almost a list of the authors of the key AI breakthroughs of the past decade. Any one of them, taken alone, is a top-tier giant in AI. What does it mean for eight of them to come together?
It means what they want to do is something too big for one person or one company to do.
What they want to do is called"Recursive Self-Improvement" (RSI)。
What does that mean?
Today, building a frontier large model requires hundreds of people working for months. Data filtering, training design, post-training alignment, research direction selection—every step relies on human researchers. But the problem is: fewer than a few thousand people in the world can do this. And the more complex the model, the closer human ability to understand and optimize it gets to its ceiling.
What Recursive wants to do is hand over data filtering, training, post-training, and research direction selection—all of it—entirely to AI itself. The whole chain closes the loop, with no humans needed from start to finish.
What does this mean? A positive feedback loop:
AI improves itself → the improved AI is better at improving itself → the loop accelerates → the speed exceeds human control
Richard Socher has a line worth pondering. He said there are three stages of neural networks:
The first stage: Neural networks learn to extract features themselves → feature engineers lose their jobs
The second stage: Unified models kill task-specific architectures → companies in niche tracks disappear
The third stage: AI learns to train itself → AI researchers lose their jobs
Do you see it? What they are doing is making the third stage a reality—and then making their own profession disappear.
Betting on your own unemployment—this is probably the craziest bet an entrepreneur can make.
But if you think carefully, this is precisely the most extreme rationality. If recursive self-improvement can really be achieved, then all of today's AI companies—including Recursive's own current model—will be overturned. But if you don't do it, someone else eventually will. Rather than wait for others to overturn you, it's better to overturn yourself.
This is a typical "ultimate risk hedge": if AI evolves to the point where it no longer needs humans, then at least it was built by "ourselves" with our own hands.
On the same day, Zhipu's stock price soared 36.9%, hitting a record high in market value. And less than 24 hours before that, founder Tang Jie posted a long tweet on X late at night.
The core is just four words:Long-Horizon Tasks。
In English it's "Long-Horizon Tasks"—meaning a task that requires spanning a longer period of time, more steps, and more intermediate states to complete. For example: from vulnerability code analysis to full attack chain deduction; from contract clause review to complete compliance report output.
Tang Jie said something I really like:
"Once a model can continuously plan, trial-and-error, judge, and deliver, what it disrupts is not just programmer efficiency, but the entire human execution layer."
How should this sentence be understood?
Over the past two years, we evaluated AI by looking at a single conversation. Whether the answer was accurate, whether the logic was clear, whether the language was fluent. This is essentially testingintelligence。
But long-horizon tasks don't need intelligence, they needexecution capability。
What is execution capability? The goal is clear but the path is uncertain, and the task is advanced continuously for hours, days, or even weeks. In the process there is trial and error, judgment, and mid-course adjustment.
Tang Jie used hackers as an example. Vulnerability discovery is a typical long-horizon task: it requires reading a large amount of code, understanding the system architecture, setting up a testing environment, constructing attack inputs, verifying the effectiveness of the vulnerability, and finally writing a technical report. There is no standard answer; it all depends on trial and error.
What is Zhipu doing? What they are doing is letting AI evolve from a "conversational tool" into "long-horizon labor." Once that is achieved—
In the past, when AI wrote a few sentences for people, everyone could still say it was just a tool.Once AI can work on its own for several days in a row, judge on its own, and deliver on its own, what it will replace is not just a certain position, but the complete replacement of an entire industry.
Think about sales assistants, market researchers, junior lawyers, auditors—what is the essence of these positions? It is not producing one perfect answer, but continuously following up on a complex task and iterating until delivery.
So Tang Jie posting a tweet in the early morning was not accidental. Zhipu's surge was not accidental either. When the capital market understood this narrative of "from tool to labor," the valuation logic completely changed.
Having talked about business models, technology trends, and the capital market, let's now talk about something closely related to us—interaction methods.
Today, at the Android Show, Google announced a new feature called "Magic Pointer."
Simply put: the mouse cursor has been given "eyes" and a "brain." It no longer only knows its X and Y coordinates on the screen, but knows "you are pointing at an expense field in a table" or "what you are pointing at is the amount number in a photo of an invoice."
Then you only need to say one sentence: "Add this to that report." AI can understand who you are pointing at, what you want to do, and how to do it.
The core of AI interaction is changing from "typing, typing, typing" to "pointing and clicking."
Why is this significant?
Over the past few years, the capabilities of large language models have surged, but the interaction method has barely changed—it is still typing input and waiting for a reply. It's like we are still using the command line to operate a modern operating system.
Think about it: when you collaborate with colleagues on a computer, how do you express yourself? You point at a page and say "there's a problem here"—rather than typing a 500-word text description. The most natural human interaction is always language combined with gestures.
Google's "Magic Pointer" is essentially doing one thing:releasing "AI understanding intent" from the narrow channel of text input into the vast space of physical interaction。
Some people may think this is just a small interaction improvement. But history tells us:every revolution in interaction methods has given rise to an entire new ecosystem。
The invention of the mouse made graphical interfaces possible. The spread of touchscreens made smartphones explode. So—when "point at the screen + say one sentence" becomes the new normal for AI interaction, what will it give rise to?
Perhaps very soon, the way we use AI will no longer be opening a chat window and typing a lot of text. Instead, we will open the screen, point and click, and talk. ChatGPT's chat box may really disappear.
Alright, today's four things have all been covered. Although they seem independent, behind them there is actually one underlying logic running through all of them.
Let me use an analogy to make it clear.
AI has three stages, just like a person's growth history.
Stage one: AI = calculator
—you give me input, I output according to rules. GPT-3 to GPT-4 were like this. Intelligent but passive.
Phase 2: AI = Intern
— You assign tasks, I execute independently. The beginning of the Agent era. Anthropic's Claude leads here.
Phase 3: AI = Expert
— You give a goal, I break it down, execute, correct errors, and deliver. This is what Tang Jie meant by long-horizon tasks.
Ultimate phase: AI = Mentor
— Not only can it do the work itself, it can also train the next generation of AI. Recursive's goal.
Today's four things correspond exactly to the leap across four phases:
· Anthropic overtakes OpenAI → marks the maturity of the "AI = Intern" business model
· Recursive founded → someone is betting that the "AI = Mentor" phase will arrive early
· Zhipu's stock price surges → the capital market recognizes the "AI = Expert" narrative
· Google's magic pointer → the interaction method is matching the evolution of AI capabilities
These four phases occur alternately and advance simultaneously. What does that mean?
It meanswe may be experiencing the most dramatic leap in the history of the AI industry。
After ChatGPT was born at the end of 2022, AI went through a two-year "carnival period"—everyone was excited that "AI can chat." From 2024 to early 2025, it entered a "cooling-off period"—people began asking, "Besides chatting, what else can it do?"
And today, in May 2026, what we see is the beginning of the "getting-things-done period"—enterprises are starting to pay for AI to actually work, entrepreneurs are starting to bet their entire fortunes on AI training itself, and interaction methods are starting to be redesigned for AI to "execute on behalf of humans."
The tool era is over. The labor era has arrived.
Putting today's four things together reminds me of a very classic concept—The Innovator's Dilemma。
Christensen said in The Innovator's Dilemma: the reason leading companies get disrupted is not that they did something wrong, but that they did everything right—only the standard for what counts as doing the right thing changed.
What did OpenAI do right? It created the world's first and most powerful general-purpose large model.
But why was the throne taken by Anthropic? Because the "standard for winning" changed. From "whose model is smarter" to "whose model can make enterprises truly treat AI as labor." And Anthropic's pay-as-you-go billing + Agent capabilities happened to satisfy the new standard.
Likewise, what Recursive is betting on is also that "the standard will change again"—from "can AI do the work" to "can AI evolve on its own."
Every change in the standard is a round of reshuffling.
So, in the end, what I want to say is:
In the AI industry, no throne is eternal.
The overlord you see today may be disrupted tomorrow by a 25-person team that has "no product but a top-tier brain."
The interaction method you find sufficient today may be replaced tomorrow by a "mouse that has grown eyes."
The only constant in the industry is change itself.
Finally, here's a question for you: Which phase of AI is your industry in? Is it a calculator, an intern, an expert, or already ready to welcome mentor-level AI?
Room 1409, Building C, Shenzhen Century Industrial Center, No. 743 Zhoushi Road, Hezhou Community, Hangcheng Sub-district, Bao'an District, Shenzhen
