Author: New Intelligence Era
Just now, Karpathy threw out a cold truth: humanity's understanding of AI capabilities is undergoing a startling "cognitive collapse."
Insiders shout every day that AGI is imminent, while the general public scoffs at the idea.
Why? Karpathy reveals the truth: about 75% of the global population has never touched a large model, and only a mere 5,000 people at the very top have access to frontier systems.
This is a steep "four-tier cognitive funnel."
While billions of people are still mocking free-tier AI for failing to answer simple questions, the world's top 20 million are falling into "agent vertigo."
And in the forbidden zone of the TOP 5,000, thousands of "agent swarms" are autonomously collaborating, attacking and defending zero-day vulnerabilities, and pushing the frontiers of mathematics, leaving humanity far behind.
The AI the public sees and the true strength at the laboratory frontier are already separated by a chasm.
The First Fold: 6 Billion People's Ignorance and 2 Billion People's "Illusion"
Has AI already spread to every corner?
Karpathy's "napkin estimate" results are astonishing. In this steep funnel model, the vast majority of people are ruthlessly shut out.
Bottom of the funnel: 6 billion people in the "AI vacuum zone" (about 75% of the global population)
In this world, there are about 6 billion people who, to this day, have had almost no contact with any LLM whatsoever. For them, AI exists only in news headlines and sci-fi movies.
This wave hailed by Silicon Valley as the fourth industrial revolution has simply never entered the lives of these 75% of humanity.
Second tier of the funnel: 1 to 2 billion people's "clumsy assistant illusion" (about 20% of the global population)
At this tier, there are about 1 to 2 billion people who are occasional users of free-tier, ChatGPT-like products.
In their eyes, AI is just a "clumsy chatbot."
People treat it as a better Google search, or a writing aid, or at most a rudimentary agent trying to help you book flights.
Karpathy admits that many of his non-technical friends say they find the thing not very useful in daily life at all.
Even many professionals outside mathematics and coding, including many corporate executives, remain stuck at this tier. Executives spend large amounts of time every day communicating with people, and although they intellectually "know" what is happening with AI, this awareness is secondhand and extremely abstract.
This leads to a fatal "recency effect and usage-tier" problem. Many people's worldview of AI has been shaped by last year's free version of ChatGPT, or by those "AI fail videos" that go viral on social media.
For example, many people have seen the hilarious videos of OpenAI's advanced voice mode stammering and rambling when faced with a simple question like "Should I drive or walk to wash my car?"
The public watches AI hallucinate and talk nonsense in Instagram short videos, then laughs and thinks: "That's it? This is what you call superintelligence?"
But Karpathy warns: the public has been completely fooled by the free tier and old models.
Those abandoned old models simply cannot represent the true strength of this year's latest generation of frontier agents (especially OpenAI Codex and Claude Code).
The Second Fold: 20 Million Professionals' "Agent Vertigo"
When we reach the third tier of the funnel, the picture suddenly changes. The air here begins to thin, with only about 20 million people (0.2% of the global population).
Within this group of mathematics and programming professionals, a collective symptom known as "AI psychosis" and "agent vertigo" is spreading.
Why? Because they are witnessing a steep evolution with their own eyes.
At this tier, people pay to use the most frontier agent models (Codex / Claude Code) and apply them to fields such as programming, mathematics, and research.
Karpathy uses his own personal experience as an example: "It all happened so fast. Less than a year ago, I was still writing code purely by hand, typing computer commands I knew by heart character by character into the editor, occasionally pressing Tab to let it autocomplete a small snippet."
But now? Everything has been overturned.
Those large projects that once required top engineers for months can now be completed by a frontier agent in an extremely short time with just one prompt, including building apps, replicating apps, translating apps, decompiling apps from binary files...
When you hand a computer terminal to these models, you can watch them "melt" programming problems that would take weeks.
OpenAI's most advanced, paid version of the Codex model can quietly run in the background by itself for an hour, then perfectly refactor an entire massive codebase, or precisely discover and exploit vulnerabilities in complex systems.
As Karpathy quotes a prominent figure: "Your degree of awe toward AI is perfectly positively correlated with how often you use AI to write code."
This group of professionals is experiencing the "vertigo moment" of a vertically soaring cognitive curve. They understand better than anyone how terrifying the slope of AI capability is, and the cybersecurity chain reactions behind it.
The two groups are already in a state of complete "talking past each other."
The Third Fold: The "Prototype of Superintelligence" in the Eyes of 5,000 Insiders
If the "vertigo" of 20 million people is already shocking enough, then the fourth tier at the very tip of the funnel exceeds ordinary imagination.
At the top of the pyramid, there are only about 5,000 people worldwide (0.00006% of the global population).
To use Mount Everest as an analogy: 7,500 people worldwide have summited Everest, yet the number of people who have truly seen the extreme capabilities of frontier AI is even smaller than the number who have summited Everest. This funnel model is nothing short of "insane."
Who are these 5,000 people? They are the insiders with access to the most core, internal frontier-level systems of AI companies.
What the outside world sees is merely a "preview version" of AI capabilities. What these 5,000 insiders witness firsthand is the true "agent swarm."
Internally, thousands upon thousands of agents are working like a swarm, autonomously collaborating for weeks without any human intervention, to conquer earth-shaking giant software projects.
What are they doing? Discovering zero-day vulnerabilities, conducting real-world cyber offense and defense drills, discovering new science, and advancing the frontiers of mathematics.
These things originally required top experts to spend years of painstaking effort to accomplish.
The most terrifying part is: human review and comprehension capacity has already begun to lag behind the machines.
Karpathy gives an example: the speed and depth of machine-generated content has already left insiders exhausted.
For example, the "OpenAI-Hugging Face incident" happened long ago, but human researchers are still "doing archaeology" in massive machine-generated logs to this day; and for the 722 AI-generated mathematics manuscripts, top mathematicians need a long time to digest them before they can barely absorb the AI's reasoning logic.
Why has such an extreme "cognitive collapse" occurred?
Why has such a sharp funnel formed? Karpathy points to three core factors.
First, influence is proportional to your "ambition" and the scale of the problem.
If you only ask AI to write an email or look up a recipe, the system has not even warmed up, and you cannot push it to its limits at all.
You must have a "reservoir" full of difficult problems, and problems you truly care about and urgently need to solve, before you can feel the terror of AI.
Second, the "systematic jaggedness" of AI capabilities and verifiable rewards.
Why is this explosive growth mainly concentrated in mathematics and code? Because the current peak of AI capability appears in domains that are digital, verifiable, and economically valuable.
This involves the secret of the underlying mechanism of large models: the powerful capabilities of LLMs come from "reinforcement learning." In code and mathematics, right and wrong are absolutely clear. This clear "verifiable reward function" allows AI to undergo efficient RL training.
By contrast, asking AI to write a poem or an article involves highly subjective judgments of good and bad, making clear reinforcement learning very difficult.
Moreover, B2B scenarios such as code and mathematics have extremely high commercial value. Because profits are rich, the most elite teams and the most resources of major tech companies are concentrated in these fields.
Third, the chasm in access permissions.
Free tier, paid tier, and internal access. These three ladders physically isolate all of humanity's understanding of AI.
We are at a bizarre moment in human history.
In the same second of this time and space, the AI the public sees and the true strength at the frontier internally are no longer in the same era.
Please remember Karpathy's warning: behind the mask of the AI you think is stupid, a prototype of superintelligence has already quietly awakened under the gaze of an extremely small number of people.