When ChatGPT burst onto the scene in 2022, AI quickly became what appeared to be an accessible technology with the potential to benefit everyone. In China, for example, the Doubao app reportedly has 380 million monthly active users, while an estimated 700 million people across the country have used AI.
However, Andrej Karpathy, a former OpenAI researcher, recently shared a rather different perspective on X (formerly Twitter):As AI capabilities advance at an unprecedented pace, the gap between people who know how to use AI effectively and those who don’t is growing rapidly. In the future, only a small fraction of people may truly understand AI’s potential and know how to harness it effectively.
Note: Andrej Karpathy was a founding member of OpenAI and previously served as Director of AI and Autopilot Vision at Tesla. He is currently at Anthropic, where he leads Claude’s pretraining research team. He is also credited with coining the term “vibe coding.”
Below is Andrej Karpathy’s original post:
I’m resurfacing this post from April because the gap in people’s understanding of LLM capabilities is growing wider.
The situation is increasingly less like two groups talking past each other and more like a distinct funnel.
Here’s a rough breakdown:
- Around 6 billion people (approximately 75% of the global population) have had virtually no exposure to LLMs.
- Another 1–2 billion people (approximately 20% of the population) use free, ChatGPT-like products occasionally.
What these users encounter are relatively clunky chatbots, which they treat as slightly better versions of Google Search, writing assistants, and similar tools. Perhaps they might even use one to help book a flight.
Some of my non-technical friends, whose perspectives I find entirely reasonable, tell me they simply have very little use for LLMs.
This group also includes many professionals outside mathematics and programming. Executives and people in other business functions, for example, spend much of their time communicating with others. While they may understand LLMs in theory, their understanding remains largely indirect and abstract.
Now, let’s move to the frontier of professional applications in mathematics and software engineering.
Around 20 million people (approximately 0.2% of the global population) have witnessed firsthand how large, complex projects that once took weeks or even months can now be completed by AI agents with a single prompt.
Building applications, replicating them, translating them, reverse-engineering them from binaries…
All of this has happened incredibly quickly, and not that long ago. Less than a year ago, I was still manually writing code, typing out computer instructions from memory, character by character, occasionally pressing Tab to autocomplete a small snippet.
Finally, there’s another group:approximately 5,000 people (just 0.00006% of the global population) who have internal access to cutting-edge systems.
The outside world has only seen glimpses of what’s possible.
Imagine thousands of agents working together over several weeks to develop large-scale software projects, discover zero-day vulnerabilities, execute cyberattacks and defenses at machine speed, make new scientific discoveries, and push the boundaries of mathematics.
These are tasks that would previously have required the world’s leading experts years to accomplish.
Meanwhile, the speed of human review and comprehension is beginning to fall behind.
People are still piecing together the “archaeological record” of the OpenAI-HF incident from several months ago. Mathematicians may need considerable time to carefully examine the 722 manuscripts on frontier mathematics.
This funnel is driven by several factors:
1. Impact Is Proportional to the Scale of the Problem
The impact of AI is proportional to the goals, scale of the problem, and time horizon involved.
A question that can be answered in a single paragraph places almost no real demands on the system.
You need large, genuinely important, and extraordinarily difficult problems that you deeply care about.
This factor primarily drives the consumer-versus-professional dimension of the funnel.
2. The Complexity of the Environment
I’ve written separately about this in greater detail.
AI systems perform best in domains that are digital, verifiable, and economically valuable.
This is because LLM capabilities are increasingly driven by reinforcement learning with verifiable rewards, trained across carefully curated mixtures of environments where the potential for measurable returns provides a strong incentive.
This factor primarily drives the domain dimension of the funnel, particularly in areas such as mathematics and software engineering.
3. Access to Cutting-Edge Systems
Free-tier access. Paid-tier access. Internal access.
This is where things get particularly strange.
Most people have never interacted with these systems. Even when they do, the experience often feels like talking to a clunky chatbot.
Most professionals experience only modest productivity improvements.
Meanwhile, a tiny fraction of people are already experiencing the dizzying acceleration that comes with an almost vertical capability curve.
All of this is happening simultaneously.