Karpathy Warns of a Massive Divide in How People Understand AI
Andrej Karpathy says the gap in AI understanding is widening into a sharp funnel, leaving the public blind to vertical progress.
Models · Source: Hacker News
What happened
Andrej Karpathy points out a widening gap in how people perceive AI capabilities. He describes this divide as a sharp funnel rather than just two groups talking past each other. At the top of this funnel, roughly six billion people have zero contact with large language models. Another one to two billion are casual users who treat free-tier chatbots like a better search engine or a simple writing aid. These users, including many executives, only see derpy chatbots. Their understanding of AI remains second-hand and abstract.
The reality shifts drastically for professionals in math and code. About twenty million people now use frontier models to build, copy, translate, and decompile apps with simple prompts. Karpathy notes that less than a year ago, he was still typing code manually character by character. Now, agents handle complex projects that used to take weeks or months. This group sees the direct economic value of AI first-hand.
At the very bottom of the funnel are about five thousand insiders with internal access to frontier systems. These users watch swarms of thousands of agents collaborate over weeks on software mega-projects. They are minting zero-day exploits, running cyber attacks and defenses at machine speeds, and discovering new science. Human comprehension is already falling behind. People are still investigating the OpenAI-HF incident from months ago. Mathematicians are stuck poring over 722 manuscripts on frontier mathematics generated by these systems.
Key facts
- 6B — People who have barely come in contact with LLMs
- 1-2B — Casual users of free-tier ChatGPT-like products
- 20M — Professionals using frontier-grade LLMs in math and code
- 5,000 — People with internal access to frontier-grade systems
- 722 — Manuscripts on frontier mathematics that human mathematicians are reviewing
Why it matters
Most founders are building for an audience that still views AI as a basic chatbot. If you build products based on what the general public understands, you are aiming at a moving target that is already obsolete. The real capability of AI peaks in digital, verifiable, and economically valuable domains. This is because language model capability emerges from reinforcement learning on verifiable rewards. The system needs a curated environment driven by revenue potential. Small questions barely stress the system. You need a reservoir of massive, difficult problems to see the true power of these models.
The second-order effect is a massive crisis in human review and comprehension. Karpathy explicitly notes that human review is starting to fall behind machine output. When AI systems generate complex, high-value output at scale, humans become the immediate bottleneck. We see this with the mathematicians analyzing hundreds of frontier math manuscripts. Builders who solve this verification problem will win massive contracts. The gap between machine generation and human verification is the next great market opportunity.
For builders
Build for verifiable and digital domains
Language model capability peaks where rewards are perfectly verifiable and economically valuable. Focus your products entirely on math, code, and digital environments. Users in these specific fields will pay heavy premiums for agents that replace weeks of manual work.
Solve the human review bottleneck
Insiders are generating thousands of complex outputs like frontier math manuscripts and zero-day exploits. Human comprehension is rapidly falling behind machine speed. If you build tools that help human experts verify AI output faster, enterprise companies and research labs will buy them immediately.
Stop building for casual chatbot users
Casual users only see a modest uplift and treat AI like a basic writing aid. They will not pay premium prices for your products. Target the twenty million professionals who need a reservoir of big, difficult problems solved right now.
My take
I see founders pitching AI wrappers to casual users every single day. They are fighting for pennies at the top of Karpathy's funnel while ignoring the real shift. The real money is at the bottom. You must build verifiable agent swarms for the few million professionals who actually understand that the curve is going vertical. Stop building for the six billion who do not care.
Original reporting: Hacker News. This is my rewrite and opinion.