Karpathy: ~5,000 people see frontier AI. 75% of humanity has not met it

On October 11, Andrej Karpathy posted that the gap in shared understanding of LLM capability is widening, and that it now looks less like two camps talking past each other and more like a sharp funnel. By his own napkin math, about 6B people (~75% of the population) have barely touched an LLM. Around 1-2B (~20%) use free-tier chatbots casually. Roughly 20M (0.2%) use frontier-grade models professionally in math and code. And about 5,000 people (~0.00006%) have internal access to frontier-grade systems. These are his estimates, not measured data.

The four tiers, and what each one sees

Karpathy's claim is that each tier forms its view from a different machine. The casual tier sees what he calls derpy chatbots, used like a better search engine or writing aid. Many professionals outside math and code sit here too; he notes that executives spend much of their time talking to people, so what they know about AI is second-hand and abstract.

The 20M professional tier is where he says the curve has gone vertical. He describes projects that once took weeks or months being completed by agents from a prompt: building, copying, translating and decompiling apps. He says that less than a year ago he was still typing code by hand. The top tier, in his account, is visible to outsiders only as a preview: swarms of thousands of agents collaborating for weeks on software mega projects, finding zero days (previously unknown security flaws), running cyber attacks and defenses at machine speed, and advancing mathematics. He adds that human review is starting to fall behind, citing ongoing "archeology" of the OpenAI-HF incident and the 722 manuscripts on frontier mathematics that mathematicians may be reading for a while.

Why the funnel has this shape

Karpathy names three drivers. First, impact scales with ambition: a question with a paragraph-long answer barely stresses the system, so you need a reservoir of big, hard problems you care about to see what it can do. Second, jaggedness: capability is uneven, peaking in domains that are digital, verifiable and economically valuable. His explanation is that capability emerges from reinforcement learning on verifiable rewards, meaning training where the model is scored automatically on whether its answer was right (code that passes tests, a proof that checks), over a mix of training environments chosen for revenue potential. Where answers can't be checked automatically, progress is slower. Third, access: free tier, paid tier, internal.

The Playbook: what to do in the next month

The practical risk is deciding from the wrong tier. If your team's opinion of AI comes from a free chatbot tried long ago, it is probably describing a different product than the one Karpathy says professionals now use. If you work in code or math-like, checkable tasks, his argument implies the uplift may be far larger than elsewhere. If your work is mostly conversation and judgment, a modest gain may be the honest answer, and he calls that view reasonable.

Sensible moves: have someone on your team run a real, large, multi-day task on a current paid frontier tool and report back. Pick tasks with automatic checks. Do not extrapolate from one tier to another. The mistake is the reverse of each: dismissing the technology from a chatbot demo, or assuming top-tier agent swarms describe what you can buy today.

Questions You Should Be Asking

• When did the people forming our AI view last use a paid, frontier-grade tool, and on what size of problem?
• Is our work digital and verifiable enough to sit where capability peaks, or are we assuming coding-style gains apply to it?
• If a vendor cites results from systems with internal-only access, what can we actually buy and test today?
• If agents produce work faster than we can review it, who is accountable for what ships?
• Are our competitors in the professional tier while we evaluate from the casual one?

What To Watch Next

Watch whether the top-tier work Karpathy describes becomes independently checkable: published review of those 722 manuscripts, or findings from the OpenAI-HF incident, will show whether human comprehension is catching up or falling further behind.