# LLM Capability Funnel By greenlig Topic: Society & Business [Open the interactive Jig](https://ask.jig.com/j/rIHLdRnuwzUn) ## Guide ### 1. A funnel, not a divide Andrej Karpathy argues that the gap in how people understand LLM capability is widening into a sharp funnel\. Each glowing ring is a population, with its radius on a logarithmic scale: about 6 billion people at the mouth, about 5,000 at the core\. All four tiers exist at the same time\. Tier sizes are Karpathy's napkin estimates, not survey data\. ### 2. About 6 billion: barely contacted The wide, dim mouth holds roughly 6 billion people, which the post puts at about 75% of humanity, who have had little or no contact with LLMs\. Against a mid\-2026 world population of 8\.3 billion, 6 billion is closer to 72%\. The ITU still counts about 2\.2 billion people who are not online at all\. Sources: the post; UN World Population Prospects; ITU Facts and Figures 2025\. ### 3. 1 to 2 billion: the chatbot tier About 20% use free\-tier chatbots casually: a better search engine, a writing aid, perhaps an agent that tries to book a flight\. The tasks are short, so a paragraph answer barely stresses the model and it can look derpy\. Many professionals outside math and code, executives included, sit here: they grasp the change intellectually, but second\-hand\. OpenAI reported about 900 million weekly ChatGPT users in early 2026\. ### 4. About 20 million: projects from a prompt Roughly 0\.2% use frontier\-grade models professionally in math and code\. They see agents complete projects that used to take weeks or months from a prompt: building, copying and translating apps, even decompiling them from binaries\. Less than a year earlier, Karpathy still typed code by hand, with occasional Tab autocomplete\. JetBrains estimates about 20\.8 million professional developers worldwide\. ### 5. About 5,000: the internal frontier At the core, about 0\.00006% of people use frontier systems internally\. Karpathy describes swarms of thousands of agents collaborating for weeks on software mega\-projects: zero\-days, cyber attack and defence at machine speed, new science and frontier mathematics\. Human review is falling behind; mathematicians may spend a long time on 722 frontier\-maths manuscripts\. These are Karpathy's descriptions; they are not independently verified here\. ### 6. Why the radius is logarithmic The slider switches from radius proportional to log10 of the head count to area proportional to people\. On a true\-area scale the 20 million ring shrinks to a 3\.6 cm radius and the 5,000 core to 0\.6 mm, about one 1,200,000th of the mouth's area\. Neighbouring tiers differ by roughly 4:1, 75:1 and 4,000:1, so only a log scale shows them all\. ### 7. Driver 1: ambition and horizon Impact scales with ambition, problem size and horizon\. The probe on the task\-horizon rail runs from a 10\-second question to years of expert work and sinks to the tier that routinely delegates work of that length\. Uplift stays invisible unless you hold a reservoir of big, difficult problems you really care about\. The mapping from horizon to tier is illustrative, following the post's argument\. ### 8. Driver 2: jagged capability Capability is jagged: it peaks where work is digital, verifiable and economically valuable\. Reinforcement learning from verifiable rewards trains on environments with automatic checkers, such as unit tests or exact answers, in a mixture weighted by revenue potential\. Code and math tower over fields judged by people, so the funnel narrows by domain as well as by person\. Pillar heights are schematic, not measured scores\. ### 9. Driver 3: access, all at once Access is the third filter: free tier, paid frontier tier and internal systems each light a different part of the funnel\. The public mostly meets a derpy chatbot and most professionals a modest uplift\. A thin sliver feels the vertigo of the curve going vertical, and it is all happening at the same time\.