Andrej Karpathy, former Tesla AI lead and OpenAI researcher, has reframed the AI discourse as a problem of experiential stratification. His model is a funnel: roughly 6 billion people (75% of the global population) have barely touched LLMs at all. Another 1-2 billion use free-tier chatbots casually — better Google, writing assistant, novelty toy. The capability they encounter is real but modest, and many professionals outside math and code reasonably report that the tools don't change their work much. The funnel narrows fast. Approximately 20 million people — 0.2% of humanity — are professional users of frontier-grade LLMs in math and code. Karpathy's description of their experience is stark: large, complex projects that previously required weeks or months of skilled work are now completed by agents from a prompt. He notes that less than a year ago he was manually typing memorized code commands character by character. That era is over for this cohort. At the very tip: roughly 5,000 people with internal access to frontier systems. What they see, per Karpathy, includes swarms of thousands of agents collaborating over weeks on software mega-projects, minting zero-day exploits, running cyber operations at machine speed, discovering new science, and advancing frontier mathematics. Human review and comprehension are beginning to fall behind — he cites ongoing 'archeology' of the OpenAI-HF incident and mathematicians still working through 722 manuscripts of AI-generated frontier math. Three structural factors drive the funnel's shape. First, impact scales with problem ambition and horizon — a paragraph-length question barely stresses the system; you need big, difficult problems you care about. Second, capability is jagged: it peaks in digital, verifiable, economically valuable domains because LLM capability emerges from reinforcement learning on verifiable rewards in curated environments driven by revenue potential. Third, access itself is tiered — free, paid, internal — and each tier sees a fundamentally different system. The GRIN lens reveals something uncomfortable here. This is a description of an information asymmetry so extreme it resembles a new class structure forming in real time. The 5,000 insiders — overwhelmingly employees of a handful of frontier labs — possess operational knowledge about capabilities that the other 7.99 billion people literally cannot access or verify. Policy, regulation, labor markets, education systems, and public discourse are all operating on information from the bottom of the funnel while decisions that affect everyone are being made at the top. Karpathy frames this as 'the weirdness of the moment,' but the structural read is harsher. When a technology's most consequential capabilities are visible only to the people who build and sell it, the information asymmetry IS the product moat. The funnel doesn't just describe unequal access — it describes unequal capacity to make informed decisions about AI governance, career planning, investment, and institutional adaptation. The general public is being asked to form opinions about a technology most of them have interacted with in its weakest form. The twenty-year question is whether this funnel widens or calcifies. If frontier capabilities diffuse broadly and quickly, the asymmetry is temporary and the current moment is just the awkward leading edge of a genuine capability revolution. If access remains stratified — as it does today, and as the economic incentives of frontier labs encourage — then a permanent epistemic class divide is forming, where a few thousand people understand what is actually happening and everyone else is navigating by rumor, hype, and second-hand accounts.
