Karpathy: LLM Capability Gap Is a Sharp Funnel
Original: The gap in shared understanding of LLM capability is widening
Why This Matters
The perception gap shapes policy, adoption, and public debate on AI risk.
Andrej Karpathy posted on X that the gap in understanding LLM capability is widening into a "sharp funnel": 6B people untouched, 1-2B casual users, 20M power users, and ~5,000 with access to internal frontier systems.
Karpathy broke down global LLM exposure into four tiers. Roughly 6 billion people (~75%) have had almost no contact with LLMs. About 1-2 billion (~20%) are casual free-tier users who see these tools as a smarter search or writing aid. A much smaller group — around 20 million (0.2%) — are professional users in math and code who are watching agents compress weeks of complex project work into a single prompt. Less than a year ago, Karpathy writes, he was typing code character by character. At the extreme tip: approximately 5,000 people (0.00006%) have internal access to frontier-grade systems, where he describes "swarms of thousands of agents" collaborating on software mega-projects, running offensive and defensive cyber operations at machine speed, and advancing mathematics — producing outputs that exceed human review capacity. He points to OpenAI's HF incident and 722 frontier math manuscripts as examples where human comprehension is already lagging. Karpathy identifies three drivers of the funnel: impact scales with problem ambition; capability peaks in digital, verifiable, economically valuable domains due to RL on verifiable rewards; and raw access tiers. His core observation — the same technology looks like a "derpy chatbot" to most people while a tiny sliver experiences "the vertigo of the curve going vertical."