Quick Take
- Summary: Former OpenAI researcher Andrej Karpathy has sketched out a widening “sharp funnel” of AI capability awareness. He puts the number of people who truly grasp today’s frontier at roughly 5,000 insiders, while billions sit essentially untouched by the technology.
- What happened: In a recent breakdown, Karpathy described the gap between public understanding and what labs can actually do as splitting into four rough tiers: about 6 billion people with no real exposure at all, 1–2 billion who use chatbots casually, 20 million professionals who lean on the tools for work, and those 5,000 lab insiders already watching AI agents coordinate on complex software and math tasks.
- Why it matters now: Market bets, enterprise plans, and regulatory moves are still being shaped by what most people see—rate-limited, safety-tuned chatbots. Yet the internal reality at the labs suggests autonomous agents could soon compress years of expert work into weeks. That mismatch leaves public expectations anchored far below where the technology actually sits.
- Who is most affected: Enterprise CTOs, AI policymakers, and knowledge workers who may be underestimating the productivity shift ahead, still treating today’s tools as the outer edge of what’s possible.
- The under-reported angle: Karpathy’s numbers are informal napkin math, but they point to a real structural problem. Democratic debate about AI risk and adoption is unfolding between groups living in entirely different technological realities.
Deep Dive
Have you ever noticed how conversations about AI often feel like they’re happening on different planets? The disconnect isn’t mainly about politics or values. It’s an exposure gap that’s grown too wide to ignore.
According to Andrej Karpathy, most of the public is still debating a limited preview version of the technology—treating large language models as slightly upgraded search engines. Meanwhile, inside frontier labs like OpenAI, Anthropic, and Google DeepMind, a small group of insiders is already observing agentic systems handling multi-step engineering, cybersecurity, and advanced math problems.
Karpathy frames this as a “sharp funnel.” At the wide end sit roughly 6 billion people who’ve never meaningfully used an LLM. Next come 1 to 2 billion casual users prompting chatbots for everyday text tasks. Then around 20 million professionals applying the tools to coding and analysis. At the narrow tip are perhaps 5,000 people with direct access to the raw, unfiltered frontier models. What they see isn’t just faster text output—it’s a genuine compression of timelines, where systems can reportedly finish what used to take expert teams years in a matter of weeks.
The exact counts are estimates, of course, but the signal matters. Public models are heavily constrained by safety rules, compute limits, and alignment work. What reaches most users is a watered-down, throttled version. As a result, media coverage and everyday assumptions tend to treat the current ChatGPT or Claude experience as the ceiling, rather than a deliberately restricted slice.
This gap carries real consequences for infrastructure and governance. If only a few thousand people have a clear view of how much compute, hardware, and autonomous behavior next-generation agents will demand, the broader market is probably misjudging future needs for data centers, GPUs, and power. The builds happening today are based on the capabilities visible only at the bottom of that funnel.
The same divide complicates policy. Lawmakers work from the tools they can access, while labs push systems that already move past those assumptions. When the internal agentic capabilities eventually reach the market, the economic and workforce effects could arrive with little warning for most of the world.
Stakeholders & Impact
- AI / LLM Providers — Impact: High. Insight: The gap between what internal agents can do and what gets released publicly will shape every launch decision going forward.
- Enterprise Strategy & CTOs — Impact: High. Insight: Many teams risk planning around today’s public APIs instead of preparing for the autonomous workflows that are already running in labs.
- Knowledge Workers — Impact: High. Insight: Underestimating how quickly these systems can handle multi-stage projects leaves people exposed when the timelines compress.
- Regulators & Policy — Impact: Significant. Insight: Rules are being written based on widespread public experience rather than the frontier capabilities labs are already scaling.
About the analysis
Researched and written with AI assistance under the editorial responsibility of the i10X News team. The title image is an AI-generated illustration. Read our AI Use Policy. Spotted an error? Request a correction.
This is an independent, research-based take drawing from market discussion, expert estimates, and coverage of the AI ecosystem. It aims to help developers, enterprise leaders, and policymakers separate everyday perception from the actual direction of frontier systems.
i10x Perspective
The widening capability gap shows that the most important stretch of the AI race remains invisible to nearly everyone. As models move from chat interfaces toward deeper reasoning agents, the real challenge won’t just be securing GPUs. It will be managing the jolt when those lab-only capabilities reach the wider market.
Over the next five years, the central task for labs like OpenAI and Anthropic will be bridging the distance between their small circle of insiders and the billions still unprepared for what follows. Oversight struggles when the people meant to govern can’t see the systems they’re trying to steer.
