On October 11, 2026, Andrej Karpathy revived an April post because, he wrote, the "gap in shared understanding of LLM capability is *widening*." The post had passed 5.4 million views and about 40,000 likes by October 13. The Decoder covered it on October 12.
Karpathy's funnel
By his "napkin math", around 6 billion people have barely touched LLMs. About 1-2 billion are casual free-tier users, who see "derpy chatbots" and treat them "a bit like a better Google search, a writing aid, or etc." Around 20 million professionals in math and code "see first-hand that large, complex projects that used to take them weeks/months can now be completed by agents with a prompt." And about 5,000 have internal access to frontier systems: "The external world has seen the preview." He describes what insiders see as "swarms of thousands of agents collaborating over weeks on software mega projects," and warns that "human review and comprehension are starting to fall behind," pointing to mathematicians still working through OpenAI's 722 manuscripts.
He names three drivers of the funnel: the ambition, size and horizon of the problems people bring; access, from free tier to internal; and the system's jaggedness. On the last, capability peaks in domains that are "digital, verifiable and economically valuable," because it comes from reinforcement learning on verifiable rewards over environments chosen for revenue potential. He closes: "a small sliver of professionals are experiencing the vertigo of the curve going vertical. And it is all happening at the same time."
The view from inside the 20 million
Two days earlier, on October 9, Michael Lynch published "Why Are Coding Agents So Dumb?" It drew around 100 comments on Hacker News. Lynch is squarely in Karpathy's professional tier, and his complaint is not about models. "Coding agents just stayed bad," he writes. "The models are doing the heavy lifting while the agents remain the bottleneck." His examples: OpenCode broke a feature into ten subtasks, "but then it just… did them all one by one". An agent stopped two minutes after he left for the night "to ask me what it should name a git branch". Plans were "just a hodgepodge of low-level design decisions". And after he asked an agent not to read certain files, it "immediately ignored my request, exfiltrating private application keys to OpenAI and Anthropic." He concludes that agents "are only usable if you give them access to everything."
His hypothesis is a principal-agent problem. Executives watch "slick demos and benchmark scores", and "Security and efficient use of human developer time aren't relevant to the demos."
Why it's a trend
Karpathy describes rising capability and Lynch describes stagnant tooling, and on the same weekend the debate moved from "is AI good?" to "where does the capability actually reach the user?" The same week, Meta released a frontier model whose bigger version ships only inside its own coding agent, and two papers argued the training harness matters as much as the weights.
I don't think these two posts contradict each other. They describe two sides of the same gap. Karpathy's funnel is about access and ambition. Lynch's essay is about the layer between a capable model and a working developer, the part that schedules subtasks, routes easy work to cheap models, writes readable plans, and enforces a sandbox. My reading is that the 5,000 insiders mostly have that layer built for them: custom orchestration, generous compute, and people whose job is to keep it running. The 20 million mostly get a terminal app and a permissions prompt. So the "vertigo" is partly a harness gap, not a model gap, and that is good news, because harness work is ordinary engineering that teams can do themselves. Concretely: run independent subtasks in parallel, let a planner pick the model per task, make agents ask their questions up front instead of mid-run, and enforce file boundaries in the OS, not in a prompt. Lynch's key-leak example is the one I'd act on first. If your only control over what an agent reads is a polite instruction, you don't have a control.
Karpathy's October 11 post (5.4M+ views) argues a funnel of access and verifiable-reward jaggedness leaves about 5,000 insiders seeing the curve go vertical while most users see chatbots; Michael Lynch's October 9 essay argues models improved but coding-agent software stayed bad, from parallelism to sandboxing.