Most people have not seen AI’s real capabilities. Andrej Karpathy, an AI researcher and OpenAI cofounder, says only 5,000 people have.
That estimate points to a gap between what these systems can do and what the public has experienced firsthand. The general public has mostly not interacted with these systems, according to Karpathy, so its understanding remains shaped by limited exposure rather than the full range of LLM capability.
When people do encounter AI, the experience often fails to reveal much. “When they have, it looks like a derpy chatbot,” Karpathy said. The description is blunt, but it captures the problem: a narrow interface can make a powerful system look ordinary, unreliable, or merely amusing.
A chatbot window presents one version of AI interaction. Karpathy’s point is that this visible version does not represent everything these systems can do. If the public sees only a derpy chatbot, then the public view of AI capability stays far behind the systems themselves.
That distance is not holding steady. Karpathy revisited a post from April with a warning: “Dusting off this tweet from April because this gap in shared understanding of LLM capability is widening.” The emphasis matters because he is describing a growing divide, not a temporary misunderstanding.
The phrase “shared understanding” puts the issue beyond personal curiosity. A shared understanding requires people to have some common view of what LLMs can do; Karpathy says that common view is weakening as the systems’ capabilities move beyond what most people have seen.
The number at the center of his argument is strikingly small: 5,000 people. Against the general public, that group represents a limited circle of people who have seen what AI can really do, leaving most others with the chatbot version as their main point of reference.
That creates an unusual communication problem. AI systems may be discussed in broad terms, but discussion cannot close a gap that direct interaction has barely opened. The public has mostly not interacted with the systems, and the interactions it has seen often look like a “derpy chatbot.”
Karpathy’s comments, revisited after just over six months, frame the issue as a widening mismatch between access and understanding. The systems exist, but the experience needed to understand their capability remains concentrated among 5,000 people.
His warning does not claim that the public has formed a false picture from nothing. It says the public has seen too little—and what it has seen presents AI in a form that undersells the underlying capability. That is how a shared understanding gets thinner while the gap gets wider.
The result is a public conversation built around an incomplete view of LLMs. Until more people interact with these systems beyond the derpy-chatbot experience, Karpathy’s central point remains difficult to escape: only 5,000 people have seen what AI can really do.
Based on
