Andrej Karpathy says only 5,000 people have seen what AI can really do

OpenAI founding member Andrej Karpathy, who joined Anthropic in May, thinks most of the world has little idea of what AI can do, and in a Sunday post on X, he said the gap is getting wider. By his own napkin math, roughly 6 billion people have barely used an LLM, 1 to 2 billion use chatbots casually, about 20 million use frontier models professionally in math and code, and around 5,000 have access to frontier systems inside the labs.

Box CEO Aaron Levie sees an opportunity in that gap. “This is a good way to understand how early we actually are in AI diffusion right now,” he wrote in his response on X.

He argued that most people still think of AI as a chatbot or a next-generation search engine while barely exploring its more advanced opportunities.

Box bets on agent swarms

“Agent swarms represent possibly the ultimate form factor of what agents will do inside of our workflows,” Levie wrote, envisioning systems that divide complex work among agents running in parallel.

“Agent swarms represent possibly the ultimate form factor of what agents will do inside of our workflows”

For now, “less than a fraction of a fraction of the population uses AI in this way,” Levie noted, and he identified the obstacles keeping it that way. We’re still early. Models aren’t equally capable across every kind of work, running them at scale is expensive, and tools to manage multiple agents are still catching up. Most people don’t even have access to the most powerful models yet.

Rethinking how work gets done with AI is a fraction of what companies have to consider. Connecting their data and figuring out security and compliance are just a few other challenges. Levie expects “a huge amount of change management” before most industries outside software and AI research can take advantage of these agents.

That puts Box squarely in the conversation, since the company has been expanding its AI offerings with products such as Box Automate, which connects AI-powered automation to document-heavy business processes. Levie has also said Box doesn’t intend to compete directly with model providers.

“A lot of this is where the applied layer of AI will really shine in the enterprise because of the domain specificity of this type of work,” Levie wrote, placing the opportunity with the companies building industry-specific applications and infrastructure on top of frontier models.

Chatbots versus coding agents

Karpathy sees casual users and developers as looking at opposite ends of the same technology. In his original April post, he contrasted OpenAI’s free voice mode, which he said “will fumble the dumbest questions in your Instagram’s reels,” with the company’s top-tier Codex model, which can spend an hour restructuring an entire codebase.

Karpathy wrote that less than a year ago he was still writing code by hand, occasionally relying on autocomplete. By April, he was describing agents that “melt programming problems that you’d normally expect to take days/weeks of work.”

Back in February, Karpathy moved beyond vibe coding to “agentic engineering,” with developers spending more time directing and reviewing agents than writing code.

Why verifiable work wins

“Capability peaks in domains that are digital, verifiable and economically valuable,” Karpathy wrote. Coding and math are well suited to reinforcement learning because results can be tested and improvements measured. Writing and advice are harder to evaluate, and there’s more money to be made improving models for technical work. As Karpathy put it in April, “The goldmines are elsewhere, and the focus comes along.”

“Capability peaks in domains that are digital, verifiable and economically valuable.”

Karpathy also thinks people aren’t giving these systems enough to do. “A question with a paragraph answer barely stresses the system,” he wrote. The real gains become apparent when agents can spend hours working through a problem, testing changes and trying again.

In March, Karpathy’s AutoResearch project ran 50 experiments overnight on a single GPU, found a better learning rate and committed the change to Git. Each run produced a validation score, so the agent could tell whether a change was worth keeping and move on to the next experiment.

From capability to deployment

Karpathy calls it “the weirdness of the moment.” Most people still judge AI by their experiences with chatbots, while a small group of developers watches capabilities improve at a staggering pace. Levie sees the same gap, but from the enterprise side. Companies still have to figure out how to make these systems affordable, secure, and reliable enough to use at scale. For Box, that’s where the opportunity lies.

Most people still judge AI by their experiences with chatbots, while a small group of developers watches capabilities improve at a staggering pace.