Andrej Karpathy, an OpenAI founding member who joined Anthropic in May, believes the world has little understanding of what AI can accomplish, and according to a Sunday post on X, that divide is widening. Using rough estimates, Karpathy calculates that approximately 6 billion people have scarcely touched a large language model, between 1 and 2 billion interact with chatbots casually, roughly 20 million employ frontier models professionally for mathematics and coding work, and only around 5,000 have access to frontier systems within research labs. Box CEO Aaron Levie responded that this divide reveals how early the industry remains in AI adoption.
Karpathy highlighted the stark contrast between what casual users experience versus what developers can achieve with the same underlying technology. In an earlier April post, he compared 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, capable of spending an hour restructuring an entire codebase. Less than a year before April, Karpathy was still writing code manually with occasional autocomplete assistance, but by spring he described agents that "melt programming problems that you'd normally expect to take days/weeks of work." By February, he had transitioned to what he calls "agentic engineering," where developers spend more time directing and reviewing agents than actually writing code themselves.
Karpathy wrote that "capability peaks in domains that are digital, verifiable and economically valuable," noting that coding and mathematics suit reinforcement learning well because outcomes can be tested and progress measured. Writing and advice prove harder to evaluate, he explained, and greater financial incentives exist for improving models in technical domains. Levie argued that most people still perceive AI as either a chatbot or an upgraded search engine while barely investigating more sophisticated applications. He pointed to agent swarms as "possibly the ultimate form factor of what agents will do inside of our workflows," picturing systems that distribute complicated tasks among agents operating simultaneously. Right now, "less than a fraction of a fraction of the population uses AI in this way," Levie observed, identifying several barriers including uneven model performance across different work types, expensive large-scale operation, and tools for managing multiple agents that haven't caught up yet.
Levie sees the capability gap as a business opportunity for companies building industry-specific applications and infrastructure atop frontier models, rather than competing with model providers directly. He expects "a huge amount of change management" before most industries beyond software and AI research can capitalize on these agents, since companies must still solve challenges around affordability, security, and reliability at scale. Box has expanded its AI products with offerings like Box Automate, which connects AI-powered automation to document-intensive business processes. Karpathy also suggested users aren't demanding enough from these systems, writing that "a question with a paragraph answer barely stresses the system," and that genuine advantages emerge when agents spend hours working through problems, testing modifications, and iterating. In March, his AutoResearch project ran 50 experiments overnight on one GPU, discovered an improved learning rate, and committed the update to Git, with each run generating a validation score so the agent could judge whether a change merited keeping. Karpathy calls it "the weirdness of the moment"—most people judge AI by chatbot experiences while a small developer community watches capabilities advance at remarkable speed, a gap Levie believes positions Box and similar enterprise platforms to bridge the distance between frontier models and practical deployment. The divergence between perception and capability suggests that enterprise adoption will hinge less on technical breakthroughs than on whether platform providers can translate advanced agent architectures into workflows that non-technical organizations trust enough to deploy.