The disparity between what AI can accomplish and what most people believe it can do continues to widen, according to Andrej Karpathy, an OpenAI founding member who joined Anthropic in May. In a post on X, Karpathy broke down global AI exposure using rough estimates: approximately 6 billion people have scarcely interacted with large language models, between 1 and 2 billion use chatbots casually, roughly 20 million leverage frontier models professionally for mathematics and coding tasks, and around 5,000 individuals have direct access to frontier systems within research laboratories.
Box CEO Aaron Levie responded to Karpathy's analysis, framing the observation as revealing how nascent AI adoption truly remains. "This is a good way to understand how early we actually are in AI diffusion right now," he wrote on X. Levie contended that widespread perception still confines AI to chatbots or advanced search tools, overlooking substantially more sophisticated applications.
Box bets on agent swarms
Levie identified agent swarms as a forthcoming frontier, describing them as systems capable of distributing intricate tasks across multiple agents executing simultaneously. "Agent swarms represent possibly the ultimate form factor of what agents will do inside of our workflows," he stated. Currently, adoption remains negligible, with Levie noting that "less than a fraction of a fraction of the population uses AI in this way."
Multiple obstacles impede broader deployment. Model capabilities remain inconsistent across different domains, operational costs at scale remain prohibitive, and infrastructure for coordinating multiple agents continues maturing. Additionally, most organizations lack access to the most capable models available. Beyond technical constraints, enterprises must address data integration, security protocols and regulatory compliance. Levie anticipates "a huge amount of change management" will precede widespread adoption outside software development and AI research sectors.
Box has positioned itself within this emerging landscape through products like Box Automate, which integrates AI-driven automation with document-centric business workflows. Levie has clarified that Box pursues a different strategy than model developers. "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," he wrote, suggesting the real opportunity belongs to companies constructing industry-specific solutions atop frontier models.
Chatbots versus coding agents
Karpathy observes that casual users and professional developers inhabit opposite ends of the same technological spectrum. He previously contrasted OpenAI's free voice mode, which he said "will fumble the dumbest questions in your Instagram's reels," with the company's advanced Codex model, capable of spending an hour restructuring an entire codebase. Less than a year before his analysis, Karpathy was still writing code manually with occasional autocomplete assistance. By April, he described agents that "melt programming problems that you'd normally expect to take days/weeks of work."
In February, Karpathy transitioned from what he termed "vibe coding" to "agentic engineering," a paradigm where developers dedicate more effort to directing and validating agent work than to writing code themselves.
Why verifiable work wins
Karpathy identified a pattern in where AI excels: "Capability peaks in domains that are digital, verifiable and economically valuable." Coding and mathematics align naturally with reinforcement learning because outcomes can be tested and improvements quantified objectively. Writing and advisory services present evaluation challenges, and greater financial incentives exist for advancing technical capabilities. As Karpathy explained, "The goldmines are elsewhere, and the focus comes along."
Karpathy also contends that current applications underutilize these systems. "A question with a paragraph answer barely stresses the system," he wrote. Substantial value emerges when agents operate over extended periods, iterating through problems, validating modifications and refining approaches. His AutoResearch project demonstrated this potential by executing 50 experiments overnight on a single GPU, identifying an improved learning rate and committing the modification to Git. Each experiment produced measurable validation scores, enabling the agent to assess whether changes warranted retention.
From capability to deployment
Karpathy characterizes the current moment as "the weirdness of the moment." Most people evaluate AI through their chatbot interactions, while a smaller cohort of developers witnesses capabilities advancing dramatically. Levie identifies the same divide from an enterprise perspective: organizations must still determine how to deploy these systems affordably, securely and reliably at scale. For Box, this deployment gap represents the central opportunity.