The AI Capability Chasm: Why Most of the World Doesn't Know What's Possible

The disparity between those who understand artificial intelligence's actual capabilities and the general population is expanding, 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, 1 to 2 billion engage with chatbots casually, roughly 20 million employ frontier models professionally for mathematical and coding tasks, and around 5,000 individuals have direct access to cutting-edge systems within research laboratories.

Box CEO Aaron Levie identified this gap as evidence of how nascent AI adoption remains. "This is a good way to understand how early we actually are in AI diffusion right now," he wrote in response on X. Levie contended that most people continue viewing AI through the lens of chatbots or advanced search technology, overlooking the more sophisticated applications emerging in the field.

Box bets on agent swarms

Levie envisions agent swarms as the next frontier, describing them as systems capable of distributing intricate tasks across multiple agents operating simultaneously. "Agent swarms represent possibly the ultimate form factor of what agents will do inside of our workflows," Levie stated.

Currently, adoption of AI in this manner remains negligible. Levie identified several barriers to widespread deployment: models lack uniform capability across different domains, operating them at enterprise scale carries substantial costs, and infrastructure for managing multiple agents remains underdeveloped. Additionally, most organizations lack access to the most advanced models available.

Beyond technical hurdles, enterprises face additional complexities. Data integration, security protocols, and compliance requirements demand attention. Levie anticipates "a huge amount of change management" will be necessary before industries outside software and AI research can effectively leverage these agent systems.

Box has positioned itself within this emerging landscape through products like Box Automate, which integrates AI-driven automation into document-centric business workflows. Levie has clarified that Box does not seek direct competition with 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," Levie wrote, suggesting the real opportunity belongs to organizations constructing industry-specific applications and infrastructure atop frontier models.

Chatbots versus coding agents

Karpathy distinguishes between casual users and professional developers as occupying opposite ends of the AI capability spectrum. In an earlier April post, he contrasted OpenAI's free voice mode, which he characterized as struggling with elementary questions, against the company's advanced Codex model, capable of restructuring entire codebases over the course of an hour.

Karpathy's own experience illustrates this transformation. Less than a year prior, he was still writing code manually with occasional autocomplete assistance. By April, he was describing agents that "melt programming problems that you'd normally expect to take days/weeks of work."

In February, Karpathy advanced beyond what he termed "vibe coding" toward "agentic engineering," where developers dedicate more effort to directing and evaluating agents than to writing code themselves.

Why verifiable work wins

"Capability peaks in domains that are digital, verifiable and economically valuable," Karpathy wrote. Coding and mathematics prove particularly suited to reinforcement learning because outcomes are testable and improvements quantifiable. Conversely, writing and advisory work present evaluation challenges, and greater financial incentives exist for enhancing technical models. As Karpathy noted in April, "The goldmines are elsewhere, and the focus comes along."

Karpathy believes these systems are being underutilized. "A question with a paragraph answer barely stresses the system," he observed. Substantial improvements emerge when agents dedicate hours to problem-solving, implementing modifications, and iterating.

Karpathy's AutoResearch initiative demonstrated this potential in March, executing 50 experiments overnight on a single GPU, identifying an improved learning rate, and committing the modification to Git. Each experiment produced a validation score, enabling the agent to assess whether modifications warranted retention before proceeding to subsequent experiments.

From capability to deployment

Karpathy characterizes the present moment as "the weirdness of the moment." The majority of people evaluate AI based on their chatbot interactions, while a select group of developers observes capabilities advancing at extraordinary velocity. Levie perceives this same divide from an enterprise perspective, where organizations must determine how to deploy these systems affordably, securely, and reliably at scale. For Box, this represents the core opportunity.