Karpathy's napkin math counts 5,000 people inside frontier labs and 20 million professional users

The spread between those two tiers is 4,000 to 1 ^[13]. By Karpathy's own account the figures are napkin math ^[1]. The coverage does not say how many of the 20 million run more than one agent at a time, and Levie's swarm thesis depends on that count.

Karpathy says what the 5,000 see "looks like swarms of thousands of agents collaborating over weeks on software mega projects" ^[14]. That is his description of systems outsiders cannot inspect. For it to transfer to a company's workflow, the work would need the property that makes the lab examples tractable.

That property is a check. Karpathy says capability peaks in domains that are "digital, verifiable and economically valuable" because it "emerges from reinforcement learning on verifiable rewards" ^[3]. His AutoResearch run shows the loop. In one night on one GPU, the agent worked through 50 experiments, landed on a better learning rate and committed it to Git ^[9]. Every run returned a validation score, and that score told the agent whether a change should stay.

A swarm multiplies the branches. In our view, where every branch has a score the system prunes itself, and where none does a person reads each one. Karpathy wrote that "human review and comprehension are starting to fall behind" ^[15].

Levie's reading is operational. He wrote that "This is a good way to understand how early we actually are in AI diffusion right now" ^[4]. He also wrote that "Agent swarms represent possibly the ultimate form factor of what agents will do inside of our workflows" ^[5], and he envisions systems that divide complex work among agents running in parallel ^[6]. His list of what holds it back is concrete: uneven model capability, expensive scale, immature tools for managing several agents, and limited access to the strongest models ^[7]. He expects "a huge amount of change management" before most industries outside software and AI research can use these agents ^[8].

Levie also has a product in the gap. Box sells Box Automate, which connects AI-powered automation to document-heavy business processes, and Levie has said Box does not intend to compete directly with model providers ^[11]. He wrote that "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" ^[10].

The diffusion reading follows from Karpathy's numbers if they hold. We would treat the applied-layer claim as a thesis from a vendor that sells in that layer. Levie put current usage at "less than a fraction of a fraction of the population" ^[12].