- On October 12, 2026, The New Stack reported remarks by Andrej Karpathy, a founding member of OpenAI who currently works at Anthropic, stating that most of the world does not truly understand the capabilities of advanced AI.
- According to Karpathy’s personal estimates, around 6 billion people have barely used large language models, 1–2 billion use conventional chatbots, approximately 20 million use advanced models for programming and mathematics, and only around 5,000 have access to the most powerful AI systems in research labs.
- These figures are Karpathy’s rough estimates, not the results of a statistical survey.
- Box CEO Aaron Levie argues that this gap reflects the very early stage of AI adoption, with most users still treating AI as a chatbot or an advanced search engine.
- Levie predicts that “agent swarms”—groups of AI agents coordinating to perform multiple tasks in parallel—could become an important form of enterprise AI adoption.
- However, very few people currently use AI this way. The main barriers include operating costs, inconsistent model capabilities, a lack of orchestration tools, and limited access to advanced systems.
- Businesses must also address data integration, security, compliance, and workflow changes before deploying AI agents at scale.
- Karpathy emphasizes the enormous difference between a free chatbot that answers simple questions and an advanced AI agent that can spend hours restructuring software or solving complex problems.
- He says that, within a short period, he shifted from writing code himself with autocomplete tools to using AI agents for programming tasks that previously required days or weeks.
- Karpathy calls this new approach “agentic engineering,” in which engineers focus on assigning tasks, providing direction, and reviewing results rather than writing every line of code themselves.
- According to Karpathy, AI makes the strongest progress in fields with three characteristics: work takes place in digital environments, results can be verified, and the work has high economic value.
- Programming and mathematics are particularly well suited to reinforcement learning because systems can check their own results, detect errors, and improve through repeated experiments.
- Karpathy’s AutoResearch project illustrates this approach: AI conducted 50 experiments overnight on a single GPU, found better learning-rate parameters, and updated the source code automatically.
- Karpathy argues that asking AI to answer only short questions fails to unlock its full technological potential. Greater value emerges when AI is assigned tasks lasting hours, with the ability to experiment and adjust its own approach.
📌 Karpathy’s and Levie’s observations highlight a crucial divide in the AI era: the difference lies not only in whether people use AI, but also in how deeply they leverage its capabilities. While billions of people are just beginning to use chatbots, a very small group has access to systems capable of conducting research, programming, and solving complex tasks autonomously. The next trend is a shift from AI that assists with individual actions to coordinated groups of AI agents that execute entire workflows. For businesses, the priority is not simply to spread AI literacy, but also to develop the ability to formulate problems, organize workflows, verify results, and govern autonomous AI systems.
