Former Tesla head of Artificial Intelligence Andrej Karpathy argues that a distinct category of applications is emerging on top of large language models, pointing to the Artificial Intelligence code editor Cursor as a leading example. In his view, Cursor has “convincingly revealed a new layer of an ‘LLM app'” and sparked a wave of “Cursor for X” pitches, where startups try to adapt the same idea to different verticals. Rather than seeing themselves as direct competitors to large language model providers, Karpathy suggests that these startups should position as specialists serving particular markets or workflows.
Karpathy defines LLM apps as tools that bundle and orchestrate language model calls for specific verticals, and he identifies four core functions that make this new layer work. First, they perform “context engineering” by preparing and structuring the information fed into the language model so that users do not have to handle that complex setup themselves. Second, they orchestrate multiple LLM calls behind the scenes, “strung into increasingly more complex DAGs, carefully balancing performance and cost tradeoffs.” Third, they provide an “application-specific GUI for the human in the loop” that is tailored to a concrete domain rather than a generic chat box. Fourth, they offer an “autonomy slider” that lets users decide how much control to hand over and how independently the Artificial Intelligence should act.
These ideas feed into a broader debate inside the Artificial Intelligence industry about the thickness of the app layer on top of models from OpenAI, Anthropic, Google and others. Karpathy expects the major labs to focus on training what he calls the “generally capable college student” type of systems: versatile models that are broadly useful but not deeply specialized for a single industry. He sees room for LLM apps to take these general models and turn them into deployed professionals by layering on domain-specific context, workflows, and interfaces. The key, he argues, lies in access to private data, tools for concrete actions, and real-world feedback, since “anyone who can feed information to an AI and let it do things” such as triggering orders, sending messages, or controlling machines can build defensible products. At the same time, he notes that OpenAI wants to cover the entire Artificial Intelligence value chain from chips to apps, and Anthropic and Google are steadily expanding their chatbots to handle more everyday tasks, which means startups will have to lean on their vertical focus to stay competitive.
