Karpathy sees Cursor as evidence that a new category of AI applications is emerging. Startups should see themselves not as competitors to large language modeling labs, but as specialists serving vertical markets.
So-called “AI wrappers” have been a topic of discussion ever since apps started being built on language models from major AI labs. It is an application that is optimized for a specific task or audience, but its core functionality comes from the underlying language model. The question is: Can these apps differentiate themselves enough to survive?
Former Tesla AI chief Andrej Karpathy recently shared his thoughts on the future of these AI startups. He finds the rise of Cursor, an AI-powered code editor, particularly impressive. According to Karpathy, the tool “compellingly revealed a new layer of 'LLM apps.'” People are now talking about “Cursor for X,” a sign that a new paradigm is taking root.
Why the new app layer works
According to Karpathy, LLM apps like Cursor bundle and tailor LLM calls for specific industries. He identifies four core capabilities. First, these apps perform a function known as “context engineering.” They prepare and structure the context that is fed into the language model and receive this arduous but important work from the user.
Second, multiple LLM calls are orchestrated internally and “incorporated into increasingly complex DAGs, carefully balancing performance and cost tradeoffs.”
Third, the LLM app provides “an application-specific GUI for humans participating in the loop.” Fourth, it provides an “autonomy slider.” Users can decide how much control they hand over and how independently the AI operates.
Startups vs. Big Labs: Who Will Win?
The AI industry has been fiercely debating just how “thick” this new app layer really is ever since the “wrapper” first hit the market. The central question is: Can large language modeling labs such as OpenAI, Anthropic, and Google themselves cover all applications, or is there room for specialist providers to step in?
Karpathy suspects that the LLM Institute tends to train what he calls “generally competent college students,” models that are general but not specialized. LLM apps, on the other hand, organize and fine-tune teams of these models, turning them into experts placed in specific industries.
The key lies in personal data, tools for action, and real-world feedback. According to Karpathy, anyone who can feed information into AI to trigger commands, send messages, or control machines has a good chance of taking on big labs.
However, the institute will not back down easily. OpenAI has made it clear that it wants to cover the entire AI value chain, from chips to apps. Anthropic and Google are also continually developing AI chatbots to handle an increasing number of daily tasks.
AI News Without the Hype – Curated by Humans
as The Decoder Subscriberyou can read without ads. Weekly AI Newsletterexclusive “AI Radar” Frontier Report 6 times a yearaccess to comments, and Complete archive.
Subscribe now
