Dec 23, 2025

Nano Banana Pro prompted by THE DECODER

• Andrej Karpathy, former head of AI at Tesla, views the AI code editor Cursor as evidence of a new category of applications that bundle and orchestrate language models for specific industries.
• According to Karpathy, these "LLM apps" serve four core functions: preparing context for language models, chaining multiple LLM calls into complex workflows, providing application-specific interfaces, and letting users control their own level of AI autonomy.
• Karpathy suggests that startups can set themselves apart through private data, specialized tools, and real-world feedback.

Karpathy sees Cursor as proof that a new category of AI applications is emerging. Startups shouldn’t view themselves as competitors to the large language model labs but as specialists serving vertical markets.

Ever since apps started building on top of language models from the major AI labs, there’s been an ongoing debate about so-called "AI wrappers" - applications optimized for specific tasks or audiences, but whose core capabilities come from underlying language models. The question: can these apps differentiate enough to survive?

Former Tesla AI chief Andrej Karpathy recently shared his take on the future of these AI startups. He finds the rise of Cursor, the AI-powered code editor, particularly striking. According to Karpathy, the tool "convincingly revealed a new layer of an ‘LLM app.’" People now talk about "Cursor for X," a sign that a new paradigm is taking hold.

What makes the new app layer work

Karpathy says LLM apps like Cursor bundle and orchestrate LLM calls for specific verticals. He identifies four core functions: first, these apps perform what’s known as "context engineering." They prepare and structure the context fed to the language model, taking this strenuous but important work from the user.

Second, they orchestrate multiple LLM calls under the hood, "strung into increasingly more complex DAGs, carefully balancing performance and cost tradeoffs."

Third, LLM apps provide an "application-specific GUI for the human in the loop." Fourth, they offer an "autonomy slider": users can decide how much control to hand over and how independently the AI should act.

Startups vs. the major labs: who wins?

The AI industry has been intensely debating how "thick" this new app layer really is ever since "wrappers" first appeared on the market. The central question: will the large language model labs like OpenAI, Anthropic, or Google cover all applications themselves, or is there room for specialized providers?

Karpathy suspects the LLM labs will tend to train what he calls the "generally capable college student" - versatile but not specialized models. LLM apps, on the other hand, would organize and fine-tune teams of these models, turning them into deployed professionals in specific industries.

The key lies in private data, tools for action, and real-world feedback. According to Karpathy, anyone who can feed information to an AI and let it do things—like trigger orders, send messages, or control machines—has a good chance of holding their own against the major labs.

But the labs won’t back down easily: OpenAI has made crystal clear the company wants to cover the entire AI value chain, from chips to apps. Anthropic and Google are also continuously developing their AI chatbots to handle more and more everyday tasks.

AI News Without the Hype – Curated by Humans

As a THE DECODER subscriber, you get ad-free reading, our weekly AI newsletter, the exclusive "AI Radar" Frontier Report 6× per year, access to comments, and our complete archive.

Subscribe now