Computing has had three eras. The PC era, defined by Windows and macOS, gave everyone a file system, a mouse, and applications that did one thing each. The mobile era, defined by iOS and Android, put a computer in every pocket and replaced menus with touch. Both eras were defined not by the devices themselves, but by the operating systems that ran on them.

We are at the beginning of the third era. And this time, the operating system is the AI.

For the rest of us: what is an operating system?

Before the architecture, a quick grounding.

An operating system is the software layer between you and the hardware. When you open a browser, the OS decides how much memory it gets. When an application needs to save a file, the OS handles that. The OS is the thing everything else runs on top of — the foundation that makes everything above it possible.

The idea Karpathy is making, and that an entire generation of engineers is now building around, is that the LLM is the new OS. Not a browser. Not an app. The underlying layer that everything else connects to.

The Karpathy thesis

Andrej Karpathy — co-founder of OpenAI, former director of AI at Tesla — has been the clearest voice on what’s actually changing. “Looking at LLMs as chatbots is the same as looking at early computers as calculators,” he wrote in a post that reframed how many engineers think about this. The large language model, in his framing, is not an application. It’s the kernel.

The CPU/RAM analogy maps cleanly. In a traditional operating system, the CPU is the central processor — the unit that executes all commands. The LLM is the new processor. The context window is RAM — it holds the current working set of information the system needs to operate. Tools and integrations (browsers, code interpreters, file systems, external APIs) are the peripherals and drivers. AI agents are the long-running processes — programs that use the kernel’s resources to complete complex tasks over time.

This is not a metaphor for convenience. It is the actual architecture of how modern AI systems are being built.

Three eras of computing

Every major technological era has been democratised by a defining operating system. DOS gave way to Windows and the GUI — and suddenly people who couldn’t type commands could use a computer. iOS and Android moved computing off the desk and into the pocket — and suddenly computing happened everywhere, all the time.

The third shift is just as fundamental. The interface is changing from menus and icons to natural language. The paradigm is changing from applications that do specific things to agents that handle open-ended tasks. The underlying assumption — that a computer waits for an explicit command — is being replaced by systems that anticipate, plan, and act.

What’s different this time is the pace. The shift from PC to mobile took roughly 25 years. The AI OS transition is moving faster, in part because it doesn’t require new hardware — it runs on infrastructure that already exists.

The interface shift no one planned for

Every OS shift changed who could use a computer and how. Moving from DOS to Windows didn’t just add a GUI — it changed who had access. Moving from desktop to mobile didn’t just add touch — it changed when and where computing happened.

The AI OS is changing what computing means. The generation entering the workforce now doesn’t think of computers as desktops with files and folders. They think of them as entities you talk to. For them, the query box is not a search interface — it’s a reasoning interface. When they need something done, the question isn’t which application to open. It’s what to ask.

This shift is already measurable. Research on computing behaviour shows clear differences between people who grew up with file-based, application-centric paradigms and those who grew up with AI-native tools. It’s not about skill level. It’s a different mental model of what a computer is for.

What this means if you’re building software

If the LLM is the kernel, then everything built on top of it follows a different logic. Applications become agents. Files become context. Interfaces become conversations. The workflow layer, the application stack, the knowledge management systems — all of them are being rebuilt for a different underlying model.

The clearest signal is where developers are going. Cursor and Claude Code are replacing IDEs, not sitting alongside them. This is kernel-level substitution, not application-layer addition. The same pattern is showing up across enterprise software: the question is no longer “which SaaS tool handles this workflow?” It’s “which agent do we configure to handle this?”

For organisations thinking about AI strategy, this reframe matters. You’re not evaluating features in an application. You’re choosing an operating system layer — and that choice compounds over time in the same way that Windows vs macOS choices compounded through the 90s.

Still early

We’re at the equivalent of 1984 for the PC — the paradigm exists but the ecosystem is forming. Most software today still runs on the old model. Most enterprise interactions are still command-interface, not natural language. The native applications for this OS largely don’t exist yet.

But the direction is set. Every major computing era has been defined by a new OS. The organisations that understood this early — and built for the new paradigm rather than the old one — are the ones that set the terms for everyone else.

The question is not whether AI becomes the new computing layer. It’s who builds it, who governs it, and what obligations come with owning infrastructure that everyone runs on.

References

  • Andrej Karpathy, X post on LLM OS (September 2023) — x.com/karpathy
  • Andrej Karpathy, “Intro to Large Language Models” talk (2023)
  • a16z, “Emerging Architectures for LLM Applications” (2024) — a16z.com
  • Rutgers University, AIOS/LLMOS paper (2024) — formalising the LLM-as-OS paradigm
  • InAirSpace, “Current Trends in Human-Computer Interaction 2025” — inairspace.com