Andrej Karpathy envisions a computing paradigm where neural networks become the host operating system, relegating CPUs to deterministic co-processors. Speaking on Sequoia Capital's YouTube channel, he argued that most classical software is a historical artifact of the 'calculator path' computers took in the 1950s.
Key facts
- Karpathy spoke on Sequoia Capital's YouTube channel.
- He describes current AI as 'running virtualized' on classical hardware.
- Classical computing is called the 'calculator path' from the 1950s.
- Future UIs would be generated by diffusion, not built by product teams.
- No major hardware vendor has announced a neural-first architecture yet.
Andrej Karpathy, in a recent interview on Sequoia Capital's YouTube channel, laid out a vision where the traditional hierarchy of computing inverts. Instead of a CPU running an operating system that hosts applications — with AI bolted on as a tool or API — Karpathy imagines the neural network itself becoming the primary runtime.
"You could basically imagine, completely neural computers in a certain sense," Karpathy said. "Imagine a device that takes raw videos or audio into basically what is a neural net, and uses diffusion to render a UI that is unique for that moment."
This is not a prediction about faster app development. It is a structural claim about the nature of computation. Karpathy traced the current paradigm to a historical fork: "In the 50s and 60s, it was not really obvious which way it would go. Of course, we went down the calculator path and ended up building classical computing."
The implication is that most software today — every app, every interface, every deterministic API call — is an artifact of a world where computers could not reason. Once a sufficiently capable neural network becomes the host process, the need for pre-built interfaces and explicit instruction sequences collapses.
[According to @rohanpaul_ai], who summarized the interview, Karpathy's point is "not simply that AI will help us build apps faster; it is that many apps may be artifacts of a world where computers needed every intermediate step spelled out."
Karpathy's vision aligns with emerging product directions from companies like OpenAI (GPT-4o's real-time vision and voice) and Google (Project Astra), both of which demonstrate systems that ingest raw sensor data and produce contextual responses without a rigid app shell. However, no major hardware vendor has publicly committed to a neural-first architecture.
The key tension: today's AI models run virtualized on classical hardware. Karpathy acknowledges this, calling it "neural nets currently running virtualized on existing computers." The transition to neural-as-host would require either new chip architectures (like Groq's LPUs or Tenstorrent's RISC-V-based designs) or a software abstraction layer that makes existing hardware behave like a neural co-processor.
Where the deterministic tail remains
Karpathy is not arguing for the complete elimination of classical code. He positions conventional software as a "small deterministic accessory for tasks where exactness still matters." This is a nuanced position: arithmetic, cryptographic verification, memory safety, and low-level drivers will likely remain in the classical domain. But the user-facing logic — the app itself — becomes ephemeral, generated by the neural host on demand.
What to watch
Watch for any major hardware vendor (Nvidia, AMD, Intel, or a startup like Groq) to announce a chip designed specifically for running neural networks as a host OS rather than as an accelerator. Also track whether OpenAI or Google ships a product that eliminates traditional app shells entirely.
