From LLMs as OS to Outcomes
date
Jul 5, 2025
slug
llms-as-os-outcomes
status
Published
tags
Principle
LLM
summary
LLMs as operating systems for computational intelligence, with prompts serving as the new system calls. This Software 3.0 approach allows for free-at-the-margin intelligence, with coding emerging as the first killer application due to its repeatable and verifiable nature.
type
Post
Andrej Karpathy's talk is a great watch. One idea from his presentation keeps echoing in my head: LLMs as the operating system for computational intelligence, and the autonomy dial. These are already flipping our approach to work.
The Operating System of Intelligence
What if we think of LLMs not as smarter autocompletes but as the abstraction layer for computational intelligence? Just as traditional operating systems provide a clean interface that hides the messy complexity of hardware from applications, LLMs hide the incomprehensible complexity of those transformer weights and neural networks behind a simple, shared, natural interface: human language. Brilliant.
This is what Karpathy calls Software 3.0. Where instead of writing explicit code, we access raw computational intelligence through natural language. The transformer architecture and those billions of trained weights? They're the "hardware"—the actual computational substrate where intelligence lives. The LLM is the OS that makes this intelligence accessible.
And prompts are the modern syscalls of this new operating system. Instead of calling functions like malloc() or read(), we make requests in English. "Analyze this data," "write this function," "deepsearch this problem"—these are our new system calls to the intelligence layer underneath.
We now have almost free-at-the-margin intelligence. And the exciting thing? We're at the same stage with LLM as OS as we were with mainframe computers eras of the early 70s where we accessed computers through timeshares.
We have the infrastructure—now we're hunting for the killer apps. And the first one? It's already here.
Coding: The First Killer App
Arguably, the first breakout application for GenAI isn't customer service chatbots or content generation (even though those are moderately successful), it's agentic coding. The Product Market Fit is highest for this category (as demonstrated by the sky high valuations of startups such as AnySphere/Windsurf).
AI agents are particularly effective in software development because code follows defined patterns and is ultimately very repeatable and trainable. It's also verifiable. Through automated tests, agents can iterate using test results as feedback, and output quality can be measured objectively.
This is where Karpathy's autonomy dial concept fits nicely. The progression:
- Low autonomy: AI-powered autocomplete (like GitHub Copilot)
- Medium autonomy: Chat-based code generation and debugging
- High autonomy: AI agents that understand requirements and implement complete solutions
We're already seeing AI agents resolve real GitHub issues and generate complete workflows from simple prompts. It's clear the future of software engineering will include AI agents in one shape or another.
What's Next? The Excel and Photoshop Moments
But coding is just the beginning. Think about the PC revolution: first came the operating system (DOS, then Windows), then the killer app that proved the concept (VisiCalc, then Excel), and finally the creative applications that expanded what was possible (Photoshop, PageMaker).
We're living through that exact sequence right now, just compressed into a much faster timeline.
The infrastructure is live. The first killer app is proven. Now we're racing toward the applications that will define the next decade of work.
And honestly? I can't wait to see what we build.