Andrej Karpathy Just Redefined Software, and It Blew My Mind
I just watched a talk by Andrej Karpathy (formerly of Tesla and OpenAI), and I have to say, it's fundamentally shifted how I see the future of coding. He argues that we're in the middle of a massive change, the biggest in 70 years. He calls it Software 3.0.
If you're a developer, an engineering student, or just tech savvy, this is a concept you need to understand.
The Evolution: From 1.0 to 3.0
Karpathy's breakdown is brilliant in its simplicity. I believe he's nailed it:
- Software 1.0: This is the code we write by hand.
if/elsestatements, functions, classes. It's the world of Python, C++, and Java that we all know. - Software 2.0: This was the deep learning revolution. The "program" isn't our script; it's the millions of weights in a trained neural network. We became curators of data, not just writers of logic.
- Software 3.0: This is the new frontier, and it's powered by LLMs. Here, the computer itself is a general purpose reasoning engine. And the programming language? English. A simple prompt is now a program.
LLMs Aren't Just Tools—They're a New Kind of OS
This is the part that really clicked for me. Karpathy suggests we stop thinking of LLMs as just another API. We should see them as a new kind of Operating System.
- The LLM is the CPU.
- The Context Window is its RAM.
- Prompting is like using the command line.
We're basically back in the 1960s mainframe era—compute is expensive, centralized, and we're all just "time sharing" on a massive, intelligent machine. From this perspective, it all makes sense.
The Superhuman Intern with Amnesia
So, what are we programming? Karpathy calls LLMs "people spirits"—stochastic simulations of humans. I agree, and that means they come with some serious quirks.
They have superpowers (encyclopedic knowledge) but also huge deficits: they hallucinate, are terrible at simple logic, and, most importantly, they have no long term memory. Every conversation starts from scratch. They are brilliant but flawed, like a superhuman intern who forgets everything every morning.
The Real Opportunity: Iron Man Suits, Not a Robot Army
So, if full automation is still a long way off, what's our job now?
Karpathy's answer is perfect: we should be building Iron Man suits, not an army of Iron Man robots.
We need to create tools that augment humans, not replace them. Think of the coding assistant Cursor—it gives the developer an "autonomy slider." You can ask for a little help or a lot, but you are always in control.
The goal is to speed up the generation verification loop. Let the AI generate, but make it incredibly fast and easy for a human to verify. This is where great UI/UX becomes more important than ever. A visual diff in a code editor is infinitely better than a wall of text from a chatbot.
My Final Take
Karpathy's talk confirmed what many of us have been feeling: the ground is shifting beneath our feet. Our job is evolving from just writing code to architecting human-AI systems. We need to build our apps, our docs, and our APIs to be understood by both humans and these new "people spirits."
It's a massive challenge, but I find it incredibly exciting.
What do you think? Are you already building "Iron Man suits"? Is Software 3.0 just hype, or is it the future?