Andrej Karpathy says software is changing (again)
The OpenAI founder and Tesla Autopilot leader has a lot to say about how LLMs are changing software
Andrej Karpathy is quickly becoming my favorite voice in AI. He was influential at both OpenAI and Tesla, and now spends a lot of time giving talks and making YouTube videos about AI. He recently gave this talk for YC, and it’s making it’s way around the internet:
It’s worth a watch (like most of his videos), but it’s rather long (again, like most of his videos) so I’ll unpack it for you here!
Software 3.0
First, he quickly goes over the first two types of software and then introduces a third.
Software 1.0 is traditional programming. Think the code that most software engineers write.
Software 2.0 is the era of neural networks. The “instructions” are encoded in weights, not human-readable.
Software 3.0 is programming by prompting LLMs. Natural language has become just as valuable as a typical program.
Software 3.0 is often the result of AI agents in tools like Copilot or Cursor generating traditional code, enabling developers to move faster or unlocking software for non-developers. He makes a nod here to Vibe Coding, a term he originally coined!
LLMs are like electricity
Another thing he touched on, that admittedly felt out of place in the talk, was how he views LLMs a lot like utility providers. He points out a few major similarities:
AI labs spend capex to train LLMs (like electricity companies building the grid)
AI labs spend opex to serve intelligence over increasingly homogenous API (operating the power grid)
LLMs provide metered access (pay per token like pay per Kw/hr)
There is intense demand for low latency, high uptime, consistent quality (across both)
Intelligence brownouts (openAI goes down) are becoming increasingly disruptive
I don’t think these comparisons are well connected to the rest of the talk, but I never thought about it this way and I’m glad he included this bit.
If you’re new here, we publish weekly deep dives into how software developers are using AI at work. Consider becoming a subscriber! If you want to check out the library of existing guides first, this is a good place to start:
AI is a new type of user on the internet
A big takeaway from this talk is acknowledging that we should prepare for a new wave of “people” on the internet. LLMs!
Some companies, like Vercel and Stripe are already providing documentation in a markdown format optimized for LLM context on their websites.
He describes how his last vibe coding project was incredibly fast to develop, but took forever to get past authentication, payments, and deployment. This follows exactly what I saw when I used AI to make a micro-SaaS last week, so I chuckled hearing this. But he had an important insight that I didn’t:
The docs are telling me what to do, like you do it! Why am I doing this? What the hell? I had to follow all these instructions, it's crazy
Vercel is taking a stab at solving this by adding a cURL request everywhere in their documentation that says “click” so an LLM can take UI actions on your behalf.
If I were a developer tools company, this would be a trend I’d be very interested in.
There is plenty of opportunity, including for developers
Here’s the part that made me happy.
Andrej says he’s “super bullish on software development. Just because everyone is a programmer because of “Software 3.0” doesn’t mean there isn’t a need for extraordinarily talented folks in the software space. Good use of AI tools, combined with hard-won experience, will be even more valuable as software continues to eat the world.