Y COMBINATOR · EXTRACTED

Software Is Changing (Again) ft. Andrej Karpathy

Andrej Karpathy on the three eras of software, the new LLM stack, and where the real leverage sits in 2026.

"LLMs are a new kind of computer, and you program them in English." — Andrej Karpathy
Andrej Karpathy, on the episode

Karpathy spent years at OpenAI and Tesla building AI systems that ended up in production. He left to teach. This Y Combinator keynote is where he laid out the Software 1.0 / 2.0 / 3.0 framework that's now everywhere: classical hand-written code, trained neural network weights, and natural-language LLM programs. The talk isn't a forecast. It's a working theory of where the leverage is right now, what the stack actually looks like, and what kinds of products are buildable today that weren't 18 months ago. If you build with AI for a living, this is the operating manual.

Tactic 01

Andrej Karpathy, Y Combinator

Pick the right era of software for the problem

The play

For each feature, decide whether it should be code, a trained network, or a prompt.

Andrej Karpathy says

Software 1.0 is the code you write for the computer, and Software 2.0 is the weights of a neural network. In Software 3.0, prompts are programs that program the LLM, and they are written in English. These are three completely different programming paradigms, and it is a good idea to be fluent in all of them because each has pros and cons, so a piece of functionality might be built in 1.0, 2.0 or 3.0.

Try this

Pick one feature you are building and write down whether it should be explicit code, a trained neural net, or a prompt to an LLM, and why.

Tactic 02

Andrej Karpathy, Y Combinator

Treat LLMs as 1960s infrastructure

The play

Design for LLMs that live in the cloud and are reached over the network, not on the device.

Andrej Karpathy says

LLMs are in a 1960s-like era: compute is still very expensive, which keeps them centralized in the cloud. Everyone is a thin client reaching them over the network, time sharing the computer as one dimension of the batch. The personal computing revolution for LLMs has not happened yet, because it is not economical.

Try this

List the features in your product that assume an LLM runs on the user's device, and plan how each works when the model is reached over the network instead.

Tactic 03

Andrej Karpathy, Y Combinator

Build the suit before you build the robot

The play

Build partial autonomy products that a person drives, not fully autonomous agents.

Andrej Karpathy says

The Iron Man suit is both an augmentation that Tony Stark drives and an agent. Working with fallible LLMs at this stage, you want to build fewer Iron Man robots and more Iron Man suits, with less flashy demos of autonomous agents and more partial autonomy products. Humans need to stay in the loop, and this has to be done carefully.

Try this

For each AI action in your product, mark whether a person approves it or it acts alone, and bring one that acts alone back under a person's control.

Tactic 04

Andrej Karpathy, Y Combinator

Make the autonomy slider visible

The play

Let users choose how much autonomy the AI gets, task by task.

Andrej Karpathy says

In Cursor, tab completion leaves you mostly in charge, while command I lets the agent do whatever it wants in the entire repo, the full autonomy version. You are in charge of the autonomy slider and tune how much autonomy to give up based on the complexity of the task. There should be an autonomy slider in your product, and you should think about how to make the product more autonomous over time.

Try this

Pick one AI feature and write down the lowest, middle and highest levels of autonomy a user could choose for it.

Tactic 05

Andrej Karpathy, Y Combinator

Ship the verification UI, not just the model

The play

Make AI output fast to check visually, and keep each change small.

Andrej Karpathy says

The AI does the generation and the human does the verification, so it pays to make that loop as fast as possible. GUIs speed up verification a lot, because reading text is effortful while looking at things is fun. Keep the AI on the leash: a diff of 10,000 lines does not help when the human is still the bottleneck, so work in small incremental chunks.

Try this

Pick one AI output in your product, design a way to check it at a glance, and limit how much the AI can change in one step.

Tactic 06

Andrej Karpathy, Y Combinator

Write your docs for LLMs to read

The play

Offer your docs in markdown for LLMs, and swap click instructions for commands an agent can run.

Andrej Karpathy says

A simple markdown file can tell LLMs what a domain is about, in a form they read easily. Much documentation is written for people, while markdown is super easy for LLMs to understand. Any time docs say click, an LLM cannot take that action, so Vercel is replacing every click with an equivalent curl command an agent can run.

Try this

Pick one page of your documentation, publish a markdown version of it, and replace each click instruction with a command an agent could run.

Tactic 07

Andrej Karpathy, Y Combinator

A perfect demo is not a finished product

The play

Treat a working demo as the start, and plan for the long road to a reliable product.

Andrej Karpathy says

In 2013 Karpathy had a perfect self-driving drive with zero interventions, and the demo made self-driving feel imminent. Twelve years later the work on autonomy is still going, with a lot of teleoperation and human in the loop. Software is tricky in the same way driving is.

Try this

Write down the cases your AI product still gets wrong after its best demo, and pick one to fix this week.

Your action plan

the plays, one checklist.

01
Pick the right era of software for the problem
For each feature, decide whether it should be code, a trained network, or a prompt.
02
Treat LLMs as 1960s infrastructure
Design for LLMs that live in the cloud and are reached over the network, not on the device.
03
Build the suit before you build the robot
Build partial autonomy products that a person drives, not fully autonomous agents.
04
Make the autonomy slider visible
Let users choose how much autonomy the AI gets, task by task.
05
Ship the verification UI, not just the model
Make AI output fast to check visually, and keep each change small.
06
Write your docs for LLMs to read
Offer your docs in markdown for LLMs, and swap click instructions for commands an agent can run.
07
A perfect demo is not a finished product
Treat a working demo as the start, and plan for the long road to a reliable product.