Andrej Karpathy: Software Is Changing (Again)

In plain English

Karpathy argues that software is changing for the third time. First came hand-written code. Then came systems trained on data. Now we can instruct AI models in plain English. He describes these models as powerful but flawed: they know a huge amount, but they make odd mistakes, forget everything between sessions and can be tricked.

So the best products keep a person in charge, with an easy way to check the AI’s work and a dial for how much freedom it gets. He also says we should make websites and documentation easy for AI agents to read.

Three ideas worth keeping

• AI generates and people check, so make checking fast, for example by showing changes visually instead of as walls of text.
• Keep the AI “on a leash” with small, specific requests, because huge outputs just make you the bottleneck.
• Build “Iron Man suits” that augment a person, not flashy fully autonomous robots, and expect agents to take a decade, not a year.

What it means for your business

Be wary of demos where the AI seems to do everything. Karpathy built a working app in hours, then needed another week to add logins, payments and hosting. Ask any vendor how a person reviews the AI’s work and how you can turn its freedom up or down.

Plain-English instructions also mean your own staff can build useful small tools.

What it means for software teams

Work in small steps and keep each piece of AI output easy to verify. Treat documentation and systems as something an AI agent will read, so write for it directly. Pick how much autonomy to grant by task, and widen it gradually.

What we found, and what we did

We built partial autonomy into how code lands in our own platform, Mibble. Low-risk changes that pass every automated check merge on their own, and anything hard to undo waits for one person’s approval.

Our rule is “grant autonomy, never assume it”, and an automatic brake pauses automatic merging if too many recent merges have had to be undone.

In September about 20 changes landed a day and people approved about 70% of them; now about 45 land a day and over 90% go through with no person approving. Our guide Vibe Engineering: How to Ship Production Code with AI and the field note [Vibe Engineering vs.

Vibe Coding](/blog/vibe-engineering-vs-vibe-coding/) take the same stance: use the tools, but keep checking what they produce.

Worth watching

• 09:26: Why he compares AI models to a new kind of operating system.
• 22:10: Why checking AI output fast, and keeping it on a leash, matters.
• 31:28: His menu app: a demo in hours, a week to make it real.