Andrej Karpathy

There's over half a million downloads, tens of thousands of people using it.

So fairly quickly, Andrej Karpathy, co-founder of OpenAI, took notice and started tweeting about it, said it slightly blew his mind, the approach we took with minimalism and And so we kind of let our customers go from our native marketing agency and went all in and are now building NanoCo, which is the company behind NanoClaw.

Very cool.

Very cool story.

It's a very reasonable question to ask, and, and in fact, we asked that question as well, and we ran a prompt optimization loop on, on GPT 5.5 Codex.

It's effectively just an Andrej Karpathy auto research loop where we say, "Okay, you can see the rubric judge score, uh, and, uh, you can do many, many iterations on the prompt." And it builds up this very complicated prompt which effectively just shouts at GPT 5.5 and says, "Hey, don't do all these cheating things.

Be a rigorous scientist.

Um, really make sure that you try and iterate and do many...

And right now, uh, the latest update that we've seen, which is actually a pretty old one, is that 92% of US developers use AI coding tools at work every single day.

But even Andrej Karpathy, who coined the term vibe coding, has now declared it passe after only one year.

It's because we realize that it's not always all good.

So on today's show, here's what we're gonna be diving into and what you'll learn if you stick with us for the next 20-ish minutes.

You describe something in plain English or whatever language that you speak, and an AI builds software, right? Uh, so it handles all the database, the deployment, the coding, even certain platforms, handle the authentication, payment platforms, all of that.

Uh, right.

So this is a term, uh, that OpenAI co-founder Andrej Karpathy coined in February twenty-twenty-five, which is crazy to think about 'cause it seems like we've been talking about vibe coding for like five years, uh, but it's been like fifteen months.

Uh, right.

I mean, my most important assistant right now is Claude Cowork.

I use it with, basically, right now I'm using Andrej Karpathy's LLM wiki method of writing.

managing its memory, right? So essentially, I tell cloud code work to maintain knowledge in a wiki like structure, right? So it's an index plus a bunch of markdown files that it maintains to sort of keep track of important concepts that I have.

And then I tell it to ingest new information into that wiki, right? So maybe I've read an interesting article, or listen to an interesting industrially I podcast with Peter, And then I tell Claude, you know, can you please take the transcript and then ingest the information onto the right sub pages inside by wiki so that and then anytime I prompt it to do any task, it sort of looks at the wiki first, is there any context in here that is relevant to this task, and then only then performs that task.

Now this cost curve is great for the consumer, but it's made competition even more intense as a lot of different foundational models begin to compete for user attention and hence the whole commoditized thesis that I have.

John, I don't know if you saw this, but there was this really funny story that came out late 2025 when Andrej Karpathy, the co-founder of OpenAI and former head of AI at Tesla, spent a weekend building a tool that routes any questions simultaneously across all these different models.

So he looked at ChatGPT, Gemini, Claude, Grok, and so on.

And he called this his LLM Council.

And if it thinks it's something I'm going to use more in the future, like the email one, it will create like a meta concept page.

That is created, I use a skill called llm-wiki, which is basically based on what Karpathy shared, this idea of like building a wiki that a human reads, but it's only updated, touched and created by an llm.

And in this case, basically what it will do is it will have concepts that are abstract, created from all of the others.

And it will automatically create them every time I have a new call.