The only skill left | Field Notes from a Centaur

I figured it out together with half of Twitter at the end of 2023, right around the time language models got good enough that you could copy and paste code between ChatGPT and your IDE, before there was Cursor (an AI code editor) or anything. And the past three years I spent every waking hour building things.

I was taught that my life trajectory would include learning about the world until a certain age, and then apply those learnings to become a useful member of society throughout the rest of my life.

At no point was there a conversation about how to keep learning about learning and the importance of changing my abilities to keep pace with the world around me.

AI shatters this concept and creates space for unlimited upskilling in areas of my life I never considered possible. In the past year, that has meant becoming a full-time software developer. No longer limited by my impatience for writing software line by line, instead I build complexity by endlessly iterating and improving tools for myself and others.

I just copy and paste things. Copy errors into the agent, copy requests into the agent. You just know where to ask. That's the only skill left.

I know every week I'm impressed, but really, it's absurd what Claude does, and how far ahead we are in doing this. Because anyone can do it, but nobody is doing it, not to this degree. We always were technical, even before the tools allowed it. Being a research institute that works with software makes a lot easier. It lets us work in ways that many people don't even know are possible.

I haven't really been writing the newsletter. I've been coding. It's ridiculous. Every Monday I'm like, I should write about how much I've been coding. And then I just go back to coding and I forget there's people.

How the cost of software has gone away allows us to do things with the right people that were just not possible, because suddenly the development cost is truly zero. It's ridiculous.

These tools are incredible for building prototypes. None of that complexity has gone away. It's still incredibly difficult to do technical things like authentication and data storage, and it gets messy very quickly. But of course we get from zero to something in no time. And the models are so powerful now that if you don't direct them exactly, you can really mess things up. So specs are doubly important.

The biggest learning has been how we can build with these tools, and how radically different it is from everything that preceded it. We're confident in our ability to build just about anything. It's crazy how software has become a non-issue. That doesn't mean issues have gone away. They've moved somewhere else in the stack. How do you find users? How do you find funding? That's become the thing. But the time it would've taken us to do wireframes is the time it takes us to do the app now.

The poverty of imagination, the failure of imagination to see what the future actually is, is a real thing. It was always hard for me to accept, but the more I see, the more I notice how difficult a time people have imagining different futures. And I think the vibe coding tooling can be extremely powerful to change people's minds about what is possible. I've trained maybe 500 people in the last year and a half.

Until recently, I would recommend most leaders and the tech-curious to keep a “360º” approach to technology at large: spend time getting to know different branches of emerging tech, in order to get a sense of what might affect your business or work. Nowadays however the only sensible recommendation I give people is: establish a personal (or collective) AI approach. How are you integrating it into your workflows? How are you personally upskilling and developing technical competencies in the space? What are you waiting for until you start doing any of this?

If you have many of these systems working together, you end up with something called agents. In the long term, I think we finally get to artificial general intelligence as multiple agents taking over parts of the economy. The short version of what I think is happening is that we're moving towards job replacement.

On agents built into operating systems, and Signal's Meredith Whittaker's warnings about them: I used to be more optimistic. Now I agree more and more with her, and people like her, in terms of where the puck seems to be heading.

I am sure you are also experiencing AI fatigue. The news can be overwhelming, and the conversations about its manifold implications can feel draining. I am not here to fan the flames. Rather I think we should spend more time experimenting with it. The only race worth mentioning is the one towards greater, better use of these tools. Depending on your use, they can be borderline magical.

There is no doubt whether you are going to see more AI in the months and years ahead. I think that’s a certainty. The only question worth asking is what role you want to play in this transition.

I feel like I've said the same thing so many times that I just want to stop talking about it. But there's more to it. I have to get back to talking a little bit less to Claude and a little bit more to my readers, because that's what it comes down to.