The marginal cost of making software is dropping to effectively zero, and that changes what craft is for. Plenty of what gets built with AI still breaks, but working software is quickly becoming table stakes, and table stakes don’t set anyone apart. We’ve seen a run of stories saying this is the end of intellectual property in software, since anything can be reverse engineered, and that the only moats left are network effects and distribution. I think that’s a fair read of the business and it misses the point for the people doing the building. When anything can be made, the thing you make stops lasting. What lasts is the practice of pushing past what’s possible, outside the distribution that models give us.

What has always separated art from design is who it’s for. Design was for the person using it. It was never verifiable, but you could tell whether it worked, because you could see whether it actually solved somebody’s problem. Art was for the person making it, and in its truest form it only mattered in the eye of the creator. So once function gets cheap, design inherits art’s question. The question becomes whether you can see yourself in what you made, whether it’s a true expression of who you are as a creator, and not just whether it meets your intent as a prompter. That’s how we keep our sovereignty as creators and builders.

This isn’t about turning every project into an art exhibit. When you build things other people are expected to use, function still meets form, and the user gets something specific from the maker being present in the work, which is care. Care is a signal that the product will keep improving. It also shows up as a level of quality that catches what a specification misses, and that really only comes from obsession with the details. Whether or not people can tell AI was in the loop, they can feel the difference between something the creator cared about almost irrationally and something that was yoloed, shipped fast and never thought about again. Care isn’t polish, though. There’s plenty of prior art showing that aesthetics alone don’t make a product more usable, and maximalism often gets in the way of the thing being made.

Abundance cuts the other way too. Users can now prompt their way to software that matches their own intent, so why would they use yours? Partly because you can’t prompt for a taste or an articulation you don’t have. And partly because software made for one person loses something that shared software has. A version built just for you is bespoke to a narrow slice of a workflow, while a product that millions of people use gives them a shared experience and, with it, a community. Figma did this for designers, and Config is where the software and the people who use it come together. That’s a separate matter from personalization inside existing software, which is a different conversation altogether.

The real threat is to the maker. Your judgment gets baked into the weights of future models. That isn’t necessarily bad for society, and the most altruistic version of these models is that they gather up the wisdom of everyone and push our collective good forward. But at an individual level it means that what you can do today, a model will do tomorrow. I’ve seen this really clearly with my own work. I estimate that what took 25 engineers a year to build at URX, my first startup, I could now build myself in a few weeks. I think models pull everything toward the center of their distribution, which is basically prior art, and it’s way too easy to live there. If you aren’t pushing past the current frontier, you become part of that distribution, and whatever made your contribution unique gets commoditized.

URX then and now, in engineer-weeks A grid of 1,300 small grey squares, 25 rows of engineers by 52 columns of weeks, each square one engineer-week: what it took to build URX. In the bottom-left corner, 3 squares are green: one person for a few weeks, what it would take now. 52 weeks25 engineers now: one person, a few weeks then: 1,300 engineer-weeks
One square is one engineer-week. The grey is what URX took. The green is my estimate for building it alone today.

So the edge is a treadmill. Every new model widens the distribution, and nothing you build out there stays outside it for long, so I don’t think any single artifact is durable anymore. The pushing is the thing that’s actually durable. If the moats in software really are eroding, the moat moves from the product to the person who makes it. In 1967 John Culkin wrote, in an essay about McLuhan, that we shape our tools and thereafter they shape us. We’re always building on the shoulders of giants, and if the tools now shape us, we need to keep pushing past what we could do before and redefine our own role, in the economy and in the creation process.

Outside the distribution Three nested bell curves share one centre, each wider than the last: what models could make then, later, and now. The oldest region is hatched densest. A green mark for your work sits just past the edge of each curve, and an arrow hops it outward each time the next, wider curve swallows it. The last hop, still dashed, lands in the empty space outside the widest curve. what models can make outside the distribution your work then later now next
Hatching is what models can already make, densest where it's been true longest. Each new model swallows the work just past the last one's edge.

Part of the reason to push is for yourself, which was always art’s motive. If you want to be one of the best at what you do, and that is in fact how you see your career and your ambitions, then when the tools catch up to what you used to be able to do, you should aspire to push even further past it. That’s an ethic of excellence. The other part is for the user, which was always design’s motive. If you call yourself user obsessed, I believe you have a moral obligation to care for the people using what you build, and to help them create joy in a way that wasn’t possible before.

The risk is that all of this becomes a treadmill with no direction, and that’s explicitly not what I mean. The direction comes from being truth-seeking, from understanding and articulating real problems as they exist in what’s been made, which means staying grounded in the reality of the product and the people using it. In practice I get outside the distribution by burning through the existing permutations really fast. I can explore 100 variations in the time it used to take to make two, see where the gaps are, and find approaches to the root problem that maybe haven’t been tried. It also lets me understand what I’m building at a deeper level than was possible before.

So pushing can lead to polishing, and the clearest example I have is the launch of Canva AI 2.0. Hundreds of engineers were working on a system many of them hadn’t touched before, with lots of separate systems being orchestrated together. To a user it felt like a single prompt, but how the result got generated wasn’t known to many of the people building it. Everyone knew the orchestration mattered, and it was such a large undertaking that no single person felt they could own it. I decided to own it. I built a simulation environment to map every permutation of how these systems could interact, found the edge cases between them, and made judgment calls on routing across each one, on the homepage, in the editor and in the side panel, and on how the tools an agent used would work together.

Users felt it as fewer paper cuts. As an example, if you asked for help ideating or concepting, the system would read that as a document creation intent and route you into making a document, when it should have stayed in the chat. Seams like that are exactly what nobody really owns on a big team, and they’re what users feel even when they can’t name it. The work outlasted my time at Canva, and it kept the team thinking about what the end user would experience instead of treating the system as a machine. When one person owns the seams, everyone else gets to own their parts without the whole thing falling apart, and I think AI shrinking teams makes that kind of craft more possible again, as long as leaders make sure faster timelines leave room for it instead of turning into slop cannons.

One paper cut, before and after A prompt asking to brainstorm ideas for a spring launch goes into the router. Before, the router read its intent as a document and opened a new document. After, it read the intent as ideation and the conversation stayed in the chat. prompt brainstorm some ideasfor our spring launch router intent: documentintent: ideation BEFOREAFTER opens a documentstays in chat
Before, a request for ideas got read as a request for a document and opened one. After the routing work, it stayed in the chat.

I’m still kind of working out how all of this applies to my own craft as a product leader. On big teams it’s easy to spend all your time on the machinery of getting people to work together, making sure the specs and requirements are right and the product does what you hoped. If that’s all you do now, you miss the forest for the trees. Durable user value, delighting customers and helping them do something they couldn’t before still come first. But that part is getting easier, and once it’s done, you get to work outside the distribution, in the space that used to belong to art. Part of the joy we’ve gotten from software over the last half century is the act of making, sure, but a lot of it is creating something that’s never existed before and having it live outside of yourself as a reflection of your values, your principles and what you care about. To me, that’s one of the main drivers of my own work.