Some things are hard, even with AI

There’s a whole lot of discussion about what’s easier with AI, but let’s talk about what’s still hard. Everything that’s easy will get commoditized. Great, you can build an app in 2 hours. So can everyone else. Value will accrue to the things that are still hard, so let’s talk about what those are.

Figuring out what you should build. AI can help you build faster, but it can’t tell you what’s worth building. And I think AI actually makes this harder because it artificially reduces the perception of how hard it is to build things. Feature creep gets easier. Scope creep gets easier. It’s easier to get distracted by the next shiny thing because the cost of chasing it feels low.

The cost of building used to be a natural filter. When everything is expensive, you’re forced to prioritize ruthlessly. That scarcity drove focus. AI removes the scarcity but it doesn’t remove the need for focus. If anything, discipline around what NOT to build matters more now than it ever did.

Getting AI to work reliably for your specific use case. There’s a big gap between “this model can do impressive things” and “this model consistently produces good output for my product.” Closing that gap takes a lot of prompt engineering, refinement, and iteration. The demo is easy. The last 20% is still very much not.

Good usability and user flow. AI can generate a UI in seconds but “does this feel right” is a different question than “does this work.” The gap between functional and intuitive is still enormous and mostly invisible to the people building the product.

Scalability and security. Pretty unsexy, pretty important. These are fundamentally adversarial problems (especially security) that don’t get easier just because you can write code faster. If anything, more code faster means more surface area to defend.

But the big one is distribution. And this is the one I think people are underestimating the most. Everything else on this list, AI can at least theoretically help everyone get better at simultaneously. Distribution is different. It’s much closer to zero-sum. There’s a fixed amount of attention and shelf space, and if everyone’s using AI to compete for it, you’re in an arms race, not a rising tide.

Andrew Chen wrote about the law of shitty clickthroughs years ago and it’s never been more relevant. Every distribution channel degrades over time as more people flood into it. AI accelerates that cycle dramatically. It’s not enough to be good at distribution, you need to stay one step ahead of everyone else. And that gap resets constantly.

The winners of the next era are going to be the ones who figure out how to get these capable new models to deliver reliable results in the right product format while carving out a distribution advantage. If you take a step back, that sounds remarkably like the AI labs themselves. Which is why they’re winning.

But if you’re building something on your own, the questions are the same. You have to find a unique product or distribution niche and you have to figure out how to fit these AI tools into a reliable workflow within it. The people who solve both of those are the ones that are going to win this next era.