In his conversation with Kent C. Dodds, recorded while Kent was in Buenos Aires for Nerdearla, Gabriel Benmergui, founder of Silver.dev, pushed back on one of the most common promises about AI: that soon you'll just tell your agents "make an Uber clone" and that's it. You type a prompt, the agents build it, and you have a company.
It's not going to work that way.
Coding agents are getting very good, and they really are changing how fast software gets built. But building software and building a successful product are two different problems. AI is solving the first one much faster than the second.
Uber was never just the code
Getting an agent to write a ride-hailing app is believable today: maps, a matching screen, payments, a driver view. That's the part people picture when they say "make an Uber clone."
But the code isn't what made Uber work. What made it work was everything around the code:
- Real market data. Which cities, which prices, which drivers, which riders, and what actually happens when you launch there.
- Reacting to users. Seeing how people actually use the product, where they get stuck, what they ignore, and what they ask for that you never planned.
- Timing. It has to be the right moment for the product. The same app built a few years earlier or later is a different bet.
None of that lives in a repository. An agent can only build what you tell it to build, and knowing what to build is the hard part.
The loop that doesn't compress
Building a successful digital product is a back and forth. You ship something, people use it, you learn from what they do, and you change the product. Then you do it again.
Coding agents make the "ship something" step much faster. They don't make the rest of the loop faster, because the rest of the loop depends on people. You need users on your platform to get the information that tells you what to do next, and users take time. They have to find the product, try it, come back (or not), and show you what matters to them.
So no matter how good coding agents get, that back and forth still takes time. Faster code means more turns of the loop, not fewer loops.
Kent took the point further in the same conversation. Borrowing the title of the Google paper that kicked off modern AI, he put it this way: "Attention of people is all you need." Distribution, marketing and figuring out what users actually want are really difficult, and they take time. He added a second ingredient from a chat with TanStack creator Tanner Linsley: trust. Attention and trust are human things, and for now they are where the real differentiation lives.
What this means if you're building a startup
If writing code stops being the bottleneck, the bottleneck moves somewhere else. It moves to the decisions about what to build.
That changes what you should look for in the people you build with, including engineers:
- Engineers who look at usage. People who open the analytics, read the support tickets, and talk to users, not just close tickets.
- Engineers who use agents to run more experiments. The point of shipping faster is to learn faster. The best people use the extra speed to test more ideas against real users, not to pile up features nobody asked for.
- Product judgment. Knowing when something is working, when to kill it, and when the timing is wrong is now a bigger part of the job.
Everyone is getting access to the same powerful coding tools. What still separates teams is how well they listen to the market and how quickly they act on what they hear.
This is also how we think about AI at Silver.dev when we evaluate engineers. Our AI Adoption guide (in Spanish) starts from the same idea: use AI to move faster without giving up control, because you can't outsource understanding. Hard tasks are still hard; what changed is that many easy tasks can now run in parallel. Our AI Adoption interview checks exactly that: whether a candidate can turn an ambiguous problem into clear work for agents and still own and explain the result.
Building product is still hard
AI is making it easier to build software. It isn't making it easy to build a product people want. That still takes real market data, real users, the right timing, and the patience to iterate.
The prompt "make an Uber clone" gets you an app. It doesn't get you Uber.
From Gabriel Benmergui's conversation with Kent C. Dodds. Watch the clip and read its transcript on this page, or watch the full interview on YouTube.