I recently watched this fantastic video on how to make chicken soup from rotisserie chickens using a pressure cooker, and what struck me was the anecdote from Heston Blumenthal's Fat Duck during their process of acquiring their third Michelin star. A process that was well-known for decades, combined with modern technology and a lot of determination completely changed how they made stock. They could have said that the method is unorthodox, but they cared about the result first and foremost, and that's what made all of the difference.

It struck a chord with me because of how my relationship with AI-generated code has shifted. For the longest time, I was not a big fan of ChatGPT's code. I remember saying "I don't think I'm bottlenecked by my ability to type." This has completely changed. Modern agentic systems can churn out thousands of lines of high quality code in minutes. I will never come close to this by using my hands.

Over Christmas break, building side projects with the newest models, I noticed that the level of detail I added to specifications, how much feedback I gave when something went wrong, how often I would iterate, was only tied to how much I cared about the result. Some projects meant more to me: I went back many times to adapt small things. With others, I was happy with the first shot. And that is not going away. Someone still has to care enough to tell the agents what to do, where to improve, what to change. These decisions are not solvable by gradient descent. They are about conscious choices to move in a certain direction - and that is a matter of taste. Where do you point the gradient?

An example from the software world is Linear - you may think it's an issue tracker, how hard can it be to build? But underneath is a deep technological foundation paired with an obsessive desire to make the experience great. It shows that they care.

Everyone has home-cooked software now, everyone knows how bad highly-processed software can be, but with these agents, we can expend immense effort getting the smallest details right, if only we care.

I think there will be tiers of software: At the top, there will be "Michelin Star Software", crafted by people who truly care, with strict quality controls, for users who love a specific product, and reflecting different tastes just as restaurants serve different needs. Then there is "home-cooked software", AI-generated by and for people who have a problem themselves. For example, I wanted an app to crunch my Apple Watch swimming data, asked Claude to make it, and the result was perfectly fine for me: not published, but convenient. At the bottom is "slop", software generated by people who don’t care and that won’t be used. Wasted energy, but probably not something we should worry about that much.

This opens new markets too. We'll be able to produce much more good software than before, because output so far was limited by the intersection of technical talent and the ability to care. One of these is becoming less important. Caring is the bottleneck now.