https://azimmermann.com/posts Practical AI notes, project observations, comparisons, and lessons from Adrian Zimmermann. en Mon, 24 Aug 2026 22:47:51 GMT https://azimmermann.com/p/future-jobs https://azimmermann.com/p/future-jobs Thu, 20 Aug 2026 10:00:00 GMT there will only be gamers, artists, influencers and coaches https://azimmermann.com/p/theses https://azimmermann.com/p/theses Wed, 19 Aug 2026 23:02:05 GMT I keep changing my mind about some of this because the tools keep changing. But these are nine things I currently think. 1. I built an AI product, and then a fairly small configuration for a general coding agent could do a surprising amount of what the product did. This will probably happen to many software products. 2. Companies may adopt AI slowly. Their competitors do not have to. A new company can start with the new tools and none of the old processes. 3. Most people do not want another AI tool. They want the finished result. The useful product is not an AI for doing research or designing something. It is the research done or the thing designed. 4. My bottleneck is increasingly how many good tasks I can think of, how clearly I can describe them, and whether I notice when the result is wrong. The AI can keep working for much longer than I can keep giving it useful direction. 5. In science, the next bottleneck may be the physical experiment. A model can propose thousands of things, but a laboratory still has to test them in the real world. 6. General intelligence may create more specialized robots, not fewer. The intelligence can be reused, but a robot for a hospital should have a different body from one that cleans a pool or works on a building site. 7. Software may become specific to one company, one department, or even one person. Instead of choosing which existing application is closest to what you need, you describe what you need and make that. 8. Building for the AI capabilities of six months from now already seems almost impossible. Even a few weeks can change what is worth building. So it is probably better to stay close to real problems and make things that are easy to change. 9. There is an overhang of capabilities in the current models. They can already do useful work in areas where almost nobody has seriously tried them. I found this with mechanical design. I am sure there are many other examples. I am not sure all nine will survive the next year. That is partly why I want to write them down now. https://azimmermann.com/p/direction https://azimmermann.com/p/direction Wed, 19 Aug 2026 10:00:00 GMT The agents working on Earth Economy Simulator have probably accumulated three or four weeks of runtime by now, basically running 24/7. For something that runs that long, giving it direction still matters a lot. If I stop paying attention for too long, it can build a lot of perfectly reasonable things in the wrong direction. But the directions are becoming more and more high level. I do not spend much time deciding which architecture or technology it should use. The AI can usually figure that stuff out. What I still have to provide are the things that only exist in my head. What should the simulator feel like? What should be visible? What matters more? When is a technically correct result still not the thing I meant? Maybe that is what my job is becoming. Not choosing the architecture, but getting the thing in my head out clearly enough that the agents build the right thing. Which is still surprisingly hard.