The ‘great’ thing with AI coding tools is that any kid can write 10,000 lines of code that appear to work. Could be Python, C++, Go, Java… doesn’t matter. And there won’t be any English typos in the comments. Well, when I say « write », I should say « have a tool write ». And you can even get a novel-length description of the work. Every open source project can get multiple 10,000-line pull requests a week. The intern you just hired can generate 50,000 lines of code in his first three weeks. Brilliant. Now, who wants to read all that? The cost of 10,000 lines of C++ that can compile and run some tests just went to the floor. Nobody cares about it. You can replicate the problem in other domains such as science. It has become trivial to generate a 12-page scientific paper, complete with graphs, using AI. A little bit of prompting and off you go. Want to write a 250-page PhD thesis? A few days of prompting and ChatGPT will get it done. Just make sure to pay for your subscription. Who wants to read all that? Sure, we can use AI to read the work but… what you learn in mathematics is that not all iterative processes converge. Adding more layers of AI increases the latency, but it may not get you where you want to be. In the world of generative AI: producing content is cheap… but producing content that is well above the quality of ChatGPT, well, that’s something else. Take this post that you are reading. I could have prompted Grok to write it. Just one request, and I get Grok to write tens of paragraphs. Done. The cost of an English post written in correct English and making reasonable statements… just went to the floor. It is now effectively free. But, you see, I write my posts myself because I think I can be clearly more insightful and deeper than ChatGPT or Grok. And similarly for software. I spend several days a week coding, and though I have been an early adopter of copilot, I mostly code by hand because I often want to maximize quality, not volume. I do enjoy using AI. For example, we recently added a C API to our simdutf library. It was boring work, mapping C++ functions to C functions, so I had this part of the code done by AI (I could also have scripted it). There is a lot of boring work in programming that I am really happy to delegate. But for high-quality work, I'll mostly write my own code. I am happy to use AI as a fancy autocomplete, but I won't rely on agents. It is not just about quality. Besides quality, I also care about understanding. To really understand what is going on, you often need to spend a lot of time thinking, trying things and so forth. To be fair, AI can help you understand what is going on better. You can interact with AI to check your understanding. But if you delegate to your AI, your understanding will be shallow. I was listening to someone from Anthropic this summer. She was saying that, unavoidably, we are going to end up with more software of lesser quality. Maybe it is already what we are getting today. So here is what I am doing. Here is what I am going to keep doing. I am going to keep typing my own words, my own code. I am no luddite and I love cool toys. But I believe that the tools should enhance my voice, not replace it.