I’ve been noticing something weird lately. Not with AI itself. AI is obviously getting better: the models are getting better, the tooling is getting better, and things that sounded completely ridiculous two years ago are now normal things to do on a Tuesday. I’m talking about the people talking about it. Spend enough time on Tech Twitter and you start seeing the same ideas over and over again, just with different words. Someone discovers that an LLM can call a tool: it’s an agent. Someone puts two agents together: it’s multi-agent orchestration. Someone gives the agent memory: it’s cognitive architecture. Someone lets it modify its own files: it’s autonomous software engineering. Then somebody draws a diagram with twelve boxes and suddenly we’ve apparently invented computers again. The funny thing is that most of these things are genuinely useful. The problem isn’t the technology. It’s the mythology around the technology. Understanding is still the work There seems to be a growing belief that because AI can do something, understanding that thing is no longer necessary. I’ve seen people genuinely argue that you don’t need to learn programming anymore because AI can write code; you don’t need to understand databases because an agent can work with PostgreSQL; you don’t need to understand frontend development because AI can generate a UI; and you don’t need to understand infrastructure because an agent can deploy it. At some point, you don’t need to understand anything because there is an API for everything. I’m not sure what we’re optimizing for at that point. If the goal is simply to produce software, then sure, this is fantastic. But if the goal is to become someone who can build software, I don’t think skipping all the fundamentals gets you there. There is a difference between not doing something and not knowing how something works. I don’t manually manage memory in every program I write, but that doesn’t mean I don’t want to understand memory. I don’t write HTTP servers from scratch every time, but that doesn’t mean I don’t want to understand HTTP. I don’t implement a database engine before building an application. That would be insane. Abstractions are good. That’s what software engineering has always been about: we build things on top of other things so that we don’t have to think about everything simultaneously. The problem starts when the abstraction becomes so opaque that you don’t know what to do when it breaks. And everything breaks. AI is particularly good at hiding this problem. You can ask it to build something and get back a surprisingly good application. Sometimes it is actually excellent. The first time you experience this, it feels like magic: you describe what you want, and a few minutes later there is a working application sitting in front of you. It’s easy to make the next mental jump: “If it can do this, maybe I don’t need to know how any of it works.” That’s the trap. Because the application working once isn’t the same thing as the system being understood. The difficult part of software usually isn’t getting the happy path to work. It’s everything around it: what happens when two requests arrive at the same time, when the network disappears, when the database is slow, when the user has ten million records, or when your third-party API changes its response? What happens when the model confidently invents an API that doesn’t exist, or when your bill goes from $40 to $4,000? The machine is very good at producing the first version. I’m still not convinced we’ve solved the second part. What the internet rewards And then there’s the startup problem. Open enough AI startup websites and they start blending together: some sort of black background, a large statement about the future, and a chat interface. “Your autonomous AI workforce.” “AI employees for every business.” “Agents that actually get work done.” Maybe there’s a little animation showing several colorful circles talking to each other. And then the pricing page: $20 per seat. Enterprise: Contact us. I don’t think these companies are necessarily bad. Some of them will probably become enormous. But the sheer number of companies doing roughly the same thing is interesting. AI has made software incredibly cheap to produce. That sounds entirely positive until you realize that everyone now has the same advantage. If everyone can build a product in a weekend, the product itself becomes less interesting. The difficult question becomes: Why this product? And that’s a much harder question to answer with an LLM. This is where I think social media makes the problem worse. The internet doesn’t reward understanding very well; it rewards things that are easy to understand quickly. A detailed explanation of why a distributed system fails under a particular workload isn’t going to outperform: “SOFTWARE ENGINEERS ARE FINISHED.” The second one gets clicked. People argue about it, someone screenshots it, someone makes a response video, and then someone else writes a thread explaining why the original person is wrong. Now there are twelve pieces of content where previously there was one bad opinion. The machine is fed. Everyone wins—except the person who wanted to learn something. I’ve started noticing that the people I actually learn the most from rarely make the biggest claims. They’ll say, “I tried this. It worked surprisingly well. Here’s where it broke. Here’s the benchmark. Here’s the repository. Here’s what I’d do differently.” That’s it. No prediction that an entire industry is dead, no grand theory about the future of human intelligence, and no diagram explaining why their three-agent startup is the next operating system. Just evidence. It’s almost boring, which is probably why it doesn’t do very well online. Use the machine, keep the curiosity I also don’t think the answer is to become anti-AI. That seems equally stupid. I’m using AI constantly. It’s one of the best tools I’ve ever had as a developer. I can throw an unfamiliar codebase at it and get a decent explanation. I can prototype something in an afternoon that would have taken me days before. I can use it to write repetitive code, investigate errors, explore an API, generate tests, and get unstuck when I’m staring at something stupid for too long. That’s incredible. I don’t want to go back. I just don’t want to confuse using a tool with outsourcing your understanding to the tool. Those are very different things. Maybe the real shift isn’t that programming is disappearing. Maybe programming is becoming less about writing code and more about knowing what the code means. That would actually be a pretty good future. Let the machine write the boring stuff. Let it generate the first draft. Let it handle the repetitive parts. But keep the ability to open the hood: read the code, read the logs, look at the database, understand the network, question the output, and throw the whole thing away when it’s built on the wrong assumption. That’s still engineering. And I suspect it will remain engineering for quite a while. The thing I find most uncomfortable about the current AI conversation is that we’re starting to treat curiosity as inefficient. Why learn how something works if an agent can figure it out? Why read the documentation if you can ask the model? Why understand the system if the model can operate it? Because sometimes I want to know. That’s reason enough. You don’t learn how computers work because you expect to manually flip bits for the rest of your life. You learn because understanding the machine changes what you’re capable of doing with it. That’s the part I don’t want to lose. AI can make me faster. It can make me more productive. It can probably make me a better programmer. But if using it eventually makes me less curious, less willing to understand things, and more comfortable accepting answers I can’t verify, then we’ve made a strange trade. We’ve made the computer more capable while making the person operating it less capable. I don’t think that’s progress. I think the interesting future is somewhere else: use the machine, let it do more, and learn more anyway. And when somebody tells you that you don’t need to understand something anymore because AI can do it for you, the correct response might simply be: Cool. Now show me how it works.