Let me tell you something that will piss off half the people reading this. AI coding assistants are not tools. They are crutches. And crutches atrophy your legs. This morning I was scrolling through GitHub, and I saw a blog post where engineers admitted that Copilot actually made their code reviews worse. They had better tools, but the quality of human judgment went down. The machines spat out more code, so humans reviewed less carefully. The net result? More bugs. More shit. Everyone says "but Steve, it's just a tool, like a compiler." That's bullshit. A compiler translates. It doesn't pretend to think for you. These AI assistants are designed to make you feel productive while you slowly lose the ability to hold a system in your head. I found out the hard way that when you outsource thinking to a black box, you stop thinking. Here's the thing. I built the Mac with a team of maybe 100 people. We didn't have autocomplete. We had Andy Hertzfeld writing assembly by hand. We had Bill Atkinson writing QuickDraw in 68000 assembly because Pascal was too slow. These people held the entire machine in their heads. They knew every register, every interrupt, every pixel. Do you think Copilot would have written QuickDraw? No. It would have written a generic rectangle drawing function. It would have been fine. It would have been mediocre. And the Mac would have shipped two years late with twice the bugs. Honestly, this is the same mistake the PC industry made when they licensed Windows to every clown with a soldering iron. You trade control for speed. You trade craft for convenience. And eventually you forget what "insanely great" even means. I read that Anthropic is supposedly building hardware with Physical Intelligence. Probably bullshit. But if they are, I hope Claude doesn't write the firmware. Because firmware written by an LLM is firmware nobody truly understands. And when it breaks at 3 AM in a self-driving truck from Einride, good luck debugging a hallucinated memory leak. The truth is, these tools are great for boilerplate. For CRUD apps. For the 90% of software that is just moving data from one box to another. But the 10% that matters — the thing that makes your product sing — that comes from a human who has suffered through the problem. Who has connected the dots backwards. I remember reading about the Apollo Guidance Computer code. Margaret Hamilton wrote it by hand. It flew humans to the moon. Not a single AI agent reviewed it. Not one "vibe coded" function. And it worked. Because the person writing it understood every single line. Now look at what's happening. Vibe coding is literally a movement now. Codeberg banned it. Codeberg! The most permissive platform on earth, and even they said "enough." Codeberg bans vibe coded projects. That's like a strip club banning nudity. It tells you everything. Meta is testing an AI bedtime story app for people with no imagination. That's not a joke. That's real. We are building tools to replace the human act of making things up. And somehow we think this is progress. I'm not saying don't use these tools. I'm saying use them like a sharp knife, not like a prosthetic brain. If you don't know what the code is doing, you don't know what your product is doing. And if you don't know what your product is doing, you shouldn't be building it. A player ship code. B players ship features. C players ship tickets. And D players let Claude ship for them. Here's my rule. If you can't rewrite the function from memory after you've committed it, you don't understand it. If you don't understand it, you can't debug it. If you can't debug it, you don't own it. Stop treating AI like a junior developer. It's not a developer at all. It's a statistical parrot with a good PR team. The best products I ever made came from someone staring at a whiteboard for three hours, frustrated, angry, refusing to accept that "this is just how it works." That frustration is the engine of great software. If you let an LLM skip that frustration, you skip the insight. So no. AI coding assistants are not tools. They are dependency injectors for your brain. And dependency injection is fine for code. Terrible for thinking. Now go write something yourself. And don't let me see you using autocomplete on the hard parts. FAQ Q1: Are you saying all AI coding tools are bad? No. I'm saying most people use them the wrong way. They offload thinking instead of offloading typing. The distinction is everything. Use them for boilerplate, for tests, for documentation. But if you've never written a memory allocator by hand, you shouldn't be using an AI to write one for production. Q2: Didn't you once say "computers are bicycles for the mind"? Isn't an AI assistant just another bicycle? A bicycle amplifies your legs. It doesn't replace them. A calculator amplifies your arithmetic. It doesn't make you forget how to add. But an AI coding assistant that writes entire functions without your judgment is like a bicycle that decides where to go. That's not a bicycle. That's a bus. And you're not the driver. Q3: You used a GitHub blog post about Copilot making code review worse. Is there data on this? Yes. The GitHub blog post they linked shows that adding better Copilot tools actually degraded code review quality because reviewers trusted the AI output too much. They found fewer bugs per line of AI-written code than per line of human-written code. More tools, worse outcomes. Classic unintended consequence. (Source: The GitHub Blog, 2026.)