Stanford's Chris Piech runs Code in Place, an intro programming class with 17,000 students and over 1,000 teachers. They've run it both before and after Cursor and Claude Code arrived, and enrollment basically doubled. It turns out that more people, not less, want to learn to code now that AI can write it.Why learn the thing the machine can now do almost flawlessly? Six reasons out of an interview with him.1. Syntax is the disposable half; problem-solving is the durable onePiech splits learning to program into two parts: the syntax (how you tell a computer to do things) and the problem-solving (how you break a big problem into small pieces and set up data to talk to algorithms). AI is very good at syntax. There is less need to memorize it.The half worth building is the one that transfers: problem solving. Decomposition, naming the sub-problems, deciding what data structure the algorithm needs. That is the problem solving part you carry into every language and every tool, including AI. The syntax is disposable; the problem-solving is durable.2. Without foundations, you can't catch AI's bad decisionsPiech says he constantly ships with AI, and he's explicit about why it works for him: he already knows architecture. Strip that away and the model starts making poor structural calls you often only notice later on in the app's lifecycle.I see this all the time working with LLMs: happy path and call it done. The edge cases show up when real users start hitting it, or when you step back from the first prototype and look hard at what you actually built. At that point you really need to understand the architecture underneath.This is the whole argument for keeping fundamentals. AI is an accelerator, not a compass. It can get you to a solution faster, but it can't tell you if the solution is good.3. Outsource too much and the muscle atrophies"If you have AI write too much of your code, at what point can you no longer do that valuable piece of the architecture?"And the insidious thing is that this will happen quietly. Using AI is fine, Piech says, as long as you stay self-aware about whether you're growing alongside it or handing away the growth. I made the same case from a different angle in Guardrails Protect Your Codebase. What Protects Your Judgment?: tooling can protect the coding part (e.g. with harnesses), but only making the hard decisions yourself protects your skill and judgment.4. Coding is the best problem-solving gym you haveCode gives you immediate, falsifiable feedback. Your logic is wrong, the thing breaks, you see it, you iterate, you learn.Apply problem-solving to life and the feedback cycle is much slower: you make a decision, and the consequences show up months later, often entangled in a complex web of other decisions.Coding on the other hand is the rare domain where you can run the loop hundreds of times a week. (Deliberate practice is how you get the reps.) That's what trains judgment, and judgment is the skill AI can't hand you.5. The high-order skill is knowing what's worth buildingPiech frames the durable skill as interfacing between what computers can do and what humans actually need. What's the valuable problem? What feature helps a user make progress?It has always been a critical high-order skill. What's changed is that more junior engineers can engage with it now, and if you're a junior his advice is to start on it today. You don't need the senior title first.It's the same reason I keep arguing design beats code: I once reduced a 1,069-line AI-built app down to 156 lines because I put design before code.6. You can now learn faster than everThis is the flip side of reason 3: used well, AI can be an extraordinarily powerful tutor.Piech's own move if he were starting today: build a lot of prototypes with Claude Code, then ask it to teach him the most important concepts behind each one, and iterate.Foundations first, then learn to code with AI, in that order. That's why I'm building a Python Foundations course: the workflow layer around the language, learned by shipping one real tool. More soon.The tool that can erode your skills can also compound them, depending entirely on whether you let it think for you or make it teach you.None of this depends on where AI lands in 2030. Self-driving cars looked close in 2012, and their progress was badly overestimated. It's the last 1% that resists the machine: the odd thing on the highway, the judgment call no one scripted. Coding is no different. Decomposition, judgment, taste, knowing what to build; those turn out to be the most human-dependent parts of the job and the ones that will be most valuable in the AI era.