Eight years ago I started an academy to teach people to program. Watching what AI has done to the field since, I've landed on a single sentence that explains most of it: AI moved the syntax dial all the way. It didn't move the architecture dial at all. The syntactical work compressed to almost nothing. The architectural thinking — the judgment about why a system is built the way it is — is exactly as hard as it always was. That gap reshapes every path into this career.
The four paths in 2026
The CS degree — still viable, higher bar. Enrollment growth basically flatlined at 0.2% (2023–24), the smallest since 2006, and recent CS grads face 6.1% unemployment against 3% overall. But deep systems work — distributed systems, compilers — still demands real fundamentals. The degree went from default safety to a deliberate choice.
The bootcamp — broken. Placement collapsed from 79% in 2022 to 58% in 2023, and time-to-employment doubled from three months to six. Codeup, BloomTech, V School, Galvanize — closed. The "learn syntax, build a portfolio, get hired" model was built on exactly the work AI now does for free.
Vibe coding — mixed. Great for prototypes and low-stakes tools. It falls apart the moment a system has to scale, stay secure, or evolve. YC's Outlit needed full rewrites by experienced engineers after the AI-generated codebase couldn't scale. It works — if you already have the judgment to know when it's producing something fragile.
The emerging path — the promising one. A programmer's mindset plus AI as a pair programmer, not an oracle. You take the time AI saves and pour it into the irreplaceable work: architecture, naming, boundaries, edge cases, tradeoffs. Carnegie Mellon's HCII found pair-programming with AI outperforms treating it as an answer machine. Judgment stays in the loop.
The evidence keeps pointing the same way
The METR study (July 2025) is the one that stuck with me: developers predicted 24% faster, actually worked 19% slower, and still believed they were 20% faster. The perception gap is enormous. GitClear found copied code overtook refactored code for the first time — syntax accelerated, architecture didn't. And engineers with 15+ years of experience report just 38% satisfaction with AI tools versus 60% for those under five years. The people who most understand architecture are the least impressed by a tool that only speeds up syntax.
The part that can't be outsourced
Judgment — knowing when a system is fragile, making the architectural call, understanding the tradeoff — doesn't come from a model. I think about the SaaS founder whose AI-built app got hacked, database wiped, paywall bypassed. The problem was never the syntax. It was the architecture.
So two questions, if you're deciding how to enter or grow in this field:
- Do you want to understand why systems work, or just build things that appear to work?
- Where will your judgment come from in five years if you skip the foundations today?
My first cohort learned syntax that's mostly obsolete now. What they kept was a way of thinking about problems that AI hasn't touched. That's the thing worth building. It always was.