This isn't the late-90s dot-com wipeout. It's more surgical. Entry-level CRUD apps, basic API glue, and routine frontend components are getting folded into AI-assisted workflows or shipped by offshore teams leveraging Copilot-level tooling. What's left paying premium rates are specialists: people who can ship production-grade vector pipelines, audit LLM security flaws, or architect agentic systems that don't hallucinate themselves into compliance disasters.
I've watched three mid-level friends get pushed out of their "full-stack generalist" roles in the last eight months, not because the company failed, but because one senior engineer armed with Claude Code and a few custom prompts could maintain what three people used to. The survivor bias in hiring right now favors T-shaped folks who can touch every layer but specialize hard on something AI can't yet replicate reliably — prompt engineering for domain-specific reasoning, real-world data curation for fine-tuning, or deployment pipelines for quantized models on edge hardware.
The counterargument is strong: software is eating everything, so demand should only grow. And sure, every company still needs people who understand distributed systems, database sharding, and the gap between staging and production. But "need" is different from "hire a body to fill a seat." The jobs that remain require sharper tooling fluency — you're not just writing Python, you're wrangling LangChain orchestrations, debugging why your RAG index returns garbage, or tuning LoRA adapters for your vertical.
The pivot is happening faster than most developers admit. The safe play isn't learning "AI stuff" in the abstract — it's picking a workflow, shipping something that actually runs in production, and building the war stories that prove you can ship under the new stack. The middle of the market isn't dying overnight, but it's getting thinner every sprint.
Next Why We Can't Just Talk to AI Like a Normal Person →