What This Is
10,000 AI job postings plus dozens of expert interviews. Andrew Ng (former Baidu chief scientist, founder of DeepLearning.AI) and his team ran a clustering analysis (auto-grouping by similarity) and landed on a judgment most ordinary developers won't love: the ability to write code is becoming less valuable; what's scarce is the ability to choose direction.
The output is called the "AI Engineering Skills Map," which distills four skills: building and deploying AI applications, software engineering fundamentals, working effectively with coding agents (AI programs that autonomously complete multi-step tasks), and shaping what to build—deciding what to do, and why.
What's worth paying attention to is the method, not the conclusion itself. Ten thousand postings capture real hiring demand across today's companies—closer to market temperature than any expert opinion.
Industry View
Supporters argue these four skills hit the mark. The biggest difference between AI applications and traditional software is unpredictable output: feed the same prompt (an instruction given to the AI) in, and the result may differ every time. A real AI engineer doesn't just understand the building blocks—LLMs (large language models), RAG (letting the model retrieve information before answering), agentic workflows (letting AI autonomously complete multi-step tasks)—they must also use statistical methods to measure, steer, and govern the system. Ng keeps coming back to one word: evals (evaluations)—the disciplined practice of running tests, surfacing bad cases, fixing them, and re-testing.
We note that the main objections come in two flavors. First, this framework assumes the reader is an engineer; it barely touches product managers, operations, sales, and other roles that are equally critical in enterprise AI adoption. Second, "shaping what to build" sounds like the old cliché of "product sense," with no operational standard. How do you define a "good direction"? Ng himself hasn't offered metrics.
There's a hidden risk worth flagging: all four skills reward individual capability, with nothing on team structure, process governance, or cross-functional collaboration. That means many companies may misread the map as "just hire someone who can do all four," overlooking the organizational overhaul required.
Impact on Regular People
For enterprise IT: Over the next 12–18 months, AI engineer hiring criteria will shift wholesale from "can you tune a model" to "can you run evals and collaborate with agents." We recommend HR update job descriptions (JDs) in parallel and bake evals capability into interview questions.
For individual careers: Non-technical roles shouldn't sit this one out. The fourth skill—"shaping what to build"—essentially requires every person to independently judge "is this worth doing" with AI assistance. Coding work is being eaten by agents; problem-definition work is gaining value.
For consumer markets: Many "AI applications" will face a shakeout. Products without evals and error-analysis loops will break when the underlying model updates; stable products will pull ahead on user experience. What ordinary users will perceive: the frequency of "sometimes works, sometimes doesn't" in AI tools will increasingly diverge across products.