On August 21, 2026, Andrew Ng — co-founder of DeepLearning.AI and Coursera — published a post on X that made a lot of AI engineers stop scrolling. Within about 48 hours it had crossed 1.1 million views, and the ratio that matters is telling: 18,451 bookmarks against 9,599 likes. When people save a post nearly twice as often as they like it, that's reference material, not clickbait.
The post is part two of Ng's "AI Engineering Skills Map" — a map built from analyzing more than 10,000 AI job postings, structured expert interviews, recruiter and hiring-manager surveys, and online data synthesis. Ng described the process as "like clustering over a massive dataset." Part one (August 14) laid out four high-level skills. Part two drills into skill number one — "Building and deploying AI applications" — and breaks it into six sub-skills.
Here's the twist worth staying for: the sub-skill Ng calls the most important is not prompt engineering. It's not RAG. It's not agentic workflows either. It's something else entirely, and we'll get there shortly.
Why "AI engineering" Is Bigger Than the Job Title "AI Engineer"
Before the six sub-skills, one point from Ng deserves its own paragraph: these are not skills reserved for people with "AI Engineer" on their business card. His analogy is cloud computing. Nearly every developer today needs cloud skills, yet very few carry the title "Cloud Engineer." AI is walking the same path.
That's true globally, and even sharper in fast-growing markets like Vietnam. Our own market research (aggregated in the 5ac.vn research note from Leos.vn and ITviec 2024 data) shows the country has more than 15,000 open AI positions but only around 8,000 qualified candidates. And 65% of hiring managers report struggling to hire because candidates routinely overstate their AI proficiency. In other words: the gap isn't a shortage of people studying AI. It's a shortage of people studying the right things. Ng's map is a direct answer to that gap.
The Starting Point: AI Output Doesn't Behave
Ng's entire framework rests on one central claim: the fundamental difference between traditional software and AI applications is that AI output is far less predictable. The same input can produce different output on every run, and quality depends on the model, the data, the prompt, even sampling parameters.
With traditional software you write code, run tests, watch them pass, ship. With AI, you have to use statistical techniques to measure, steer, and govern output quality. That forces an iterative loop: build → inspect output → decide what to do next → build again. Nobody nails the first prompt and deploys.
And to govern that uncertainty, you need a measurement mechanism. That's why evals sit at number one.
The Six Sub-Skills: Building & Deploying AI Applications
1. LLM Foundations — the ground floor
Understanding how LLMs work well enough to make sound technical decisions: tokenization, how output generation actually happens, context-window tradeoffs, cache hits, knowledge cutoffs, reasoning-effort levels, tool calling, choosing multimodal models.
The real payoff here is decision-making: which model to pick, which models to combine, whether to fine-tune, self-host, or just call an API. These decisions land directly on your monthly operating bill. For small businesses running lean AI budgets, picking wrong at the start means burning money every day after.
2. Grounding Models with Data — beyond basic RAG
RAG is where this starts, not where it ends. Ng's framework puts RAG alongside two alternatives: Knowledge Graphs and Semantic Layers for structured data. You also have to decide whether to inject data directly into the prompt or fetch it on demand via tools, and build pipelines that turn messy documents (text, PDF, HTML, images) into LLM-ready input.
One warning worth underlining: grounding fails silently when your data goes stale. No exception thrown, no red log line. The chatbot keeps answering — it's just answering wrong, and your customers discover it before you do.
3. Building Agentic Systems
Ng describes agency as a spectrum. At one end: workflows, fixed chains of LLM calls that are easy to control. At the other: agent harnesses, where the LLM decides its own next step — flexible but unpredictable. Along that spectrum sits a stack of architecture decisions: chain or parallelize, which steps run as code versus LLM calls, what fallbacks look like, which tools to expose (MCP, CLI, sandbox), memory design, long-context management, single-agent versus multi-agent.
Because agentic systems run close to real data, production hardening is mandatory: guardrails, defense against adversarial input, leak prevention, governance. This is exactly where many teams stumble moving from prototype to a system real customers depend on.
4. Evaluation-Driven Development ⭐ — "the most important trait"
This is the only sub-skill Ng explicitly labels most important, and it separates teams that ship from teams that iterate blindly. What it covers:
- Evals plus an error-analysis loop: a disciplined cycle of measuring and dissecting failures, instead of tweaking prompts by gut feel.
- Exploratory data analysis on traces and outputs before defining metrics — don't measure what you don't yet understand.
- Deciding what to measure: this requires product and business insight, not just engineering judgment.
- Choosing the eval type: code-based/deterministic checks, LLM-as-judge, or human-in-the-loop review.
- Meta-evaluation: auditing your own eval suite. Evals are living artifacts, not one-time deliverables.
- Failure-driven development: capture real errors from production, add them to the test suite, block regressions.
In the LinkedIn discussion thread on Ng's post (over 378 comments), plenty of senior engineers confirmed this from experience: eval design is a core technical skill, not compliance paperwork. One recurring theme in the replies: evals are precisely where teams fall apart.
5. Operating in Production
Deployment is where the real fight begins: observability across four axes (performance, quality, cost, latency), statistical regression testing, risk-calibrated CI/CD, incident response for model failures and prompt injection, and cost/latency optimization at scale through model selection, distillation, fine-tuning, and workflow simplification.
6. Machine Learning Foundations
It sounds paradoxical, but even if you only ever call APIs, understanding what's inside an LLM pays off. Mental models like bias/variance, error analysis, and data engineering transfer directly — they help you reason correctly when outputs get fuzzy, and weigh accuracy against training/inference speed when choosing a model.
Investment Order: What Should You Learn First?
A practical insight from the framework's dependency structure (systematized in explainx.ai's independent analysis): the six sub-skills are not equal peers. A sensible order:
- LLM Foundations first — you can't measure what you don't understand.
- Evals immediately after — because, in Ng's words, they're the mechanism that makes every other skill manageable.
- Grounding, Agentic, Production, ML foundations — expand once the eval loop is running.
Translated for a working team: don't pour effort into a flashy multi-agent pipeline before you have even a minimal eval set. Without evals, every change is a guess — and you can't tell whether a change helped or hurt.
What This Map Means for Engineers in Emerging Markets
Mapping Ng's framework against hiring signals in Vietnam (aggregated in our research note):
| Sub-skill | Vietnamese market signal |
|---|---|
| Grounding / RAG | RAG ranks among the most sought-after skills |
| Agentic Systems | LangGraph and CrewAI appear in job requirements |
| Production Ops | MLOps (Docker, FastAPI, cloud) is a common requirement |
| ML Foundations | Vietnamese LLMs (PhoBERT, ViGPT) create baseline ML demand |
On compensation: junior AI engineers earn 12–25 million VND/month (averaging 15–20 million), mid-level engineers 50–85 million, and senior engineers 85–125+ million. Verified AI capability carries a 20–30% premium over non-AI roles (Reeracoen Vietnam Q1 2026). The largest hirers include VinAI, VNG, MoMo, MB Bank, TPBank, FPT, VinBigdata, and Shopee VN.
The numbers that stay with you: 15,000 open roles, 8,000 qualified candidates, and 65% of employers complaining that applicants inflate their skills. Against that backdrop, a skills map published by Andrew Ng himself — built from 10,000+ real job postings — comes close to being an answer key. Follow the map and you don't have to guess what you're missing; walk into an interview and you share a common language for proving competence instead of claiming it.
If you're a founder hiring AI engineers rather than becoming one: use the six sub-skills as your interview checklist. Ask candidates to describe an eval loop they've actually built, or a production AI failure they've actually handled. Whoever answers concretely is worth more than every "proficient in AI" bullet on their CV.
Takeaway
Three things to carry out of this piece:
- AI engineering ≠ the AI Engineer title. Full-stack, data, and DevOps engineers all need these skills — just like cloud before it.
- Evals are skill number one, ahead of prompting and RAG. Invest in the eval loop before scaling system complexity.
- Markets are short on the right skills, not on learners. 15k jobs against 8k qualified candidates in Vietnam alone. Ng's map is a self-study roadmap with real data behind it.
At 5ac.vn, we build G-Company OS — an agentic operating system for small and mid-sized businesses — and Ng's framework mirrors almost line-for-line what we solve daily: clean data grounding, guarded agent harnesses, and evals underneath every change. If you want to start your company's AI engineering journey without a large team, follow this series on the 5ac.vn blog. Next up: building your first eval loop for a real Vietnamese-language chatbot.
Sources: Andrew Ng (@AndrewYNg) on X, Aug 14 & Aug 21, 2026; DeepLearning.AI The Batch Issues #366, #367; explainx.ai analyses; Andrew Ng LinkedIn post; Vietnamese labor market data from Leos.vn, ITviec, Reeracoen Vietnam, Robert Walters, xDev Asia. All figures traceable to research note t_fc42c7dc.