The framing: AI is creating jobs, not just deleting them
The doom narrative is easier to write, but it's incomplete. Ng's argument — laid out across his recent posts and his DeepLearning.AI guidance — is that the same wave compressing the middle of the org chart is also creating brand-new job categories that did not exist in any meaningful form three years ago. Two of them are worth understanding right now if you're considering a switch.AI Forward Deployed Engineer (FDE)
An engineer embedded inside a client company, customizing AI solutions for that specific organization — building and tuning agentic workflows, picking the right models, integrating with internal data.
AI Engineer
Builds software using AI components — LLM prompts, agentic frameworks, evals — and works fluently with AI coding agents (Claude Code, Codex, etc.). Ng's bet: this will be by far the larger category over time.
Role 1 — The AI Forward Deployed Engineer
This is the buzzy one. OpenAI, Anthropic, and the post-Series-D applied-AI startups are all building teams around it. The pattern was actually pioneered by Palantir about 20 years ago, who sent engineers onsite into government and enterprise customers and let them write code against secure internal data and workflows. The modern FDE is the AI-era version: parachute in, understand what the client actually needs, ship a working agentic system around their data.What's actually required
It's not a pure-code job. Ng has been explicit about this — FDEs need real communication and business skills. They:- · Talk to clients to figure out what problem is actually worth solving
- · Prioritize projects when the customer wants ten things and only two ship
- · Explain complex tech in plain language — to a CFO, not a CTO
- · Push back diplomatically when the client asks for something that won't work
- · Actually build and ship the agentic workflow at the end of all that
Why this matters for career switchers
If you're coming from consulting, account management, technical PM, customer engineering, or solutions architecture — that hybrid profile is a real edge over a pure-IC engineer who hates client meetings. The role explicitly rewards the soft skills the rest of engineering has long deprioritized.Perspective AI's 2026 report on 1,200 FDEs found median mid-level FDE total comp at $385K, staff-level at $610K, and principal FDEs at frontier labs clearing $1.2M. Palantir's FDSE role still anchors the low end at $215K median. Equity is now 55–70% of comp at the top of the market, up from 35–45% in 2024.
Role 2 — The AI Engineer (Ng's bigger bet)
Here's Ng's slightly contrarian take: as hyped as FDEs are right now, he expects far more AI Engineer roles to emerge. The logic is structural.Companies will accept a handful of FDEs — but want many of their own.
Vendor-neutrality is now an executive concern.
The demand signal is already in the data.
What the role actually involves
AI Engineers build software using AI components — they don't necessarily train models from scratch. The day-to-day looks more like:- ✓ Designing prompts and agentic workflows that hit a quality bar
- ✓ Writing and maintaining evals — the part that catches the AI being confidently wrong
- ✓ Working alongside AI coding agents (Claude Code, Codex, Cursor) instead of just shipping all the code by hand
- ✓ Stitching LLMs into existing product surfaces — including handling failures, costs, latency, and edge cases
"Skilled AI Engineers are in very high demand right now." The market is mostly hiring generalists. The specialization wave hasn't fully landed yet — which is exactly what makes this the right moment to enter.
The part to bet on: this category will fragment, like "Software Engineer" did
Ng's most useful prediction: "AI Engineer" today is roughly where "Software Engineer" was in the 1990s — a single title that will, over the next decade, fragment into many specialized roles. His guesses for what's coming:AI FDEs
Embedded client-side engineers customizing AI for a single organization.
LLMOps Engineers
Production reliability, monitoring, model lifecycle for AI systems at scale.
Evals Engineers
Specialists who build the test harnesses that catch silent AI failures.
AI Data Engineers
Pipelines optimized for training and retrieval, not just analytics dashboards.
Harness Engineers
The infrastructure underneath AI coding agents and autonomous workflows.
Roles not yet named
Probably the largest category by 2028. Nobody has the full map yet.
How to actually break in — by background
You don't need a CS PhD. The shortest path depends on what you're walking in with.| Background | Highest-leverage on-ramp |
|---|---|
| Software engineer | Add evals + agentic workflows + at least one shipped LLM feature. Internal transfer beats external job-hunt. |
| Data scientist / ML | Shift from training models to using foundation models. Build something agentic end-to-end you can demo. |
| Consultant / SA / TPM | Aim at FDE roles. Your client-handling skill is rarer than coding skill — and harder to fake. |
| Product manager | "AI PM" is now a real specialty. Lean into evals + agentic design — the line between AI Engineer and AI PM is thinner than you think. |
| Domain expert (law, medicine, finance, ops) | Pair with an AI Engineer. You bring the validation set the AI Engineer can't generate alone. This pair-up is its own growth role. |
The Ng takeaway
If you're eyeing a switch, the takeaway isn't "learn to code or die."
It's that there are multiple on-ramps into AI work — and several of them reward the soft skills and domain knowledge you already have.
---
Data sources: Andrew Ng's public writing and DeepLearning.AI guidance; Perspective AI's 2026 Forward Deployed Engineering Compensation Report (1,200 FDEs); LinkedIn / World Economic Forum AI-skill demand data; AI Engineer salary benchmarks (Coursera, Kore1, Axiom).