The Company That Asks Marketers 'What Did You Build?'

The Company That Asks Marketers “What Have You Built?”

On August 28th, Andrew Ng went on a podcast and pushed back hard on the idea that AI is replacing jobs. His argument: when economists break a job down into individual tasks and analyze them, AI can only handle about 30-40% of them. The remaining 60% that stays with humans actually becomes more valuable.

But later in the same conversation, he described his own company’s hiring bar like this: “All of my marketers know how to code.” When he interviews marketers, he asks what they’ve built. If they haven’t built any software, that’s a problem.

The examples that followed were concrete. One person on the marketing team built their own Mac desktop app that scans the web for source material whenever they need to pick a topic to write about. On the finance team, an executive noticed people spending hours every week opening documents and copying numbers by hand — so he wrote a script that opens the files automatically, checks the contents, and flags anything that looks off. The recruiting team just has an engineer sitting on it, full stop.

Put the two claims side by side and it feels off. He says jobs aren’t disappearing, yet he’s demanding that marketers write software. But this isn’t a contradiction. The job itself hasn’t gone anywhere. What’s been rewritten is the ticket you need to walk through the door.

---

Let’s follow the numbers Ng used as evidence

Ng pointed to software engineering job postings as proof of his claim — that, contrary to what the doomsayers say, posting volume is actually rising. I checked it myself.

According to Indeed Hiring Lab’s August data, as of August 14th, 2026, the overall job postings index sits at 101.8 — just slightly above pre-pandemic levels — while software development postings sit at 74.4, a much lower level that hasn’t recovered¹. Since 100 marks February 2020, this one job category is still down more than a quarter from where it started.

And yet Ng isn’t wrong either. The same organization’s July analysis, setting February 2025 — when Claude Code launched — as the baseline of 100, shows software development postings up about 15%, while overall postings fell 7% over the same period. But the starting point was so low that even after the rebound, the category is still 27.5% below pre-pandemic levels.

The same database produces both “it’s growing” and “a quarter of it vanished” as true statements. What creates the difference isn’t interpretation — it’s the baseline. Measure against the last year and a half, and it looks like a recovery. Measure against the last six and a half years, and it looks like a collapse. Ng picked the former. The people he’s arguing against picked the latter.

Korea’s Numbers Swing Harder on the Same Dial

Korea’s numbers run much hotter than America’s. Data compiled by JoongAng Ilbo this past June, based on the hiring platform Catch, shows that entry-level postings from large and mid-sized IT and telecom companies in March 2026 fell 73% from a year earlier.

This 73% figure is itself a product of its baseline — that’s worth flagging up front. It’s a single month measured against the same month a year earlier, so if a company’s hiring calendar slips by even a year, the number swings wildly. The same yardstick we just used to check Ung’s figure needs to apply here too, for fairness.

Still, the numbers sitting alongside it confirm the direction. Naver didn’t run its first-half new-graduate hiring drive this year, after doing so for three consecutive years. Musinsa’s junior developer hiring round — its first in 4 years — drew roughly 2,000 applicants, of whom 66 were hired. According to each company’s ESG reports, new hires fell from 599 to 231 to 258 at Naver between 2022 and 2024, and from 870 to 452 to 314 at Kakao over the same span.

University-side indicators point the same way. Comparing computer science employment rates between 2023 and 2025, Seoul National University fell from 83.8% to 72.6%, Hanyang University from 81.4% to 70.3%, and KAIST from 77.9% to 69.8%. These figures exclude graduate-school entrants and military enlistees.

What we need to look at alongside this is tenure. Over the same period, average tenure at Naver rose from 6.9 to 7.7 years, and at Kakao from 5 years to 6 years and 3 months. Normally, when work disappears, people leave along with it — but right now, the people already inside are staying longer. What’s vanished isn’t the work — it’s the door.

Company responses sharpen this picture further. In a survey Wanted Lab released last December, polling HR managers at 153 domestic companies, 74.5% of respondents said they would maintain or expand their 2026 hiring scale. It’s not that a majority of companies are shutting down hiring. Rather, in the same survey, when asked what qualities define an ideal candidate, 64.7% cited job-specific expertise and 24.2% cited AI and data skills. Headcount is holding steady — but the bar for who clears it is shifting.

What’s Shrinking Is a Job Category — What’s Growing Is the List of Requirements

The most decisive number comes from Wanted Lab’s data. After analyzing 72,793 job postings on Wanted from January 2025 through April 2026, 5 of the 6 job categories that showed a statistically significant decline in 2026 turned out to be developer roles.

The same analysis also found a metric moving in the exact opposite direction. The phrase “AI native” appeared 8 times more often in job postings, and “physical AI” 5 times more — measured by comparing each phrase’s share of total postings against the previous year.

On one side, positions are disappearing. On the other, the requirements attached to those remaining positions are being rewritten from scratch. Kisu Jeong, Head of the AI Division at Wanted Lab, explains that AI has already taken over the kind of simple tasks and coding once handed to new hires as a way of easing them in. What Ng said in the US — “every one of our marketers codes now” — and the 8-fold jump in “AI native” mentions in Korean job postings are two sides of the same event.

Ng frames this shift as an expansion of job scope. Just as developers once split between frontend and backend have merged into full-stack roles, marketers who once only coordinated campaigns are now expected to own an entire cycle from planning to execution, and recruiters are expected to see a hire through from start to finish. What’s needed, then, isn’t just knowing how to use AI. The job knowledge required to handle that expanded scope has to come along with it.

These days I find myself spending a lot more time coding. Where I used to just write strategy documents, I’m now spending more time actually building things myself — and the numbers above show this isn’t a personal preference, it’s the market shifting under me.

Selling Fear, Selling Optimism

Let’s go back to where Ng started this interview. He pointed to PR and regulatory capture² as the root of the misinformation surrounding AI. For a company that spent billions of dollars training a model, it’s a real problem when someone else builds something comparable and releases it to the world for free. So you stoke fear, pull in regulation, and suddenly the playing field tilts toward the incumbents while the open-weight³ camp gets its hands tied.

That point is accurate on its own terms. The incentive structure really does look that way.

But you have to apply the same standard to the other side. In this same interview, Ng said AI models are terrible for learning — students who use AI score higher on assignments but retain far less afterward. And in the very next question, he introduces the new company he’s built.

Coursera invested $100 million on July 28th in LearnVector, which Ng founded and now runs as CEO. That’s roughly a one-third stake on a fully diluted basis⁴. The first product is slated for early 2027, and Ng remains chairman of Coursera’s board.

The line “AI-based learning as it’s currently done is the worst” is also, conveniently, the sentence that defines the market for his new learning company. This isn’t to say Ng lied — the data really does point that way. It’s just that if the fear-sellers have an incentive, so do the optimism-sellers. Neither side is a disinterested observer.

Look again at the advice he gave college students and new graduates in that same interview, and this structure comes into focus. University curricula can’t keep pace, he said, so learn the latest skills separately online — and the examples he named were Coursera, DeepLearning.AI, and Udemy. Two of the three are organizations he built, and Coursera and Udemy merged into a single company this past May. The advice itself is sound. But when advice and business interest point in the same direction, it’s better to know that going in.

So what this piece wants to leave you with isn’t a conclusion — it’s a method for reading. The next time you come across a number about AI and jobs, check three things. First, what’s the baseline date? The same index can look like a recovery or a collapse depending on whether you’re measuring from one year ago or six. Second, what’s the denominator — number of postings, a share, or actual hires? Third, what is the person saying this actually selling? It could be a model, a regulation, a course, or fear itself.

Oswarld's Lens

There’s a reason this story feels even more uncomfortable in Korea. The requirements have already changed, but the way we judge those requirements is still anchored to the old baseline.

The practice of scanning résumés for school names and majors, then counting certifications and internship months, can’t capture “what have you actually built.” On the flip side, applicants keep hearing that they need to build a portfolio, but there’s no agreed-upon scale for how much of what counts. In the US, someone like Eugene Ahn publishes his own interview criteria, which serves as that scale. In Korea, that role is largely vacant.

What’s filling the vacancy right now is portfolio consulting and fear-based marketing. The blurrier the standard, the better it is for whoever’s selling. If employers simply spelled out, concretely, what output they want, a good chunk of this market would sort itself out. But we’re still stuck at the slogan stage of “skills over school pedigree.” Please, don’t spend your money on this… buy yourself some beef instead.

Closing

Let me break this down by situation.

If you’re currently working, the fastest way to test yourself is to build, on your own, one tool in your job that you’ve been waiting for someone else to make. What Eung’s finance team did wasn’t anything remarkable — they just automated a document check they’d been repeating every week.

If you’re on the hiring side, your job posting shouldn’t say “AI proficiency” — it should say what kind of things you want the person to have built. Without that one line, applicants have no choice but to prepare against the old baseline.

If you’re the one reading the numbers, checking just three things — the reference date, the denominator, and what the speaker is selling — will filter out most of the exaggeration. The same standard applies whether the number leans toward fear or toward optimism. As with the 74.4 and 15% we saw today, both are often true at once — it’s usually just the baseline that differs.

Reader, in the end, what this shift is really about isn’t the number of jobs — it’s the wording on the ticket of admission.

💬 Go back and reread your organization’s job postings — is there a sentence that’s different from one written 3 years ago? Tell us in the comments which phrase is new.

📨 If you know a colleague who’s writing job postings these days, or preparing to switch jobs, pass this along. Checking just these three baselines can change how they prepare.

References & Further Reading

Primary sources

• The Singju Post, “Andrew Ng: The Biggest Opportunities in AI Aren’t Where You Think”, August 31, 2026. ··· This is the full transcript of the August 28 broadcast. The marketer-hiring criterion is at the 22:20 mark.
• Indeed Hiring Lab, “US Labor Market Snapshot — August 2026”, August 24, 2026. ··· This is where the software development index reading of 74.4 comes from. It updates monthly, so it’s worth checking the latest figure before citing it.
• Indeed Hiring Lab, “AI and Job Postings: From Destruction to Creation?”, July 8, 2026. ··· This shows both the rebound and the trough, making it the best source for understanding the baseline problem.
• JoongAng Ilbo, “IT entry-level postings down 73%, computer science graduates struggling in the AI era”, June 11, 2026. ··· Includes Wanted Lab’s analysis of 72,793 job postings and computer science employment-rate data.
• Coursera, “Coursera Makes $100 Million Strategic Investment in LearnVector”, July 28, 2026. ··· The original release detailing the investment size and equity structure.
• AI Times, “Wanted Lab releases hiring trends report”, December 8, 2025. ··· A survey of 153 HR managers. The sample is small, so treat it as directional only.

Background

• FRED, “Software Development Job Postings on Indeed in the United States”. ··· You can plot the index yourself. Try shifting the baseline around — it makes this issue’s argument click.
• Class Central, “Coursera Bets $100 Million That Andrew Ng Can Do What Coursera Can’t”, July 29, 2026. ··· A skeptical take on the same investment, questioning the valuation of a company with no product.

Past issues worth reading alongside this one

• The People Selling Fear to Job Seekers ··· This issue looked at the distribution structure that turns fear into a product. This one applies the same yardstick to the speaker on the other side.
• Ferrari Designers Still Start With a Pencil ··· A story about the ladder new hires used to climb while learning the job. Read alongside this issue’s “ticket of admission,” the picture comes together.
• The Thinking Muscle That Collapses in 10 Minutes ··· This issue laid out the research behind Ng’s claim that it’s “the worst thing for learning.”

The author is Oswarld (Kwangseob Ahn). Current roles: Adjunct Professor at Sejong University, Strategy Consultant at INLEVEL9. Career, research, books, and recent work are kept current on the About page. Latest · July 2026: HEMA-2: A Consolidation-Aware Tri-Memory Architecture with Multi-Channel Scheduling for Lifelong Conversational AI.

📝 Glossary

Footnotes

1. Job Postings Index: An index Indeed built by setting the number of postings on February 1, 2020 to 100 and tracking change from there. Since it’s a ratio to that baseline rather than an absolute count, the same curve reads differently depending on where you set the starting point. ↩

2. Regulatory capture: The phenomenon where regulators end up acting in the interest of the parties they’re supposed to regulate. A classic case is when rules made in the name of safety end up protecting incumbents’ turf instead. ↩

3. Open weights: A method of releasing a model’s weight files so anyone can download and run them. This differs in scope from open source, which also opens up training data and code. ↩

4. Fully diluted: A way of calculating ownership stakes that assumes all rights not yet converted into shares — like stock options — have already become stock. This produces a more conservative figure than counting only shares actually issued. ↩