🎉 Giảm 20% cho đơn hàng đầu tiên!✦🚚 Miễn phí vận chuyển cho đơn trên 500.000₫✦⚡ Flash Sale — Số lượng có hạn!✦🎁 Quà tặng kèm cho đơn từ 1.000.000₫✦🎉 Giảm 20% cho đơn hàng đầu tiên!✦🚚 Miễn phí vận chuyển cho đơn trên 500.000₫✦⚡ Flash Sale — Số lượng có hạn!✦🎁 Quà tặng kèm cho đơn từ 1.000.000₫✦
Data Scientists at Big Tech: Complex Models or Valid Experiments — Who Gets Paid More?
Quay lại Blog

Data Scientists at Big Tech: Complex Models or Valid Experiments — Who Gets Paid More?

18 tháng 11, 20269 phút đọcVũ Khôi
English
#English#Học tập & Nghề nghiệp#Dữ liệu & Database
Vũ Khôi

Vũ Khôi

Tác giả tại Wotech Electronics Blog

Photo of a laptop running a Jupyter Notebook with a linear regression chart, next to handwritten notes on linear models

Data Scientists at Big Tech: Complex Models or Valid Experiments — Who Gets Paid More?

Budget week in Singapore, a product director has two pages on the desk: page 1 is a churn-prediction model using the newest architecture, 94% accurate; page 2 is a two-column table from a 3-week experiment showing that a feature the team built for 6 months did not increase orders at all. There is money for one direction only. An essay by Harry Vũ in the DoJobNgon group (18/11/2026) — based on the author's observations, self-reported data on levels.fyi (Meta, US and Singapore, updated November 2026), and an interview with Prof. Stefan Thomke (Harvard Business School, author of "Experimentation Works", via HBS Working Knowledge) — argues that companies buy certainty for expensive decisions, not models. The compensation figures are self-reported data, not independently verified by Wotech.

Quick answer: According to the essay, companies buy certainty for expensive decisions, not models. At Meta in the US, per self-reported levels.fyi data, at the same second rung the software engineering ladder (including ML model builders) has a median total compensation of about $298,000/year versus about $254,000/year on the Data Scientist ladder; but within the Data Scientist ladder, rung 3 is about $108,000 above rung 2 — over twice the cross-ladder gap. What drives promotion is the size of the decisions your numbers are trusted to close, not model complexity. Prof. Thomke, per the essay, says that on average only about 10% of experiments improve profit — 9 out of 10 ideas do not add profit, and unmeasured, they would still ship, cost maintenance, and be reported as successes. The turning point is rung 3: from here, the Data Scientist chooses the question, owns a product area's success metrics, and their number opens or blocks a launch. In Singapore, per levels.fyi (133 self-reports, November 2026), Data Scientist total compensation runs from about S$92,000 (lower quartile) to about S$163,000 (upper quartile), the top 10% from about S$201,000 — and the author cautions: do not convert to Vietnamese pay, take the shape, not the size. Before each project, ask: if this number comes out zero, who must change plans?

Experimentation means testing an idea with a controlled comparison — for example holding out a group of customers who do not receive discount codes (a holdout group) to measure whether the codes actually bring customers back — instead of believing a feature is good because it was built. Data Scientist in this essay means someone doing data analysis, measurement, and experimentation to answer business questions — distinct from the ML model builder on the engineering ladder.

A budget week in Singapore: a 94% model or an experiment table?

The essay's opening scene poses the choice: a churn-prediction model, newest architecture, 94% accurate — or a two-column table from a 3-week experiment showing a 6-month feature added no orders. Money for one direction only. The essay uses the scene to state its main argument: companies do not buy models, they buy certainty for expensive decisions. An accurate model is an asset; a number that lets leadership dare to stop a 6-month feature is also an asset — and the second kind, per the essay, is what the pay ladder rewards more as you climb.

What the pay table says: at the same rung, who is paid more?

Self-reported levels.fyi data (Meta, US, updated November 2026), per the essay:

Rung (Data Scientist ladder) Who asks the questions Median total compensation
Rungs 1–2 Others ask, you answer correctly About $153,000 and $254,000/year
Rung 3 You choose the question, own a product area's success metrics, your number opens or blocks a launch About $362,000/year
Rung 4 You set measurement standards across teams, arbitrate conflicting metrics, numbers feed budget allocation About $550,000 and $836,000/year

Three takeaways, per the essay. First, at the same second rung, model builders on the engineering ladder are paid more: about $298,000 versus about $254,000 on the Data Scientist ladder. Second, the step within the Data Scientist ladder is bigger than the gap between ladders: rung 3 is about $108,000 above rung 2 — over twice the $44,000 cross-ladder difference; the subsequent steps are about $101,000, $108,000, $188,000, and $286,000. Third, from rung 3, Data Scientists start out-earning engineers one rung below. What decides it, per the essay, is not the difficulty of the tools but the size of the decisions your numbers are trusted to close.

In Singapore, per levels.fyi (133 self-reports, November 2026): Data Scientist total compensation from about S$92,000 (lower quartile) to about S$163,000 (upper quartile), the top 10% from about S$201,000, Grab median about S$162,000; the same title in the same city spreads about 1.8×. The author is explicit: do not convert to Vietnamese pay — take the shape, not the size.

Why do companies buy "certainty" instead of models?

Per the essay, Prof. Stefan Thomke says that on average only about 10% of experiments improve profit: 9 out of 10 ideas do not add profit, and if unmeasured, they would still ship, cost maintenance, and be reported as successes. A valid experiment — like the 3-week one showing a 6-month feature added no orders — prevents an expensive decision. That is the "certainty" the essay says companies will pay for: not a beautiful model, but a number trustworthy enough to stop, redirect, or reallocate a budget.

Two friends, two choices: a 3-year story

The essay tells the story of two friends who joined the same Singapore e-commerce company with equal statistics skills:

Friend 1 Friend 2
Choice The famous homepage-recommendation team, building a more complex model The unwanted job: measuring whether quarterly discount codes really bring customers back
Year 1 Accuracy clearly up, excellent review — but the system was long-optimized, budget and headcount unchanged Hard: designing a holdout experiment, arguing with marketing over the "returning customer" definition, often the bearer of bad news
Year 2 No change The company moved a large part of the discount budget by her results and formed a promo-effectiveness team
Year 3 Same rung Promoted to lead that team

The story is the author's anecdote, unverified — but it illustrates the argument: the second person was trusted to close big decisions, and that is what the pay ladder rewards.

Is the model path a losing path?

No, per the essay — and this is the important counterbalance. At the same rung, model builders are paid more. Where the model IS the product — ad ranking, content recommendation — complexity has real value, but that is the engineering ladder, judged by engineering standards. Conversely, the Data Scientist seat has its own price: pressure for pretty numbers, the role of the person who ruins the meeting's mood, and it only fits people who find arguing over metric definitions worthwhile. And there is no guarantee: this path works only when the guarded budget is big and someone truly changes decisions by your numbers. The trap, per the essay, is thinking you are climbing because your tools get harder, while your decisions stay small.

Apply now: 3 questions to bring to an interview

The essay suggests bringing these 3 questions to an interview to learn whether the seat is truly a decision-making seat:

  1. When did an analysis last delay or cancel a launch?
  2. Who sets the success metrics for my area?
  3. What decisions does the person one rung above close?

And before each project, one question only: if this number comes out zero, who must change plans? Per the essay, measure a Data Scientist career by decision size, not model difficulty.

Frequently asked questions

1. At Meta in the US, who is paid more — Data Scientists or model-building engineers? Per self-reported levels.fyi data cited by the essay, at the same second rung, the engineering ladder (including ML model builders) has a median total compensation of about $298,000/year, above the Data Scientist ladder's about $254,000/year. But the step to rung 3 within the Data Scientist ladder (about $108,000) is more than twice the cross-ladder gap.

2. What does Prof. Thomke say about experiments? Per the essay, Prof. Stefan Thomke (Harvard Business School) says that on average only about 10% of experiments improve profit; 9 out of 10 ideas do not add profit, and if unmeasured, they would still ship and be reported as successes.

3. What is Data Scientist pay in Singapore? Per levels.fyi (133 self-reports, November 2026) cited by the essay: from about S$92,000 to about S$163,000, the top 10% from about S$201,000, Grab median about S$162,000. The author cautions to take the shape, not the size — do not convert to Vietnamese pay.

4. Is the model-building path a losing path? No, per the essay: at the same rung, model builders are paid more, and where the model is the product (ad ranking, content recommendation), complexity has real value. But that is the engineering ladder, judged by engineering standards.

5. How do you know whether a Data Scientist role has real decision power? The essay suggests 3 interview questions: when an analysis last delayed or cancelled a launch; who sets the success metrics for your area; what decisions the person one rung above closes.

Source: essay by Harry Vũ in the DoJobNgon group (18/11/2026), based on the author's observations, self-reported data on levels.fyi (Meta, US and Singapore, updated November 2026), and an interview with Prof. Stefan Thomke via HBS Working Knowledge. The original post ends by promoting Aspiring Hub (career and scholarship advising); Wotech notes this is the original post's promotional content and does not verify or endorse the claims in it.