Your Abilities Are Being Repriced: A Guide to Programmer Asset Revaluation in the AI Era

There are two voices in the market about AI and programmers.

One says: AI can write anything, programmers are finished. Another says: AI is just a tool, the code that needs writing still has to be written by humans.

Both are too tidy. The real world is never tidy.

Over the past two years, I've seen a more awkward and more thought-provoking picture: the entry point for junior positions is narrowing, the number of AI-related positions is exploding, and within the same group of people, some are being lowballed, some are being paid more.

This is not a binary "unemployed or not unemployed" problem. This is a revaluation of capability assets.

---

1. Two Curves: the Entry Is Collapsing, the Stock Is Growing

Look at US data first.

Stanford Digital Economy Lab used data from the largest payroll software in the US for a study: since the spread of generative AI, young workers aged 22-25 in AI-high-exposure occupations saw relative employment fall 13%. Senior employees in the same occupation, and workers in low-exposure occupations, remained stable or even continued growing. [1]

In August 2026, the study updated the data: young workers' employment in AI-high-exposure occupations is now 19% below what "trends for comparable groups would predict," and the gap has been widening since 2025; the adjustment has mainly been achieved through reduced hiring, not large-scale layoffs. The study also emphasizes: there is no evidence of a general unemployment wave. [2]

These two numbers need to be read together.

• 13% → 19% means the impact is deepening;
• "achieved through reduced hiring" means the stock is relatively safe, the entry has been closed first;
• "no general unemployment wave" means this is not a recession, it is structural substitution.

Look at China.

36Kr cited recruitment platform data: in the first half of 2026, new AI job postings grew 789% year-over-year; roughly 1 in 4 new positions is an AI position. [3] Zhaopin data shows that from January to May 2026, AI graduate positions grew 28.4% year-over-year. [4]

On one side "1 in 4 positions is an AI position", on the other "22-year-olds can't get in the door." Contradictory?

Not contradictory. Positions are moving upstream, and people are still standing still.

Figure 1: In AI-high-exposure occupations, young workers show a relative employment gap, senior workers remain stable. Source: Stanford Digital Economy Lab (2025, 2026).

---

2. Why the Entry Collapses First

Because the work of junior programmers is, in essence, "a gift bag of standardized tasks."

Writing CRUD, reproducing design mockups, integrating APIs, filling in unit tests, writing scripts, changing configuration — these tasks have clear input and output, lots of public examples, and errors are easy to find. They happen to be exactly what large models are best at.

Anthropic's January 2026 Economic Index report provided a profile: in November 2025, computer and math tasks (mostly coding-related) accounted for 34% of Claude.ai conversations and 46% of first-party API traffic; on Claude.ai, that share has fallen from the March 2025 peak of 40%, while API-side automated use is still rising. [5]

Translation: people use the chat box to "ask AI to help write", companies use the API to "let AI write directly."

Entry-level positions carry exactly the tasks that can be written away directly. Senior positions carry "deciding what to write, to what standard, and what to do when things go wrong" — these tasks have not been standardized, so they have not been absorbed directly.

Marx discussed at length in Volume 1 of Das Kapital how machines displace workers. Later economists proposed "compensation theory": new technology destroys old jobs while creating new jobs, eventually compensating back.

This theory may be right over the long cycle. But it has one fatal premise: compensation takes time. And time, landing on specific people, means months or years of unemployment, career change, pay cuts.

In Zhuangzi: The Secret of Caring for Life, the butcher Ding reaches the highest state in carving an ox:

I meet it with spirit and do not look with eyes. The senses stop, and spirit moves as it wills.

The knife follows the gaps between bones; it does not hack hard. What the eyes see is the "look with eyes" part; what the hand catches is the "meet with spirit" part.

AI is strong, but strong in "looking with eyes" — it can see code, documentation, error messages. What it cannot see: what the boss behind this requirement really wants, whether this change will anger some customer, that this technical debt cannot be paid off now, who will take the blame after this solution goes live.

The "looking with eyes" part is being automated; the "meeting with spirit" part is going up in price.

---

3. Capability Asset Revaluation Matrix

I rearranged programmer abilities along two dimensions:

• Horizontal axis: degree of standardization — can this be written into clear rules, is there a large amount of public examples;
• Vertical axis: weight of responsibility — if it goes wrong, who bears the consequence, how big is the cost.

Figure 2: Capability asset revaluation matrix. The further bottom-right, the easier to be replaced; the further top-left, the easier to earn a premium.

Four quadrants, four fates.

Bottom-right: low responsibility, high standardization — depreciation zone.

Writing boilerplate code, simple page restoration, common SQL, basic unit tests. Not "useless", just "not worth much anymore". These abilities become everyone's default configuration, just like today nobody gets a raise for "knowing how to use a search engine."

Top-right: high responsibility, high standardization — engineering zone.

SRE, security, compliance, payments, permissions. The work itself has method, but the cost of error is enormous. AI can help you do it, but it won't sign for you. Prices here are stable; the threshold is "auditable, rollback-able, accountable."

Bottom-left: low responsibility, low standardization — long-tail service zone.

Small tools, small scripts, small customization, fragmented needs in an industry. Low unit price, but real demand, scattered competition. Suitable as a first stop for a side business, not as a long-term moat.

Top-left: high responsibility, low standardization — premium zone.

Requirement definition, architecture trade-offs, cross-team coordination, online firefighting, industry know-how, customer trust. What these things have in common: no standard answer exists, but someone must be responsible for the answer.

The more ubiquitous AI becomes, the more valuable this quadrant is. Because when "making something" becomes cheap, "deciding what to make, and being responsible when it's wrong" becomes expensive.

---

4. Three Types of Appreciating Assets

1. Judgment: knowing when not to trust AI

There is a detail in Stack Overflow's 2025 survey: senior developers are the group that "trusts AI the least" — only 2.6% "highly trust", 20% "highly distrust". [6]

This is not conservatism. This is a lesson learned in production.

A beginner sees AI-generated code run, and assumes it's correct. A senior person asks first: edge cases? Concurrency? Permissions? How do we roll back if something goes wrong? This is not "knowing more syntax", it is "knowing more about cost."

The essence of judgment is sensitivity to consequences.

2. Verification: turning "looks right" into "actually right"

Veracode's 2025 report tested 80 coding tasks with 100+ large models: 45% of AI-generated code introduced OWASP Top 10 vulnerabilities, only 55% passed security tests; and this security performance did not meaningfully improve as model generation capabilities improved. [7]

This means: the more AI writes, the higher the value of testing, review, security, and observability.

The most expensive kind of engineer in the future may not be "the one who writes the most" but "the one who verifies what others write to the point it can go live."

3. Integration: plugging AI into real broken systems

In Upwork's 2026 report, the fastest-growing coding skills are AI integration (+178%) and AI chatbot development (+71%); meanwhile, full-stack remains the #1 most demanded skill. [8]

These two sentences should be read together: enterprises are not trying to "build an AI-native app from scratch", they are trying to "shove AI into a system that has run for ten years." Old systems have historical data, permissions, weird business processes, and piles of legacy code you cannot touch.

The model is responsible for "can"; the integrator is responsible for "usable".

Integration = domain knowledge × engineering ability × communication ability. These three are precisely what AI is worst at.

The Analects, chapter "Wei Zheng" says "君子不器" — a gentleman is not a vessel. The original meaning: a gentleman should not be a tool with only one use. Today, we can give this a plain reading:

Don't turn yourself into an "interface that only gets called." Interfaces are standardized, and standardized things get replaced. Be the one who decides "what to call, how to call, and who to go to when things break."

---

5. Three Steps to Complete Your Capability Revaluation

Enough theory. Here is the method; the first round can be finished in one weekend.

Step 1: Inventory — break work down into tasks, not job description

Take the last 6 months' calendar and commit history, list 30 things you actually did. Don't write "responsible for XX system development"; write concrete tasks, for example:

• Organize customer feedback into a requirements document
• Write a batch export script for operations
• Troubleshoot a production timeout
• Persuade product to cut an unreasonable requirement
• Explain the permission model to a new hire

Then score each item on two dimensions (1-5):

• Standardization score: how many public examples exist? Can the rules be written clearly?
• Responsibility score: if it goes wrong, who bears the consequence? How big is the cost?

After scoring, you'll see a cruel but useful distribution: the things you spend the most time on are not necessarily the things that make you valuable.

Step 2: Pricing — pick out the "high responsibility + low standardization" tasks

Filter for tasks with "responsibility ≥ 4, standardization ≤ 3". These are your premium zone.

Then do a translation: translate the task into a "customer-perceivable result."

What you did	Customer-perceived result	How to price it
Troubleshoot production timeout	Peak availability from 99.5% to 99.95%	Maintenance fee based on availability target
Organize customer feedback	Requirements rework rate down 30%	Diagnostic fee per project phase
Write scripts for operations	Save 40 person-hours per month	Revenue share based on cost saved
Persuade product to cut requirement	Delivery cycle shortened by two weeks	Project management fee based on delivery cycle

Note: this is not teaching you "packaging". This is teaching you to describe value in the other party's language. The same work, described as "I wrote 3,000 lines of code," vs. "I made your team work 40 fewer hours a month," receives two different quotes.

Step 3: Proof — turn ability into verifiable evidence

In the AI era, resumes are failing, because "what you know" is too easy to exaggerate, and too easy for tools to level the field.

Effective evidence comes in three types:

1. Reproducible work: a link that can be clicked, runs, has real users;
2. Comparable metrics: before/after data, with methodology written clearly;
3. Traceable process: decision records, pitfall records, retrospectives. It proves not "you can", but "you judged."

This is also why I recommend every programmer maintain a "decision log": not a running log, but recording "what the options were at the time, why you chose this one, and whether it turned out right later."

A year later, this log is worth more than any resume.

---

6. Finally: Configure Your Career with the Barbell Strategy

Taleb's Antifragile proposes the "barbell strategy": put most resources in extremely safe places, a small portion in extremely high-volatility experiments, and avoid the middle.

Applied to a career:

• 80% of energy, into a main job that provides stable cash flow and accumulates "responsibility assets." Don't ruin your main job chasing AI hotspots.
• 20% of energy, into high-volatility experiments: take a project you've never done, build a small product nobody might use, write a series of articles in a vertical domain.

The middle ground is the most dangerous: neither stable cash flow nor high-upside possibility. For example "scroll through two hours of AI news after work but do nothing."

The Book of Changes, "Xici Xia" says:

When things are exhausted, change; when changed, things are unblocked; when unblocked, they last.

It is not "if you change, you will surely be unblocked"; it is that when "exhausted" you must change. For many people, "exhausted" doesn't mean no money; it means abilities are still there, but the price is no longer rising.

Capability revaluation is not a crisis; it is more like a re-quote.

The market is relabeling every capability. What you can do is not stop it from labeling, but make sure you stand on the left side of that label — closer to consequences, farther from standard answers.

---

References

1. Stanford SIEPR, Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence (2025): https://siepr.stanford.edu/publications/working-paper/canaries-coal-mine-six-facts-about-recent-employment-effects-artificial
2. Stanford Digital Economy Lab, Canaries in the Coal Mine? (2026-08 update): https://digitaleconomy.stanford.edu/app/uploads/2026/08/Canaries_August2026.pdf
3. 36Kr, 2026 Autumn Recruitment, requiring fresh grads to have "3 years of experience" per person?: https://36kr.com/p/4025027072577414
4. China Economic Net, Internet companies release 200k+ positions, "AI+" talent demand rising (2026-07): http://www.ce.cn/xwzx/gnsz/gdxw/202607/t20260703_3065683.shtml
5. Anthropic, Economic Index report: Economic primitives (2026-01): https://www.anthropic.com/research/anthropic-economic-index-january-2026-report
6. Stack Overflow, 2025 Developer Survey — AI: https://survey.stackoverflow.co/2025/ai/
7. Veracode, 2025 GenAI Code Security Report: https://www.veracode.com/press-release/ai-generated-code-poses-major-security-risks-in-nearly-half-of-all-development-tasks-veracode-research-reveals/
8. Upwork, In-Demand Skills 2026 (2026-02): https://www.upwork.com/press/releases/upworks-in-demand-skills-2026-demand-for-top-ai-skills-more-than-doubles-as-ai-is-embedded-into-everyday-work