When AI Devalues All 'Experience', What is the Scarcity?

Opening: A Question from OpenAI CFO Puts “Experience” on Trial

On September 11, Wall Street News successively relayed a judgment from OpenAI CFO Sarah Friar: When AI devalues all “experience”, what is the scarcity for enterprises? The question for individuals is even more direct—is your experience still worth money?

This view has spread widely in communities because it doesn’t poke at abstract anxiety, but at specific scenarios happening every day: A programmer with eight years of experience finds a newcomer using AI to write a module in three days that would have taken him a week; a copywriter with ten years of experience finds they no longer need to write the first draft; a veteran data analyst of over a decade watches an intern generate charts directly using natural language. The most frequent sentence in the comments section is “Did I waste all these years?”—whether this question is worth panicking over is exactly what we need to break down in this issue.

The barriers built up by years of seniority are being dismantled in the chat box, second by second. But what is being dismantled and what is feared to be dismantled are actually not the same thing.

Shiwen: OpenAI’s CFO personally says “experience devaluation”, is your first reaction agreement or vigilance?

Yongliang: First vigilance, then agreement. Vigilance because the speaker has a motive—to pave the way for the direction of AI replacing human labor; agreement because if you remove the stance, the judgment itself holds water. But the word “experience” is used too broadly, so broadly that it almost buries the things that are truly valuable along with it.

Shiwen: Then let’s talk it through thoroughly: which part of experience is devaluing, which part isn’t, why enterprises will pay in the future, and where an ordinary person should exert their strength now.

Q1: Does the Proposition of “Experience Devaluation” Hold?

Yongliang: The proposition holds, but what is devaluing is not the part that only started devaluing when AI appeared.

I started with Java and have been in technology for seventeen years. The operational experience I accumulated over these seventeen years has been expiring non-stop: the Struts configuration skills I was proud of in my early career, the routines of manual SQL tuning, my understanding of EJB containers—more than half of these things that could fetch a salary raise back then have already entered a museum. In the intervening years, frameworks shifted from SSH to the Spring family, then to microservices and cloud-native; every blood change declared a batch of “experience” invalid. AI did not invent the devaluation of experience; it just stepped on the gas for a process that has been running for twenty years, and it stepped on it hard.

How to judge if your own experience is being devalued? I summarized a rough method: anything that can be written as an SOP, a checklist, or trained for a new recruit to pick up in three months is in the channel of devaluation. Because the essence of this part is “pattern matching”—having seen enough cases, you know what to do next. And pattern matching happens to be what models are best at; they have seen several orders of magnitude more cases than you.

What one really needs to guard against is the substitution of expressions: “experience devaluation” is being framed as “seniority devaluation”. If enterprises understand this sentence as “old employees are useless,” that is a misinterpretation—what is devaluing is the encodable, reusable operational routines, not judgment and sense of responsibility. These two things have never been written in the skills section of a resume, and have never entered a shelf life. Packing them up and throwing them away along with expired operational techniques is the most expensive mistake in this round of anxiety.

Q2: What is Truly Not Devaluing, or Even Rising in Price?

Yongliang: It is the instinct to judge “whether this solution can land,” and that moment when you step up when the system crashes.

First, let’s talk about judgment instinct. In the first minute of hearing a requirement, you know where the pits are—this client’s budget can’t support this architecture at all, that team can’t handle this process, this feature won’t be used even if it goes online. This intuition is accumulated from hundreds of failed projects; it’s not knowledge, it’s more like a sense of smell, giving conclusions without logical deduction. AI can give you ten solutions, each looking structurally complete and fully argued, but which one to choose, which one has hidden costs, which one will blow up in three months—this link still has to be human. And the more complex the organizational environment, the more valuable this instinct is, because the pits are often not technical pits, but human and process pits.

Then there is the ability to take the blame. It sounds like a joke, but it’s serious. When the system fails at 2 AM, who can judge the impact, make the call to rollback, and explain things clearly to the boss and customer service? The model can give suggestions, listing options one, two, three, four, five, but it cannot be responsible for this decision. Responsibility is a scarce commodity that cannot be outsourced—as long as the business world needs someone to sign, needs someone to be woken up by a phone call in the middle of the night, this sentence holds. I have an offbeat standard when hiring: people who have encountered major online accidents and handled them are ranked very high in priority. Because the value of such people is not in not having accidents, but in being able to hold down the losses when accidents happen.

These two things have one thing in common: they are not “learned”, they are “trodden out” (through experience). Models can learn knowledge from corpora, but they can’t trod through the pits for you.

Q3: How to Respond to the Counter-View—Give the Model Enough Context and Evaluation Means, Isn’t Judgment Just a Matter of Time?

Yongliang: It is a matter of time, but “a matter of time” is precisely the point—because making tacit knowledge explicit itself still has to be done by humans, and in the visible window, the people doing this are the scarce commodities.

Take an example from my industry. AI gives a medical informatization solution, technically flawless: reasonable architecture, mature components, performance indicators all met. But it doesn’t know which supplier made the hospital’s HIS system ten years ago, how messy the data dictionary is, that the Director of the Information Department and the Vice Dean are pushing the same thing but their statements are inconsistent, or whether the Dean really wants the rating or cost control. This context cannot be found in documents; it’s in people’s brains. For the model to digest this layer of judgment, someone first has to make all this tacit knowledge explicit, organize it into a form it can digest, and feed it to it—and the one doing this is still a human.

Some will say, then when evaluation methods mature, the model can verify the quality of the solution itself. The problem is that evaluation can only cover defined indicators, and implicit constraints are precisely not in the indicators—how do you write “Vice Dean and Director have inconsistent statements” into the evaluation? How do you write “The Dean wants the rating” into unit tests?

So my answer is: judgment will indeed eventually be partially compressed into the model, but every step of this compression process requires human porters. The scarcity in the future is not the abstract word “judgment”, but “people who can feed implicit context to AI”. The scarcity has shifted from “people with experience” to “people who can lead AI”—this shift is far more important than the four words “experience devaluation” because it points out where humans should exert their strength.

Q4: If the Scarcity Changes, the Logic of Corporate Spending Changes—Where?

Yongliang: The premium shifts from “seniority” to three things: private data, business translation capability, and accountability mechanisms.

In the past, enterprises paid a premium for seniority because seniority ≈ experience ≈ fewer pits. Experience itself has become cheaper, money will flow to three new targets.

First, high-quality private data. This is a moat that models cannot replace—models eat public corpora, they haven’t seen your transaction records, process parameters, customer history, or service work orders. Whoever has more and cleaner private data will have a faster AI landing. So when I do consulting for enterprises, my first suggestion is always: don’t rush to get on the model, first take stock of data assets.

Second, people who can translate business problems into AI tasks. The business side says “improve retention”, this is a requirement not a task; the task is “segment users silent in the last 90 days, predict churn probability, generate three tiers of recall scripts and design A/B tests”. This translation in the middle, right now few people can do well—it requires understanding both business jargon and the boundaries of model capability. When I hire, I value this the most, preferring to let go of someone who is purely technically strong.

Third, accountability mechanisms for results. The higher the participation of AI, the more decent the output, the more critical the “person who signs and takes responsibility” in the organization. Auditing, reviewing, signing chains—these links that sound bureaucratic will become more important, not less, because of AI. Conversely, pure execution layer positions will be continuously compressed—junior code positions, first draft copywriting, basic data processing, the quantity and unit price of these positions are going down. This isn’t my prediction, it’s the employment direction already implied when someone of OpenAI CFO’s status says this.

Q5: An Engineer with 5 Years of Experience Whose Skills are Being Covered by Models, What Should They Invest In Now?

Yongliang: Stop rolling in the direction of “more proficient”, walk in the direction of “harder to describe”.

Specifically three points. First, build value on specific situations: the industry you are in, the clients you have touched, the messes you have handled that aren’t written into textbooks. Models are good at fields with standard answers and massive corpora; specific situations are your breakwater. Writing the same code, the price gap between an engineer who understands hospital business processes and one who doesn’t will widen, not narrow, in the AI era.

Second, take real problems to walk the AI every week, but don’t copy the answer, review why it was wrong. Build a document, note down where it took things for granted this time, where it missed constraints, which suggestion would have caused trouble if followed. The places where it is wrong are the scale of how much human value remains. Stick to this habit for half a year, and you will see your moat earlier than those around you, and the accumulated “AI pitfall file” itself will become your private data asset.

Third, practice expression. In the future, competition between people is largely a competition of “who can explain the problem clearly”—explaining vague business demands as instructions AI can execute, explaining technical feasibility as decision bases the boss can understand. Expression is the exit of tacit knowledge; the wider this exit, the higher the quality of the context you feed AI, the greater your leverage.

Epilogue

Shiwen: Finally, summarize this issue in one sentence?

Yongliang: Every time a technology wave shouts “such and such is going to devalue”, what truly devalues in the end is often just the threshold—the door gets lower, but the room inside is even more crowded; if you want to stand firm, you still have to have something others can’t take away.

Shiwen: This sentence is for everyone. See you next time.

---

【Technical Deep Dive】Why Models Can’t Eat Up “Tacit Knowledge”: Deconstructing from the Perspective of Context Engineering

This issue repeatedly mentions “implicit context”, let me break it down for technical readers why this is a structural shortcoming of the model, not just a temporary lack of effort.

First look at the input essence of the model. The entire input of a large model is a sequence of tokens. The so-called “giving the model enough context” means in practice: someone organizes the judgments in their brain into text, retrieves them into documents, and stuffs them into the context window. This “organize—retrieve—stuff in” process is exactly the context engineering the whole industry is doing now. It precisely proves my point: every time tacit knowledge is used, it must first go through a process of artificial explicitation. Tools are evolving, from manual pasting to RAG to agent automatic retrieval, but the step of “source knowledge coming from the human brain” has never changed.

Then look at why retrieval can’t save the day. RAG’s assumption is “knowledge has already been written into documents and can be found by embedding similarity”. The most valuable judgments in organizations are precisely those not written down: why B was chosen instead of A last time, what things a certain client verbally promised, which segment of code is a mine left by the predecessor. Embedding captures semantic similarity, not causality and pros/cons—and these are the raw materials for judgment.

Finally look at the actual level of agent memory. Current agent long-term memory schemes (conversation summaries, vector memory, skill files) can remember “what happened”, but it is hard to remember “why the decision was made at that time”. The context of the decision—the constraints, trade-offs, risk preferences at the time—is almost impossible to be compressed and stored. This is also why in most enterprise AI landing projects I have seen, the bottleneck is never in model capability, but in no one being willing to spend time organizing knowledge explicitly.

In one sentence: The model’s appetite for knowledge is infinite, but the spoon feeding it has always been in human hands. This is the engineering-level explanation for why “people who can lead AI” are scarce.