Career Direction When Tech Skills Depreciate — Move from Execution to Judgment
This article comes from an anxiety that kept resurfacing for me recently: AI keeps getting stronger, I cannot out-code it, and if 99% of jobs get replaced by AI plus robots within 3-5 years, the value of a technical career goes to zero — so where should I pivot?
After taking the question apart seriously, my conclusion is: the premise is half right and half wrong, and the wrong half is exactly what decides the direction. You do not need to “change direction”; you need to change position.
1. Fix the premise first
“99% of jobs replaced in 3-5 years” is a guess, not a fact, and it is most likely wrong. Not because AI is not strong enough, but because the bottleneck of replacement speed has never been the technology:
- Physical-world jobs are gated by robot hardware cost, supply chains, installation and maintenance, and safety regulation — these move at the pace of capital expenditure and approval cycles, not model iteration speed;
- The essence of many jobs is responsibility, not labor: the doctor’s signature, the accountant’s stamp, the engineering sign-off. AI cannot go to jail, cannot pay damages, cannot be held accountable. The “labor part” of these jobs will be hollowed out, but the “signing authority” stays with humans;
- In every previous wave of automation, the gap between “technically feasible” and “socially completed replacement” was 15-30 years, not 3-5. The spreadsheet was invented in 1979; accountants still exist today.
But the half that is right is brutal: the premium of coding as a craft is collapsing, irreversibly. I do not need anyone’s forecast for this — I am the evidence. My own effort estimates already run at “4 human weeks ≈ 1-2 AI days”. The day before, I watched an agent on my Mac build a full-stack project in a single day: backend, two client apps, local model inference. The market price of “translating requirements into code” is going to zero, and first-hand data beats any report.
2. Reframe: what depreciates is not “technology” but “execution”
Take “technology” apart and you see two different things — one going to zero, one appreciating:
| Going to zero | Appreciating |
|---|---|
| Turning a clear spec into code (translation labor) | Deciding what to build and what not to (judgment/taste) |
| Execution capacity sold by the hour | The position that signs off and bears responsibility for outcomes |
| The knowledge premium of “knowing framework X” | Why customers trust you and buy from you (trust/relationships/distribution) |
| Labor that is employed | Assets that are owned (product/customers/cash flow/audience) |
The key insight: AI driving implementation cost to zero is a disaster for people selling hours, and a massive subsidy for people who own ideas, customers, and the authority to decide. The same technological shift points in opposite directions depending on where you stand.
So the question is not “which new technology to learn”, but “how to move from the position of selling execution to the position of owning judgment, trust, and assets”.
3. The blind spot inside the question itself
Look back at how I originally asked it — “which direction should I learn toward” — that is the engineer’s muscle memory: when facing uncertainty, the solution is to learn one more thing.
But none of the four appreciating things in the table above can be “learned”. All of them are trained by making real decisions, bearing real consequences, and accumulating real relationships.
And I already have two ready-made training grounds:
- Selling the product: landing 30 paying customers. The sales line trains decision-chain analysis, pricing, negotiation, trust building, and root-cause investigation of “why won’t the customer pay” — exactly the top of the AI-resistant list. I eventually realized the real tension is never whether the product is good, but whether the customer has enough motivation to switch — being able to say that sentence is itself a marker of moving from technical thinking to business judgment. Actually landing those 30 customers completes a pivot, worth more than reading 100 essays on “what to do in the AI era”.
- The content and personal-brand line. The first article on my public channel is already out — the first brick of the “distribution and audience” asset. An audience is one of the few assets that becomes more valuable as AI gets stronger: once content production cost goes to zero, attention and trust become scarcer.
As for my ongoing agent-learning main line — it does not conflict with the pivot, but its positioning needs calibration: “being able to write agent code” will depreciate in three years just the same; what does not depreciate is the systemic judgment of where agents should be used, how to orchestrate them, how to verify their work, and how to take responsibility for the results. My long-term bet on Harness Engineering is the right one — the essence of a harness is “the human steps back to the judgment and verification position and hands execution over”, which lands exactly in the right column of the table.
4. The recommendation (cut down to one axis)
Open no new direction. The pivot for the next 6-12 months = re-weight the three existing lines by their “right-column content”:
- The sales line gets the highest weight — it is the only line that directly trains “judgment + trust + responsibility”. Even if it is slow and uncomfortable (engineers instinctively retreat into code), it is the real pivot; writing code is fake safety. I set myself a red line: any week where the tooling advanced but sales did not move equals using coding to escape selling.
- Keep the content/brand line moving in small steps — monthly output is enough; the compounding pays off in three years.
- Shift the agent main line from “learning tools” to “producing judgment” — hoard fewer tutorials, accumulate more of the trade-offs and failures made in a real multi-machine mesh. People who can use agents will stop being scarce very soon; people who have crashed with agents and can articulate the trade-offs stay scarce much longer.
5. One last reminder
“Pivot anxiety → learn another new thing” is the most sophisticated disguise of the pattern “using high-intensity hands-on work to avoid low-output judgment and closure” — it feels like solving the problem, while actually fleeing the two uncomfortable things that are genuinely valuable: close the deals, and build the name.
Open to rebuttal: if you have concrete evidence for the “99% in 3-5 years” premise (a real replacement case in some industry), bring the data and let’s argue.