Stop Betting on What AI Can't Do. Bet on What Compounds.

There's a genre of advice flooding every feed right now: the list of skills AI supposedly can't replicate, usually taste, creativity, and emotional intelligence, usually attributed to unnamed CEOs, usually ending before the only useful part, which is how you'd actually build any of them. The lists aren't wrong so much as built on a frame with a short half-life. Every "AI can't do X" claim is a bet against the next model release, and the last three years have been a graveyard for those bets. Jensen Huang told the world kids may not need to learn to code; Y Combinator then reported a quarter of a batch shipping codebases that were 95% AI-generated. The limits keep moving. Careers built on them move with the limits, which is to say, backward.

So here's the sturdier filter, the one we hire against and the one I'd give my own kids: don't ask what AI can't do. Ask what gets more valuable when it does more. Does this skill compound or depreciate when the models improve 10x? Production skills (drafting, coding, decking, editing at volume) depreciate: not to zero, but to commodity, because the models compete with them directly. Complement skills appreciate, because every improvement in generation increases the leverage of whoever directs, selects, and answers for the output. The whole career question fits in that one test, and it's evergreen precisely because it doesn't depend on where the frontier is this quarter.

The four skills that pass the test

Judgment, which is what taste actually is. When a model can produce a hundred competent options in a minute, the value of producing option 101 rounds to zero and the value of knowing which one is the one goes vertical. Choosing is the skill that scales with generation capacity by definition: more options, more value per good chooser. This is the appreciating asset the taste lists gesture at without naming the mechanism.

Problem selection. The models answer; they don't yet ask well. Deciding what's worth building, which pain is real, which market moment is now, remains the highest-leverage human act, and it gets more decisive as execution gets cheaper: when anything can be built in a week, choosing the wrong thing is the only expensive mistake left. Even inside the fastest-scaling company on earth, Anthropic's growth lead describes opportunity identification and stakeholder alignment as the stubbornly human end of the pipeline, joking with a colleague that AGI will arrive before six people in a room agree on anything.

Accountability. Someone has to sign. As we've argued in the signature economy, provenance is where value concentrates when perfection is free, and accountability can't be automated for a structural reason rather than a technical one: it isn't a capability, it's a stake. A model can produce the recommendation; it cannot have something to lose if the recommendation is wrong. The person willing to be answerable becomes the scarce input in every process the machines accelerate.

The human interface. The honest version of the emotional-intelligence claim: reading the room, aligning the stakeholders, closing the deal, delivering the hard message so the relationship survives it. Some of this the machines will assist (ours already draft the kinder version of a heated Slack message), but the trust-bearing moments stay human because trust is between accountable parties, and see the previous skill.

Notice what's missing from the depreciation column: fluency itself. The viral scripts say the future skills have "nothing to do with tech," and that's the one claim we'll flatly contradict, because we published a hiring ladder that says the opposite. Judgment without leverage loses to judgment with it, every time, which is why level 4 fluency is our floor and not our ceiling. The durable stack is multiplication, not substitution: fluency times judgment. Skip the tools and your taste never touches the work; stop at the tools and you're a prompt away from replaceable.

How you actually build them, since nobody ever says

Judgment trains like pattern recognition, because that's what it is. Reps plus articulated reasons. Make real calls, write down why before you see the outcome, review the ledger monthly. We know taste is trainable because we built a product that extracts it: the Taste Decoder turns a decision-maker's yes/no history into explicit principles, which means the principles were learnable all along; the exercise is doing that extraction on yourself, continuously. Volume of exposure matters too, but only graded exposure: consuming great work while articulating why it's great, which is different from scrolling past it.

Problem selection trains in the pain archives. Lost-deal notes, support tickets, churn calls, the places companies keep their real problems. An hour a week in there beats a year of brainstorms, because the skill is recognizing which hurt is worth solving, and recognition needs specimens. Add constraint practice: artificial limits (half the time, one page, no budget) force the reframing muscle the research says abundance atrophies.

Accountability trains by exposure, and only by exposure. Ship under your name. Publish the position someone can attack. Take the client call when the work missed. There is no simulator for having skin; you acquire the skill by acquiring the stake, in increments sized to survive.

The interface trains on the reps the machines just freed up. Run the kickoff. Mediate the disagreement instead of escalating it. Ask for the deal. The gift of the automation wave is that admin no longer eats the hours, so the hard human moments can get the practice they never used to; the mistake is letting the saved hours go to more production instead.

The honest caveat, which is also the meta-skill

Some of today's complements will partially automate; we run audience simulations that already do a version of "reading the room" at scale, and we'd be foolish to pretend the boundary is fixed. That's exactly why the frame matters more than the list: "AI can't replicate X" expires, but "does X compound with model improvement" is a question you can re-ask every year, and re-asking it is the actual future-proof skill. Careers don't survive by standing where the machines aren't. They survive by standing where the machines push value, and pushing value toward judgment, selection, accountability, and trust is the most consistent thing the machines have ever done.