Execution Is the New Privilege

Leo was standing under the monkey bars in the park near our apartment. A boy, maybe eight or nine, grabbed the first bar and crossed the full set without stopping, legs swinging in a lazy rhythm that only comes from having done it a hundred times before. Leo reached up, gripped the first bar, swung to the second. At the third his fingers opened and he dropped to the rubber matting. He looked at his hands for a second, then walked to a puddle near the slide.

He wasn’t upset. He wasn’t trying again. He just found something else.

In the first article in this series, “The Judgment Collapse”, I took apart Valerio Porcu’s argument that AI has made ideas officially worthless and only execution matters. I showed that the line Porcu drew between generation and judgment is dissolving from both directions. In the second, “Kill Discipline”, I followed the money behind innovation theater and found the real organizational gap: the capability to stop. Both articles ended in the same place. If execution is what counts now, who gets to develop it?

I watched Leo walk away from the monkey bars and thought: that question is not about technology. It’s about who gets to practice.

The Ladder’s First Rung

The Burning Glass Institute, working with Harvard Business School’s Project on Managing the Future of Work, published two reports in July 2025 that describe the problem with a precision that should bother anyone who manages people. The first, “The Expertise Upheaval,” identifies a paradox at the center of how AI is reshaping work: the tasks being automated are the tasks that trained the next generation of workers.

Drafting reports. Conducting research. Building forecasts. These are not just entry-level chores. They are the work through which junior employees develop instincts about what matters, what breaks, and what the numbers are actually saying. When a junior analyst builds a financial model by hand, they learn which assumptions fail first. When a new consultant writes and rewrites a client memo, they learn what persuades and what doesn’t. AI can do both tasks faster. But the speed comes at a cost that doesn’t show up on any efficiency dashboard: the training ground disappears.

The numbers are specific. The Burning Glass report estimates that AI-based automation could make nearly 18 million entry-level U.S. jobs obsolete, roughly 12 percent of the total workforce. Their companion report, “No Country for Young Grads,” documents four forces converging at once: the AI-driven expertise upheaval itself, post-pandemic lean staffing that cut mentoring capacity, AI accelerating both trends simultaneously, and a growing surplus of graduates competing for fewer entry points.

A Stanford study led by Erik Brynjolfsson put the sharpest single number on it. His team analyzed payroll data from ADP covering millions of American workers and found a 13 percent decline in employment for young workers (ages 22 to 25) in AI-exposed occupations since late 2022. For software developers in that age bracket, the drop was steeper: nearly 20 percent. Workers over 26 in the same roles saw stable or growing employment. The authors are careful to note that other factors contributed, including the post-pandemic tech correction. But the age-specific pattern is hard to explain away: if this were just a cyclical downturn, older workers in the same roles would have declined too. They didn’t.

We are watching this play in many news, across companies from all over the world. The ladder isn’t being replaced. The bottom rungs are being removed. And companies are responding the way you’d expect: by hiring experienced professionals who don’t need the rungs. Plug-and-play hires who require no onboarding, no mentoring, no ramp-up time. It’s rational from a quarterly perspective. From a ten-year perspective, it’s the organizational equivalent of eating your seed corn.

The Power Users

Two days after Porcu published his article, Anthropic’s head of economics Peter McCrory presented data at the Axios AI summit in Washington that put a finer point on the inequality question. His finding: no widespread AI-related job displacement. Not yet. But a growing gap between the people who use AI well and everyone else. Users with six months or more of experience showed a meaningfully higher success rate in their AI interactions, a gap that persisted even after adjusting for differences in task type and use case. McCrory described AI as becoming “a technology that rewards those who already know how to use it.”

He also said something that landed harder than the headline number: “Displacement effects could materialize very quickly, so you want to establish a monitoring framework to understand that before it materializes so that we can catch it as it’s happening.”

The pattern was already visible. The Harvard Business School study I discussed in the first article tested AI advisors with 640 small business owners in Kenya. High-performing entrepreneurs who got AI access saw profits rise roughly 15 percent. Low performers saw results drop roughly 8 percent. Same tool. Same advice. Different outcomes. The difference was prior experience, not the AI. I keep coming back to that study because the implication is uncomfortable: the tool works best for the people who need it least.

A survey by the Adaptavist Group of 4,000 knowledge workers across four countries quantified the gap. Employees with more than 20 hours of AI training were over four times more likely to view AI as indispensable. Nearly half of them were saving 11-plus hours per week. Those with less than one hour of training? Seven percent saved comparable time. The divide ran along predictable lines: income, seniority, gender. Workers earning above £100,000 were twice as likely to have received serious AI training. At the C-suite, 87 percent of men reported sufficient guidance on using AI at work versus 77 percent of women. Among administrative staff, that gap widened: 57 percent versus 46 percent.

The obvious counterargument: previous technologies (personal computers, the internet, smartphones) also started with access gaps that eventually narrowed. Maybe AI will follow the same curve. But the pattern here is different in one respect. Earlier technologies widened access to information and communication. AI widens access to capability. And the people best positioned to direct that capability are the ones who already had the judgment to know what to build. The gap isn’t about access to the tool. It’s about what you bring to it.

The “just execute” narrative assumes everyone starts from the same line. The data shows they don’t.

What AI Can’t Touch

So if execution is the differentiator, which execution skills actually survive? Not the ones AI handles with ease, like drafting, coding boilerplate, and data analysis. The ones that matter now sit further up the stack. Coordinating across teams whose priorities conflict. Persuading a stakeholder who’s been burned before. Reading a room well enough to know when to push and when to stop talking. These are not skills you pick up from a tutorial or a prompt. They correlate as much with social capital, network access, and mentorship as they do with raw talent. I learned them slowly, through years of sitting in rooms where the information was incomplete and the deadline was yesterday. They’re built that way for everyone; there is no shortcut.

Jared Spataro, Microsoft’s CMO of AI at Work, described the trap: “If speed is prioritized, speed scales.” But so do its blind spots. “Fewer people will practice decision-making. Accountability will become harder to trace.” His conclusion: “The organization might become faster at execution but weaker at direction.” The junior analyst who never builds the model doesn’t develop the instinct for which assumptions break. The associate who never drafts the memo doesn’t learn what persuades.

The kill capability I described in the second article, “Kill Discipline”, the organizational muscle to stop projects that aren’t working, is exactly this kind of skill. It requires seniority, political capital, and the credibility that comes from having been right about hard calls before. Nobody develops it in year one. Nobody develops it from a dashboard. And it correlates with the kind of accumulated experience that AI is now bypassing at the entry level.

The IMF’s 2025 assessment of Italy illustrates the structural side. Small innovative Italian firms struggle to become large ones. The venture capital market is limited even compared to European peers. Bank-dependent financing doesn’t support the kind of risky, intangible investment that innovation requires. Skilled professionals are in short supply. The working-age population is projected to decline by double digits through 2050. These are structural constraints that no amount of execution culture can fix without policy change. “Just execute” assumes the field allows it. In much of the world, the field doesn’t.

Tao made the point with mathematical precision on the Dwarkesh Podcast. When asked what year he’d be twice as productive from AI, he refused the framing: “Productivity, I think, is not quite a one-dimensional quantity.” His papers include more code and numerics now because AI made those cheap. But the core work, the hardest step, still happens on pen and paper. And he was explicit about what AI can’t do: “It either solves something or it doesn’t. It cannot plant a flag halfway up a cliff and build from there the way a human mathematician would.”

That cumulative reasoning, the ability to build from partial progress, to hold a failed attempt in your head and use it to guide the next one, is the human skill that matters most. In mathematics, in business, in leadership. Tao developed it through decades of sustained mathematical practice. A senior consultant develops it through years of watching strategies succeed and fail. A first-generation college student with no professional network and a four-hour commute does not develop it on the same timeline, if at all. Who gets access to the conditions that build this kind of reasoning is not a technology question.

The Next Generation

This isn’t only a current-workforce problem. It’s being reproduced in the generation that hasn’t entered the workforce yet. And the mechanism is quieter than a layoff statistic.

A pilot study at Kennesaw State University, using think-aloud protocols with 20 undergraduates, found that students aren’t just outsourcing their writing to AI. They’re negotiating when and how it belongs in their process. Deciding what to delegate and what to keep. Rewriting AI output to claim ownership. That negotiation is itself a judgment skill. And the students who did it well were the ones who already had confidence in their own thinking.

Kristi Girdharry, a writing professor at Babson College, put it directly: “Understanding what AI does to your thinking first requires knowing what your thinking can do without it.” She described students arriving already optimizing for the grade rather than the learning, having spent years producing right answers instead of sitting with hard questions. “Before they can develop discernment about any tool,” she wrote, “they need something more foundational: a sense of their own thinking as worth trusting.”

I think about this when I watch Leo work through a problem. He’s six. He gets frustrated. He doesn’t want help, and then he wants too much of it, and finding the line between the two is the whole project of being his parent. The conditions that determine which students arrive at college with that foundation are the same privilege dynamics this article is about. Educational quality. Family environments that tolerated slow thinking and didn’t treat every wrong answer as failure. Prior exposure to unstructured problems where the point was the struggle, not the solution. The students who have that background use AI as amplification. The ones who don’t risk becoming fluent with a tool they can’t evaluate. And nobody is selecting for this skill in admissions, because nobody has a rubric for “trusts their own thinking.”

The Distribution Question

Leo was still at the puddle when I looked over. He’d found a stick and was poking at something in the water. The other kid was on his third trip across the monkey bars, that lazy swing rhythm already automatic. Nobody builds practice time into the structure. But Leo will get there. He has time, and arms that are growing, and someone standing underneath the bars while he figures out the swing.

But that’s Leo. He has a parent watching from six meters away, a home full of books, and a family that treats his questions as worth answering. The question this series has been circling is not whether Leo will be fine. It’s whether the conditions he has are the exception or the norm.

Porcu said ideas are overrated and execution is what matters. Tao said productivity isn’t one-dimensional and the infrastructure for developing expertise needs redesigning. Both are correct. But neither followed the logic to where it gets uncomfortable: if execution is the only thing that counts, and the training ground for execution is being automated away, and the skills AI can’t touch are the ones you develop through mentorship, network access, and expensive education, then the people who already have those skills aren’t the winners of a meritocracy. They’re the beneficiaries of conditions that fewer people will have access to.

The article Porcu wrote was about technology. The problem it pointed at is about distribution. It always was.

Nobody is designing the lower rungs.