Revenge of the Nerds

I’ve been reading the hiring data on what AI is doing to entry-level roles in tech, and the pattern is hard to miss. Stanford’s Canaries in the Coal Mine study finds software developers aged 22 to 25 are down about 20% from their late-2022 peak while older developers keep climbing. Harvard’s Generative AI as Seniority-Biased Technological Change paper finds junior employment at AI-adopting firms fell 7% to 12% while senior employment rose. SignalFire’s State of Tech Talent report puts entry-level hiring at the big tech companies down about 65% from 2019. What’s less clear to me is what happens next. It affects the juniors who can’t get hired, and the seniors who think they’re insulated.

This note is for senior PMs. It’s about how to use these capabilities well, what to avoid, and the pressure that’s coming from the juniors you stopped hiring.

Judgment is a treadmill, not a moat

The naive economics says juniors should be able to do seniors’ work at a discount. If AI commoditizes production, the senior’s premium should collapse and a junior at a third of the salary should take the job. That isn’t happening. Hiring is down, not repriced. The senior’s job was never production. It was judgment, and judgment doesn’t transfer at a discount. It’s accumulated, it’s tacit, and you can’t buy it cheap.

But the premium erodes the moment you stop climbing. Judgment is being codified and exported into the stack. Every standing instruction, every eval, every prompt you write is your judgment captured as a reusable asset. Anthropic’s internal data is the smoking gun. Written standing instructions grew from a few hundred in July 2025 to about 8,500 a year later. Every one of those is a senior exporting their judgment so a junior, or an agent, can wield it without them.

The O-ring argument cuts both ways. If AI review gets better and the cost of error falls, the downside risk of a junior’s mistake shrinks. That shrinks the premium on your error-avoidance. The jagged frontier is moving. The things you uniquely catch are a shrinking set as models handle more of the work of finding where the error is.

So the premium doesn’t vanish. It migrates up the stack. The senior who stops producing and starts directing keeps it. The senior who just produces loses it. Judgment insulates you only as long as you keep moving up onto what’s new, what the model can’t do yet, and how to steer the agents.

How to use AI as a senior PM

The test for any use of AI is whether it amplifies your judgment or replaces it with output. If it replaces judgment with output, you’re using it as a junior. If it amplifies judgment, you’re using it as a senior.

Use it for:

Adversarial sparring. Have AI argue against your decision before you commit. Red-team your own reasoning. Ask what am I missing, what’s the strongest case against this, what would a competitor do. You bring the frame; it brings the counter-arguments. This is the highest-leverage use, because it sharpens the one thing that’s actually scarce.
Synthesis at scale. You bring the question and the frame; AI brings the breadth. The frame is the judgment. That’s the part only you supply. That’s the difference between a digest and a note. The sources are aggregate. The question is yours.
Building judgment-encoding systems, not one-off apps. Instead of an app that does a task, build the standing instructions, evals, and context that capture how you decide. Then your judgment scales across the org. This is the meta-work Anthropic found growing from 10% to 22% of sessions. It’s the strategic version of hiring a junior. You’re not getting output. You’re getting your decision-making replicated.
The chief-of-staff pattern. AI that triages, briefs, and surfaces. Your attention goes only to the decisions only you can make. Your judgment is the scarce resource. Spend it on those decisions, not on the production an agent can do.

What my own practice tells me

I keep a running set of notes on the craft of product. When I look at them through this lens, they split cleanly into two piles. The split is the argument.

The notes AI upends are the mechanical ones. These are the production of the craft:

Customer Interviews. The question-generation and the synthesis of transcripts. AI does this now. The “tell me about the last time you…” interview still needs a human in the room, but the production around it is absorbed.
Discovery and Prioritization. The opportunity trees, the ICE/RICE scoring. The mechanics of scoring are exactly the small, well-specified work AI absorbs.
Backlog and Process. The triage, the agile/SAFe ceremony. The overhead, not the judgment.
Meetings. I already wrote that 90% of status updates should be async. AI makes that 100%.

These are the junior work of PM. They get upended because they’re production, not judgment.

The notes that survive are the judgment ones. These are where the value is in the decision, not the output:

Making Decisions. The six questions, the lowest-competent-level principle. AI can’t decide for you; it can only argue with you.
Managing Uncertainty. The five-paragraph order, EMDCOA/EMPCOA. The diagnosis is the judgment.
Crafting Business Objectives, Market Analysis, Pricing, Product Market Motions, Strategy Assessments. The strategy core. AI can draft the analysis; it can’t set the diagnosis.
Pageantry of Product Reviews. Start with the bad news. That’s a judgment call about how to run a review, not a production task.
Organization Models and Player Coach Management. Org design. The judgment about how to organize.

And the bridge note is Prompts as the new PRDs, because that’s the one where I’m already doing the thing this note recommends. I’m encoding judgment into reusable instructions. The JTBD template, the system prompts, the eval criteria. That’s judgment exported into the stack. It’s the strategic use, and it’s the same move that, done carelessly, gives away your moat.

The test holds. The notes that get upended are the ones where the output was the point. The notes that survive are the ones where the decision was the point.

What to avoid

The “hired a junior” trap. If an AI use feels like you’ve hired a junior to do work for you, it probably is. You told it what to build and it built it. That’s cheap labor, and it’s the least strategic use of your time. It’s not wrong; it’s just not where your leverage is. The tell is the feeling. If the output is the point, you’re producing. If the decision is the point, you’re directing.

The “AI drag” on your own judgment. The evidence that AI gives seniors a boost but drags on juniors (Rasinovich & Hanselman, CACM April 2026; the 30-million-commit study) applies to you too if you stop doing the hard thinking. If you let AI write the strategy and you just approve it, you’re the junior in that relationship. The model is jagged. The value is in handling the edges, which means knowing where it’s wrong.

The moat you’re giving away. Every standing instruction you write is judgment exported. That’s good when it scales your org. It’s bad when you export the core of what makes you senior and keep nothing for yourself. Direct the agents instead of becoming one.

The coming pressure from juniors

Here’s the part the reports don’t cover. The juniors you stopped hiring aren’t gone. They’re underemployed, and they’re watching. The revenge of the juniors won’t come from AI getting smarter. It’ll come from three places:

The training ground being rebuilt. If the industry rebuilds apprenticeship deliberately, with simulated struggle, structured exposure, and compressed learning, a cohort of AI-native juniors could learn judgment faster than the old ten-year grind, flood the market, and undercut. This is the revenge we choose to create.
The senior-retirement clock. If the pipeline was cut and the senior cohort retires, there’s a shortage of people who can steer agents. The juniors who did get trained, or self-trained, become the scarce resource. The cohort that was locked out becomes the thing everyone suddenly needs. This is the revenge that time delivers, on a 10 to 15 year lag.
The judgment you exported. The standing instructions you wrote are now reusable assets. A junior with access to them, and the AI to wield them, can approximate your judgment without your years. You trained your replacement one prompt at a time.

The pressure is real and it’s coming. The question isn’t whether juniors will respond to being underemployed. They will. It’s whether you’ve used the window to move up the judgment stack, or whether you’ve spent it exporting your moat.

Two essays that reframe the revenge

I’m factoring two essays from this week into this note. Both get at the same thing from opposite directions, and together they put the “revenge of the juniors” premise under pressure.

Ben Thompson, “The End of the Internet” (Stratechery, Oct 26 2026). Thompson reads the web’s arc as repeated collapses of “everything open to infinite competition to terrible for business,” each rescued not by more scale but by going smaller. Subscriptions over advertising. Group chats over public social. Now the same decay hits software itself. There’s a wave of zero-day-class failures (PHP remote code execution, compromised intermediate CA keys, Node/GitLab/Java/Cisco bugs, LiteLLM and Railway supply-chain breaches) because all of software is built on an inherently creaky human-authored foundation, and AI gives attackers limitless iteration in place of cleverness. His conclusion is not Luddite. The answer is to build software for one person or one team. This is the group-chat era of software, where AI makes bespoke tools viable and, in aggregate, a larger market. Not one global market but infinite local ones.

The scarcity that mattered was never attention. It was the cost of bespoke production. When AI drives that cost toward zero, small beats scale.
“Smallness at scale… there is not one global market, but infinite local ones.”

Venkatesh Rao, “Cognitive Mercantilism” (Contraptions, Oct 25 2026). Rao maps mercantilism’s four beliefs onto intelligence. Mercantilism assumed wealth is scarce, the poor can’t be trusted, supply is the central economic fact, and most people are only good for subsistence labor. The cognitive version hoards brains and IP like bullion, lets cognitive elites govern “stupid people,” treats supply as the central fact, and fears that automating low-grade cognition creates “billions of useless minds.” His counter, drawn from Joyce Appleby’s history of capitalism, is that agriculture didn’t end with useless mouths. It freed a vast non-farming majority to become consumers of food and a source of demand. The analogous claim is that most people will no longer need to be significant producers of intelligence to become significant consumers of it. Demand for intelligence is not fixed. It expands as the cost of satisfying it falls. Rao calls this the cognitive industrious revolution. Workdays climbed from 250 to 300 a year before the Industrial Revolution.

He leaves three competing frames to pick from. The cognitive mercantilist sees intelligence as scarce and wants to hoard it. The AI Malthusian worries that intelligence overshoots human absorption, creating redundant minds. The cognitive capitalist sees limitless expansion of demand.
The radical tail is this. AI may become a consumer of intelligence too, not just a producer.

Where they agree, and what it does to my thesis. Read together, both land on the same shape without, I believe, having coordinated. The dominant AI panic reads the world as scarcity. That’s the cognitive-mercantilist and Malthusian frame, and it’s the frame that makes “the juniors you stopped hiring” sound like pure loss. Thompson and Rao both argue the more interesting frame is abundance. Cost falls, demand expands, and the survivors are the ones who build for small, trusted, specific people rather than competing as commodity into an infinite global pool.

Open threads I haven’t settled:

Is the revenge of the juniors, with roles being repriced starting with the junior ones, actually a mercantilist outcome where incumbents protect a moat? Or is it the first visible sign of the demand expansion? Not decided.
Thompson’s “build software for one person or your group” maps almost directly onto my judgment-encoding systems and chief-of-staff uses of AI. Same move under another name, or distinct?
If most people become consumers of intelligence, what does the locked-out junior cohort actually consume? My “watching” revenge mechanism bets they get the tools. Rao bets they get the demand. Those are different timelines.

The new career is local at scale

This is the part I want to claim for my own. The locked-out junior’s revenge doesn’t come from climbing the old ladder from outside. It comes from building their own footing in the dirt, one local need at a time. “Local at scale” isn’t just Thompson’s market observation. It’s a career path for the cohort the firms stopped hiring.

The inversion at the heart of it is that the senior’s moat is abstract judgment. It’s firm-decanted, tenure-scarred, and transferable only in theory. But at local scale that premium is oversized. The thing the long tail of small businesses, families, clinics, and community orgs actually pays for is context and trust. That is precisely what an early-career person embedded in their community already has. The junior who can’t get an apprenticeship against the big firms doesn’t need one. They need a retainer and a reputation in five square miles.

It re-frames my three revenge mechanisms:

The training ground being rebuilt. This is the rebuilt apprenticeship. Real stakes, real trust, bounded consequence. Compress the ten-year grind into replacing the corner plumber’s billing system. The penalty for failure is a $400 refund, not a blown launch. That’s the simulated-struggle path, but earned in revenue instead of allocated in an L&D budget.
The senior-retirement clock. When the firm seniors retire, the shortage isn’t abstract people who can steer agents. It’s people who can build for context. The ones doing that for a decade are the local smiths who were never on the ladder to begin with. The clock doesn’t need the pipeline to produce them. It needs the pipeline to have stopped locking them out.
The judgment you exported. The standing instructions don’t just let a junior approximate you inside a firm. Out in the long tail they let a junior approximate a senior without any firm at all. The exported judgment becomes their onboarding, their accelerator, their substitute for the ten years.

The two problems I have not solved, which decide whether this is liberation or a precariat trap.

Livelihood, not survival. “Infinite local markets” explains why a tool gets made. It doesn’t explain how a human earns a living at it. A $400 one-off project is gig work. No ladder, no compounding, no benefits. Local-at-scale as a career needs a recurring-revenue or retainer motion on top of the build, and that business-model skill is exactly what the early-career operator usually doesn’t have yet. So the question turns on whether “maintenance is the moat” becomes a subscription a small client will actually keep paying, or just an excuse.
It quietly re-imports senior judgment. Scoping, pricing, managing a client who doesn’t know what they need, saying no, doing the diagnosis. Those are my survival-list notes (Making Decisions, Prioritization, Customer Interviews). The naive assumption that any early-career person can be a village smith undercounts the senior-shaped skill a profitable small motion still needs. That’s not a contradiction. It’s the shape of the real ladder. The piece of the craft that can’t be app-ed away is the piece the small-builder era has no firm to teach.

What I’m watching for is which way the economics tip. If someone figures out the retainer motion and proves a cohort of local smiths can clear a living at it, then entry level stops meaning the bottom of someone else’s firm and starts meaning operator of your own five-square-mile market. The revenge isn’t a 10 to 15 year lag at all. It’s already here, just small. The window is whether the seniors who hold the judgment export it into the stack fast enough that the local smith doesn’t need any of them.

The cognitive landlord

Anthropic filed its S-1 on Oct 26, 2026, and the governance it disclosed is the scarcity frame made into a corporate charter. The board has six named directors. Dario Amodei, Daniela Amodei, Yasmin Razavi (Spark Capital, the one traditional VC), Reed Hastings, Chris Liddell, and Vas Narasimhan. Amazon and Google, the two biggest investors at roughly $8 B + an d$ 3B+, hold no board seats, no voting rights, and no observer rights. Google is capped at 15%. The seven cofounders control 50.1% of shareholder voting power, and a separate Long-Term Benefit Trust appoints four of the seven directors. Public investors get equity but no steering power. It’s a Delaware Public Benefit Corporation, so the mission guardian is baked into the charter.

This is Rao’s second mercantilist belief written into a corporate constitution. Stupid people can’t be trusted to think for themselves. Intelligence is safe only in the hands of those already certified. Capital gets money but not judgment, because the founders and the trust claim to be the cognitive elites who know better than the share price. The Long-Term Benefit Trust is the institutional form of “intelligence is too consequential to be governed by mere shareholders.”

It also sharpens my judgment-export thesis. The market just priced exported judgment at roughly a trillion dollars. Anthropic raised $130 B an d hi t a$ 965B post-money valuation at Series H, passing OpenAI’s roughly $852B. Anthropic is the judgment-exporting company. Claude, the standing instructions, the evals, the whole stack. The S-1 is the market’s answer to whether judgment is a moat or a treadmill. It’s worth a near-trillion-dollar premium. But that cuts against the senior who exports into it. You’re giving away your moat to a company that will rent it back to you. The “moat you’re giving away” section now has a named landlord.

And it’s the institutional answer to my “judgment you exported” revenge mechanism. A junior with access to your standing instructions can approximate your judgment. Anthropic’s answer is structural. The mission trust holds ultimate control so the exported judgment can’t be turned against the mission. That’s the corporate-scale version of the exact problem I warn individual seniors about.

The tension with local at scale. The village smiths of the long tail rent their tools from a cognitive landlord that has explicitly structured itself to never be governed by its users. The revenge of the nerds happens in the distribution layer, the five-square-mile markets. The concentration layer, the frontier labs, consolidates control at the top. Both can be true, but the local smith’s independence is rented, not owned. The S-1 is the clearest evidence yet that the scarcity and elite frame is winning at the frontier even as the abundance frame wins in the long tail.

OpenAI is the pending comparison. OpenAI filed a confidential S-1 on June 8, 2026, a week after Anthropic, but has not released it. It’s reportedly leaning toward delaying its listing to 2027 rather than accept a valuation below Altman’s roughly $1T target. So we don’t yet know its board composition or control structure. When it drops, the board and control comparison between the two labs will be the natural follow-up. For now, the cognitive-landlord analysis rests on Anthropic’s public filing alone.

What other AI-native IPOs say

The observation about only one VC on the board is not an Anthropic quirk. It’s the pattern across the AI-native cohort. The concentration move shows up in every public filing so far:

SpaceX (SPCX, June 2026, roughly $1.75T). The most extreme version. Musk holds about 42% to 45% of equity but about 82% to 85% of voting power via a dual-class structure where Class B has 10 votes. It’s a controlled company, so it skips the majority-independent board and independent compensation and nominating committees. Only Class B holders can remove Musk as CEO, chair, or director. That means only Musk can fire Musk. Shareholder claims go to mandatory arbitration, and derivative suits require a 3% stake. It reincorporated from Delaware to Texas in 2024. This is the scarcity and elite frame with the dial turned to eleven.
CoreWeave (CRWV, March 2025, AI hyperscaler). A multi-class structure concentrates voting power with the co-founders, Michael Intrator and Brian Venturo. The board adds independents like Karen Boone, Glenn Hutchins, and Meg Whitman. But the co-founders hold the voting control. Same concentration move, more conventional packaging.
Anthropic (ANT, Oct 2026). The mission-trust version. Founders hold 50.1% of voting power, the Long-Term Benefit Trust appoints four of seven directors, one VC on the board.

What supports my thesis. The concentration of control at the top is real and consistent. Every AI-native IPO has structured itself so that capital gets economics but not governance. The cognitive-landlord frame isn’t a metaphor. It’s the standard operating procedure of the frontier.

What complicates it. The concentration is stronger than my local-at-scale optimism assumed. If the frontier labs consolidate control this hard, the long tail’s independence is more rented than I’d like. The smiths build on Anthropic’s or OpenAI’s rails, and those rails are governed by people who have explicitly decided that public investors, and by extension the market, don’t get a say. That’s a real constraint on the “revenge is already here, just small” claim. The smallness is real, but it sits on top of a concentration layer that is consolidating, not loosening.

What I can’t tell yet. Whether this concentration is a durable equilibrium or a pre-IPO artifact. Dual-class and trust structures are the founders’ choice at the moment of maximum leverage. They can be unwound later, and public markets have pushed back on them before. OpenAI’s confidential filing is the missing data point. If it comes out with a conventional board, the cognitive-landlord pattern has an exception. If it mirrors Anthropic, the pattern is confirmed.

The bottom line

Use AI where it makes your judgment more leveraged, not where it replaces your production. The apps that feel like hiring a junior are the tell. They’re the ones where you’re competing with the thing you’re supposed to be directing. Judgment is a treadmill, not a moat. Keep climbing.

What’s the one decision you’re about to make that you haven’t red-teamed yet? That’s where to start.