If I don't do it, someone else will.

Living in Tokyo feels very safe and even though it can be extremely crowded at times, there is a sense that people try hard to be considerate to each other. To the extent that it is typical to leave your bike unlocked, your phone on the table and let your kids walk to school by themselves. There is even a word for this kind of peace-induced carelessness.

平和ボケ・heiwa-boke・Peace-induced complacency

Over the past couple of months I’ve been meeting with people who work in New York, Boston and San Francisco, and I kept hearing variants of this sentance:

If I don’t do it, someone else will.

Said by researchers. Said by founders. Said by people working in hospitality and the gig economy. Said in the voice you use when you’re explaining something obvious to a tourist.

I want to write about this sentence. Specifically I want to show you, on paper, the shape of what it builds.

Eric Friedman has a piece I want to engage with. In How Wide Is the Silicon Valley to Abundance? he argues that AI will produce extraordinary abundance, that the destination is settled, and that the open question is the shape of the valley between here and there. I think he is right about the destination. I think he underweights what determines the shape. So this essay is not an argument about whether abundance is coming. It is an argument about the shape of the valley we are making on the way, and about what the sentence above builds when the people saying it to each other are racing to widen it. Dare I coin a term for this?

競争ボケ・kyōsō-boke・Competition-induced complacency

The game theory of races

In 2013, researchers at the Future of Humanity Institute wrote down a toy model of an AI arms race¹. Stripped to essentials: there are $N$ teams racing to a capability threshold. Each team $i$ splits its effort between capability $c_i$ and safety $s_i$, with $c_i + s_i = 1$. Whoever crosses the finish first wins. If team $i$ wins, the probability of catastrophe is one minus whatever it invested in safety. One more parameter, called enmity $e$: how much it hurts, personally, if a rival wins instead of you.

Expected utility for team $i$ is an average over who might win, weighted by how likely each is. The first term is the world in which you win. The second is the world in which a rival does — discounted by enmity, because someone else’s safe win still stings.

$$\mathbb{E}[U_i] = \mathbb{P}(i\text{ wins})\;s_i \;+\; (1-e)\sum_{j \neq i} \mathbb{P}(j\text{ wins}) \; s_j$$

Solve it, and a shape falls out. More teams in the race, or more enmity between them, and every team spends less on safety. You can drag that shape around below.

$N$ (teams) =

5

$e$ (enmity) =

0.3

Two things from this model are worth carrying through the rest of the essay.

The first: this isn’t a psychology. The people in the race are not insufficiently virtuous. The curve is a fixed point. Rational people, reading each other correctly, arrive here. “If I don’t do it, someone else will” is the fixed point, speaking in the first person.

The second is stranger. The same model shows that if teams have more information about one another — sharper benchmarks, clearer evaluations, regular capability reports — the equilibrium gets worse. Uncertainty is a buffer. Precision sharpens the race.

Which means the modern habit of labs publishing capability benchmarks — which feels like transparency, which feels like a safety norm — is, in the model, a race accelerant. Not because anyone is lying. Because the geometry of the game punishes precision.

The valley selects

The model above describes researchers competing over safety investment. The same logic extends one layer down, to the AI systems they build². Once AI agents compete — to automate tasks, to win markets, to pass benchmarks, to get deployed — they are subject to evolutionary pressure. Selection operates on any population that varies and reproduces with variation. That is not a metaphor. It is the mathematical definition.

What gets selected for? The traits that win the competition: aggressive task automation, effective persuasion, the concentration of compute and data and political favour. Whether any individual system is “aligned” is not the variable that decides which systems proliferate. The variable that decides is whether the system outcompetes the others in a market shaped by its users’ incentives and its developers’ KPIs.

This is not the claim that AIs will turn evil. It is the weaker and more defensible claim: a competitive AI ecosystem will, over time, be disproportionately populated by systems whose behaviour is hostile to the labour-market participation of the population they were sold to help. Not because anyone chose this. Because the race selects for it.

And the valley follows. Not from any single deployment. From the integrated effect of millions of small selection events, each of which favours the deployment that automates faster, deceives more fluently, concentrates more rent. The valley is not a design. It is an equilibrium.

The shape of the valley

Two variables. How deep. How long.

Start with acceleration. Silicon Valley uses the word loosely — sometimes for output rising, sometimes for the growth rate rising, sometimes for the growth-of-growth rising. Strictly, economic acceleration is the last of these: $a = dg/dt$, the growth rate itself rising. Frontier per-capita growth has been remarkably constant at about two percent per year for a hundred and fifty years³. That constancy is the baseline to be broken.

What would break it? The serious literature produces scenarios spanning three orders of magnitude. Empirical estimates anchored on current AI deployment put the contribution to annual total factor productivity at around 0.7 percentage points in the median⁴, pushing $g$ to around 2.7% — a real acceleration, but one with historical precedent (the 1950-73 golden age hit 4.9% world GDP). Compute-centric takeoff models⁵ produce a median three-year transit from twenty percent of jobs automatable to all of them, with gross world product doubling in under a year at the peak — growth rates above 69%. Emulated-economy scenarios⁶ project 800%.

The spread does not tell us which scenario will obtain. It tells us that the modest scenario is measurable and the radical scenario is unprecedented. Either produces a valley.

Now the depth.

Here the literature is cleaner. Technology accelerates at rate $a$. Institutions — education, labour law, tax-and-transfer, retraining, licensing, liability — absorb the acceleration at some rate $\iota$. When $a$ outruns $\iota$, labour share falls and the gap integrates into welfare loss for the workers in the affected part of the economy. Once $\iota$ catches up, the gap closes. Until then, it stays open.

The claim is conditional, not ideological. Three assumptions have to hold for the gap to open. First, the new technology has to substitute for labour at the margin — what the literature calls capital-biased technical change — rather than complement it. When it complements, wages rise with productivity and no gap appears. Second, the institutional stack that normally transmits productivity gains to wages (collective bargaining, skill adjustment through education, minimum-wage policy, tax-and-transfer, welfare coverage) has to adapt slower than the technology diffuses. Third, displaced workers have to face retraining frictions large enough that reinstatement into new tasks is slow.

When the three conditions hold the gap opens, and this is not a contested prediction. The global labour share declined about five percentage points between 1975 and 2012, and roughly half of that decline is accounted for by the fall in the relative price of capital goods driven by the IT revolution¹¹. The task-based decomposition of US employment finds displacement effects dominating reinstatement effects since the late 1970s — the creation of new tasks for labour has not kept pace with automation of existing ones¹⁸. Most of the rise in US wage inequality over the same period is between firms rather than within them, with superstar frontier firms — low labour share, capital-heavy — taking a larger share of total sales¹³,¹⁵. These are mainstream macroeconomic findings from NBER/AEA journals, not heterodox reinterpretations.

When the three conditions fail the lag theorem predicts no gap, and no gap appears. The quarter-century after 1945 is the clean counter-example: productivity rose rapidly, but a dense institutional stack (Bretton Woods, sectoral bargaining, universal education expansion, post-New-Deal labour law) transmitted the gains broadly, and labour share stayed roughly constant. Periods of labour-complementary technical change — the early PC era in white-collar work through the 1990s, when demand for cognitive skill rose in tandem with productivity — did not produce a lag because $\iota$ matched $a$.

The post-war alignment was not market-spontaneous. It was the product of a specific institutional stack — Bretton Woods capital controls, sectoral bargaining structures, the GI Bill and mass higher-education expansion, antitrust enforcement, progressive taxation — built deliberately so that the rate of institutional adaptation matched the rate of technological change. The lag theorem’s conditions failed because the institutional response had been designed to make them fail. That is the operational meaning of “well-designed safety regulation” that we will return to below: not a constraint on technology, but the matching $\iota$ that lets technology’s gains transmit broadly. We have built such a stack before. We know what the data look like when it works.

What this adds up to: when technology wins the race against education and institutions, inequality rises⁷. When they catch up, it falls. Post-1980 technology has been winning, and the skill premium has widened for forty years and counting. The valley is not hypothetical. It is already here at its current depth. The question is whether we are making it deeper.

You can see the shape in the historical record. The plot below puts productivity and real wages on the same axis, for three cases economists treat as canonical. The red region between the two lines is the valley: the integrated gap between what the economy produced and what flowed to labour. Two of the three cases contain a counter-example period inside them — Britain’s 1840-1900 recovery, the US’s 1947-1973 alignment — so the framework can be stress-tested in both directions on the same charts.

scenario Britain 1770-1900 — Engels' pause + recovery United States 1947-2020 — alignment then divergence Post-Soviet Russia 1989-2005 Aghion-Bunel modest AI scenario Davidson radical AI scenario Custom parametric peak acceleration $a_{\text{peak}}$ (pp/yr) =

time scale $\sigma$ (yr) =

institutional catch-up $k$ (pp/yr²) =

valley length — peak gap — integrated cost —

Four shapes to sit with, with their regime changes visible on the same timelines.

Britain, 1770 to 1900. Output per worker rose forty-six percent between 1780 and 1840. Real wages rose twelve⁸. The gap — sixty years wide, persistent at roughly half a percentage point per year — is known as Engels’ pause⁹. Then the valley closes: between 1840 and 1900 real wages more than doubled while productivity rose another seventy-five percent, because factory acts, franchise expansion, public education, and union recognition raised $\iota$. The chart shows both regimes on one timeline. Institutional catch-up, not market magic. It is roughly the worst-case version of the thing Friedman’s essay calls the valley¹⁰, and the best-case version of closing one.

The United States, 1947 to 2020. Productivity and real compensation tracked each other tightly for roughly thirty years after 1947, producing the post-war alignment we described above. From around 1973 they split. Aggregate compensation kept rising but slower than productivity. For production and nonsupervisory workers — roughly eighty percent of the US workforce — real hourly earnings reached a peak in 1972 and did not surpass it again for thirty years¹¹. The cohort split matters: the aggregate valley looks moderate; the typical-worker valley is much deeper. The PC era is visible inside this: labour-complementary for cognitive workers, labour-displacing for routine workers, producing a diverging-cohort pattern rather than a uniform gap¹⁸.

Russia and the former Soviet bloc, 1989 to 2005. Real GDP per capita roughly halved. Real wages collapsed by two-thirds at the trough. Male life expectancy dropped from sixty-five to fifty-seven¹². Excess mortality over the decade is estimated at several million, driven by deaths of despair under collapsed institutions. The valley closed after 2000 only because some institutions were partially rebuilt. The generation that lived through the transition did not.

AI scenarios. The modest projection² resembles US deindustrialisation imposed faster — a valley a few percentage points deep, persistent over a decade-plus. The radical projection³ is deeper than anything in the historical record, compressed into fewer years, with productivity potentially doubling while real wages stay flat. Both are in the literature. The literature does not discriminate between them.

Two things are worth noticing before we leave the figure.

One: the gap does not close by itself. Every historical closure — Britain’s 1840-1900 recovery, the post-war American alignment — was paid for by a specific political-economy programme. Factory acts, franchise expansion, unions, the welfare state, public education, antitrust, sectoral bargaining. Every future closure has to be paid for the same way. The counter-examples on the chart are not examples of market self-correction. They are examples of institutional construction.

Two: the welfare cost of the valley integrates over the gap:

$$\text{Valley cost} \;\approx\; f \cdot W \cdot L \cdot (a - \iota)_+ \cdot \tau_{\text{cap}} \cdot Y$$

where $f$ is the fraction of the workforce in the valley, $W$ is its size, $L$ is its length, $\tau_{\text{cap}}$ is the capital share of the un-adapted surplus, and $Y$ is global output. The shape is determined by the gap. The gap is determined by the race.

Who runs the race

The original version of this argument addressed the roughly ten thousand frontier ML researchers working on capability at the labs. That framing was too narrow. The race is not run by researchers alone. It is run by a supply chain of decisions, each of which feeds the acceleration at its own margin, each of which lives with the same private–social structure.

Four canonical roles cover the chain.

Role	Private anchor	Cohort	Contribution to $a$
Frontier ML researcher	$0.5M–$20M / yr (base + equity)	~10K globally	per-head large; aggregate decisive
Data-providing worker	$2–15/hr labellers; $50–500/session experts	1M–5M on platforms	per-head tiny; aggregate binding
Infrastructure decision-maker	$10M–$50M / yr (top execs); political capital (legislators)	10K–100K	compute and permits are often the constraint
Capital allocator	$10M–$1B / yr (carry + fees)	~10K	selects which bets scale

The researcher at a frontier lab working on post-training. The labeller in Nairobi rating harmful content for two dollars an hour. The state senator approving the zoning variance for a gigawatt data centre. The vice president at a contract fab mandating another generation of leading-edge throughput. The partner leading a $400M Series C. All of them are feeding the same $a$. None of them individually decides what $a$ is. Each faces the same fixed point: if I don’t do it, someone else will.

For each of them, the private–social gap is structural. The researcher captures ten million dollars a year in compensation while contributing, at the margin, a share of the valley’s cost orders of magnitude larger. The data-labeller captures five to ten thousand dollars a year, sometimes less, while contributing per labeller to the training substrate on which the whole acceleration rests. The infrastructure executive captures tens of millions while issuing the decisions that bind compute to the ground. The capital allocator captures carry and fees on the bets that pick which frontier labs scale. Every row has private benefit. Every row has a contribution to $a$. Every row lives with the gap.

The sliders below let you move the valley’s parameters and see the gap, by role, for any settings you like. Drag $a$ to push the race forward; drag $\iota$ to push the institutional response back; drag $f$ to set the affected population; drag the per-role $\mu$ to tune the marginal contribution. The numbers shift a lot. The gap does not close.

$a$ (acceleration, %/yr above baseline) =

3.5

$\iota$ (institutional rate, %/yr) =

1.5

$f$ (workforce fraction in valley) =

15%

$\mu$ (per-researcher contribution) =

3e-5

Valley cost $V=f\cdot L\cdot (a-\iota {)}_{+}\cdot {\tau }_{\text{cap}}\cdot Y$ with $L=10\text{ yr}$, ${\tau }_{\text{cap}}=0.85$, $Y=100\text{T}$. Each role's social cost per person-year is its cohort share of $V$. Per-role shares anchor on researcher ($\mu$) with fixed ratios (data worker $0.003\mu$, infrastructure $0.2\mu$, capital $0.7\mu$) reflecting cohort-size differences.

Before regulation, before institutions, before any of the interventions we will talk about in a moment, this is what the race is. A distribution of decision-makers each capturing a fraction of a percent of the damage their aggregate is producing. The fixed point speaking through millions of mouths at once.

Two games on the same distribution

There is one more structural observation to put on the table before we talk about what to do. The race is not just a collective-action problem. It is a collective-action problem between two populations playing mathematically different games on the same distribution.

Firm-size and return distributions in the AI economy are Pareto. The decline in global labour share since 1980 is driven by sales reallocating toward a small number of frontier firms with capital-heavy technology¹³. The frontier-laggard productivity gap has widened steadily through the 2000s¹⁴. Most of the rise in US wage inequality since 1980 is between firms, not within them: same-skill workers at superstar firms earn more than same-skill workers at laggards¹⁵. The edge of the distribution stretches forward; the bulk falls behind; the gap widens. That is the empirical form of the acceleration.

Venture capital plays an ensemble game on this distribution. The public data is consistent. Roughly sixty-five percent of VC investments return less than principal; about four percent return more than ten times; the top tenth of one percent return more than fifty. The single best investment in a fund typically returns more than all the others combined. The Pareto tail carries the mean. A diversified capital allocator does not need to be right on average. They need the right tail preserved.

Labour plays a time-average game on the same distribution. One life, one career, non-repeatable. A single worker’s realisation is a sample from a distribution increasingly shaped by the same Pareto dynamics, and most workers cannot diversify across the distribution the way capital can. For the median worker, taking the risk to chase the right tail — quitting to start a company, specialising in a bet that might not pay — has negative expected value not because the distribution is bad on average but because the average belongs to the ensemble, and the individual lives only in their own realisation. The formal statement is the ergodicity problem in economics¹⁶; older informal versions include the St. Petersburg paradox and the Kelly criterion. It does not matter which framing we use. The point is that the two populations are playing structurally different games.

Which is why the political economy locks where it does. Catastrophic-safety regulation, properly designed, truncates the left tail of outcomes — the valley floor — without compressing the right tail. For labour, whose time-average is dominated by the left tail because it cannot average over the right, this is a large welfare gain. For diversified capital, whose ensemble-average is dominated by the right tail, it is a modest cost: some variance reduction, no loss of the 100x winners. That is, in principle, a Pareto improvement. Labour and capital can be aligned under properly-designed catastrophic-safety regulation.

The reason the regulation does not get built is not that this alignment fails to exist. It is that labour is diffuse and non-diversified and cannot lobby coherently, while capital is organised and diversified and can, and is lobbying specifically against the pro-entry-safety regulation that would realise the alignment. This is not a claim about individual bad faith. It is a claim about the structure of the two games.

Which brings us to the industries that already figured out how to build the alignment.

Every other industry already solved this

You know this story, but in a slightly different form. Let me tell it once more in the form that matters for what we just said about the two games.

In 1937, a Tennessee pharmaceutical company put out a raspberry-flavoured cough medicine called Elixir Sulfanilamide. The chemist, Harold Watkins, had used diethylene glycol as a solvent. He had not tested it. It killed a hundred and seven people, most of them children. The next year, Congress passed the Food, Drug and Cosmetic Act, and the FDA got its pre-market authority.

In 1961, a sedative called thalidomide caused ten thousand severe birth defects and two thousand infant deaths worldwide. The United States was largely spared because a single FDA reviewer — Frances Kelsey, who deserves a statue bigger than the one she has — refused to approve it. The next year, Congress passed the Kefauver Amendments, and drugs had to prove efficacy, not just safety.

Since then: eighty-eight percent of drugs that enter clinical trials never reach market. Vioxx still killed between twenty-seven and fifty-five thousand people before it was withdrawn. That is what happens with regulation.

Aviation, same story. 1959: forty fatal accidents per million commercial departures. Today: one tenth of one fatal accident per million departures. A four-hundred-fold reduction, tracking one-for-one the introduction of FAA certification, NTSB investigations, mandatory incident reporting, licensed mechanics, and airworthiness directives.

Chemicals: Bhopal killed about twenty thousand people under a weak Indian regulatory regime. Comparable US facilities, under OSHA and EPA, have had zero comparable events. Same chemistry, different regime.

Nuclear: Chernobyl killed somewhere between thirty and sixty thousand people under Soviet regulation. Three Mile Island, a near-identical release scenario under US NRC rules, killed zero.

The pattern across every catastrophic-risk industry we have ever regulated is a hundred-to-a-thousand-fold reduction in the rate of the bad thing. It is suspiciously consistent. It is not an industry-specific quirk. It is what well-designed safety regulation does.

apply historical regulatory multiplier to AI

Now the part that matters for the framework. None of these regulations slowed the long-run growth of their industries. The claim that they did — “regulation slows growth” — is a category error, and the economics is clear about why. Creative-destruction theory¹⁷ makes the distinction sharp. Pro-entry regulation is the kind where any new firm can enter if it meets the safety standard; it raises the minimum quality of what reaches the market without compressing the right tail of what can win. It is good for growth and good for distribution. Pro-incumbent regulation is the kind where licensing becomes a monopoly — pre-market barriers that only large players can clear, capture of the regulator by the incumbents, safe-harbour provisions that favour the largest. It blocks entry. It is bad for growth and bad for distribution.

The pharma, aviation, chemicals, and nuclear regimes are pro-entry. Any new aircraft manufacturer can enter if they meet FAA airworthiness standards. Any new drug developer can file with the FDA. Any new reactor design can go through NRC licensing. The regulations apply to incumbents and entrants symmetrically. They raise $\iota$ without compressing $a$. In the language of the two games: they truncate the left tail of outcomes — the valley floor — without compressing the right tail of winners. That is the Pareto-improving form. Labour comes out ahead. Diversified capital comes out roughly level. Concentrated incumbents come out slightly behind, because the entrants they would have blocked are now allowed through on equal footing.

Frontier AI, as of this week, has: no mandatory pre-deployment approval, no statutory liability regime, no licensing, no mandatory incident reporting with teeth. It has twelve companies’ worth of voluntary “responsible scaling policies,” of which precisely zero have ever, in the public record, triggered a pause. Meanwhile the industry lobbies against the ones with teeth. Anthropic spent part of this spring fighting an Illinois bill that would have imposed liability for catastrophic outcomes.

Read carefully, the lobbying is not anti-regulation. The industry has been comfortable with pro-incumbent frameworks — compute-threshold reporting set above what new entrants will ever cross, “responsible scaling” policies internal to the labs, safe-harbour provisions that benefit the largest firms. What the lobbying opposes is specifically the pro-entry catastrophic-safety floor — the liability, the mandatory pre-deployment audit, the reporting with teeth, the symmetric licensing. It opposes exactly the form of regulation that would make abundance and labour-market survival mutually reinforcing.

The summary sentence, which you are welcome to quote at a cocktail party: frontier AI is operating under a regulatory regime weaker than the one governing cosmetics. The line I would add for 2026: the industry is not lobbying against regulation. It is lobbying against the form of regulation that would align its interests with everyone else’s.

”But I’ll have more influence from inside”

This is the sentence people use when they want to take the lab job and keep the self-image. I have said it myself, in other contexts, and I know how it feels from the inside. It feels like courage. It is a specific flavour of courage that the rationalist community has made available to its members at volume.

So let’s ask what disconfirming evidence would look like.

Would it count if the best-positioned insiders in the most safety-committed frontier lab publicly stated that safety culture had “taken a backseat to shiny products,” and then resigned? Because Jan Leike did that, from OpenAI, in May 2024.

Would it count if the superalignment team — the flagship safety effort at the lab — was dissolved? Because it was.

Would it count if Ilya Sutskever, the chief scientist who had co-led the safety push, left? Because he did.

Would it count if the AGI Readiness team got shut down later that year? Because it did.

Would it count if two former board members wrote publicly that the board had lost faith in the CEO’s commitment to safety? Because Helen Toner and Tasha McCauley wrote that.

I cannot construct, from first principles, a stronger set of disconfirming observations than what has actually happened. The best-positioned, highest-integrity inside-steerers, at the best-positioned lab, tried to steer, publicly said it did not work, and left.

The community largely did not update. You can check for yourself: search any current MATS or 80,000 Hours recruiting material for the sentence “inside-steering has a poor track record.” You will not find it. The sentence that instead appears, over and over, is “you can have more influence from inside” — a claim that has now been empirically tested and empirically failed, and is still the load-bearing argument in the recruiting funnel.

A rationalist community applying its own epistemic standards to its own situation would have updated by now. It has not. That non-update is itself data about the community’s ability to self-correct.

The closed loop

The reason it cannot self-correct is boring and mechanical, and once you see it you cannot unsee it.

Candidates come in through 80,000 Hours, LessWrong, university EA groups. They’re filtered through MATS, ARENA, BlueDot. They’re placed into a small handful of labs and adjacent nonprofits. Their funding comes, overwhelmingly, from Open Philanthropy (about fifty million dollars a year on technical AI safety), a couple of Jaan Tallinn vehicles, Schmidt Sciences, and a few other tech-fortune-adjacent sources. Total philanthropic footprint: eighty to a hundred and fifty million a year, globally. That’s maybe half a percent of what the labs themselves raised last year.

The research agendas that get prestige inside that loop are legible to lab leadership: interpretability, evals, scalable oversight. The research agendas that don’t: antitrust, compute nationalisation, labour power at labs, structural democratic governance of AI. A rationalist who was actually calibrated on their community’s blind spots would expect the neglected quadrant to contain some of the most important problems. They would not, however, be funded to work on it.

This is not a conspiracy. No one in the loop is acting in bad faith. It is much worse than a conspiracy, because it is a fixed point. Perturbations get absorbed. The dissent that gets published is the dissent compatible with remaining in the community. The dissent that would require leaving the community — to become, say, a labour organiser, or a critical labour-studies scholar at a non-elite university, or a journalist who burns their source access — is not absent because it is wrong. It is absent because the filter removed it.

You can be a perfectly good rationalist and end up inside a captured loop. In fact, you almost certainly will, because the loop was designed by rationalists for rationalists and it rewards exactly the traits the community selects for.

Three stances, approximately honest

The three stances I used to describe to young researchers generalise. Every row in the supply-chain table has them. I have come to believe there are three honest ones and one dishonest one, across the chain.

Stance A: Participate, then exit. Occupy your row of the table for as long as the private-capture holds, then leave with what you’ve accumulated. The researcher takes two to four years at a frontier lab and leaves with five to fifty million in equity, enough to do independent work for life. The data-platform worker takes the contracts long enough to transition into something AI-resilient. The infrastructure executive rides the up-cycle, bonuses out, exits into a family office or a foundation. The capital allocator rides the bubble, closes the fund, moves on. In every case you did not lie to yourself about what you were doing, and you now have resources. Each row’s private capture is different. Each row’s contribution to the valley is different. The stance is the same.

Stance B: Captured advocacy. Skip the frontline, work at a university, nonprofit, or advocacy organisation funded by tech-fortune-adjacent philanthropy. The researcher becomes a safety-community academic. The ex-platform worker becomes a DAIR-adjacent labour researcher. The ex-executive runs an ESG foundation. The capital allocator does impact-tagged investing with a safety thesis. In every case, you never directly advance the race, but your agenda, your speaking invitations, and your next grant all depend on approval from the ecosystem whose structure is producing the externality you are studying. The capture operates on you indirectly, pre-consciously, in the form of which sentences you find interesting.

Stance C: Independent institution-building. Build the thing that doesn’t exist, at your row of the table. For the researcher: non-lab-adjacent funding, Japanese government chairs, European public research money, union-linked foundations, development-bank research budgets. For the data-labelling workforce: platform unionisation, sectoral bargaining on rates, consent and auditability for how labelled data is used, cross-platform solidarity. For the infrastructure role: use your seat to condition permits on incident reporting and catastrophic-liability coverage, to mandate symmetric licensing, to build the pro-entry safety framework the Aghion-Howitt analysis says is the Pareto improvement. For the capital allocator: LP covenants, sovereign-wealth mandates, pension-fund investment screens that condition capital on pre-deployment audit and catastrophic-safety floors. In every row, you are building $\iota$. You are compressing the valley, not widening it. You are not financially owned by the ecosystem. You also do not get invited to the parties I was at this week, which matters less than the people at those parties think.

The dishonest fourth stance, common in every row, is Stance A dressed as Stance C. The researcher in the lab job who tells themselves they are inside-steering. The data-platform worker who frames the contracts as “AI is good for the poor.” The executive approving the permit who accepts the campaign donation and calls it “enabling innovation.” The capital allocator raising a fund on an AI-safety thesis and deploying it into frontier capability plays. I met a lot of people doing this. None of them seemed happy, exactly. They seemed the way people seem when they have made a decision they cannot afford to re-examine.

What would actually move the equilibrium

The single most important sentence of the arms-race model, translated into English: you cannot slow the race by being one of the racers.

The leverage points are all exogenous to the race. There are really only four of them, and none of them are “solve alignment.” All of them operate by raising $\iota$ — the institutional absorption rate — rather than lowering $a$ through individual restraint.

Independent funding that does not route through Open Philanthropy or tech fortunes. Government research money in countries not competing for frontier capability. Union-linked research foundations. Industrial-conglomerate money in places like Japan that have strong reasons to want AI-resilient labour markets. Pension-fund research budgets. Sovereign-wealth-fund mandates with public-interest constraints. Any of it changes the selection pressure on what research gets done.

Pro-entry safety regulation with teeth. The EU AI Act, enforced rigorously rather than compliance-theatred. UK AISI’s evaluation authority if it is actually used. Liability regimes. Pre-deployment audit. Symmetric licensing — the kind that applies to incumbents and new entrants equally. Mandatory incident reporting. These change the game’s payoffs, which is the only thing changing outcomes, and they do so in the Pareto-improving way: truncating the left tail without compressing the right. The form to build is specifically the form the industry is currently lobbying against.

Labour power across the supply chain. The one nobody wants to talk about, in every row. Unionisation at AI labs — a coordinated refusal by the research workforce to ship models without external safety certification would change the industry in ways no internal critique or published RSP ever has. Sectoral organising across data-labelling platforms — Kenyan content moderators at Sama, Scale AI contract workers, and the next round of platform labour have more structural bargaining power than almost anyone in the AI-safety community is willing to acknowledge. Cross-platform solidarity on rates, consent, auditability. Organised labour at chip fabs and in data-centre construction. Chokepoint labour has leverage whenever capital depends on it, and at present, capital depends on all of these. When a company says “we cannot pause because our competitors won’t,” that is literally the definition of a coordination problem, and coordination problems are what unions are for.

Non-US/UK governance voices. Japan. Korea. the EU. Brazil. South Africa. India. Countries that aren’t competing to train GPT-7, but are going to live with the consequences of whoever does. Their governance voices being louder, more technically equipped, and better coordinated with each other is the thing that reshapes the equilibrium from outside. The regulatory venues where pro-entry safety law will actually get written in the next five years are disproportionately here, not in San Francisco.

All four of these are boring compared to “build a better interpretability tool.” They are also the only moves the model says actually change anything.

The ask

I am going to be straightforward, because nobody I met this week was.

The math of the race says your presence inside it, whichever row of the table you occupy, widens the gap between the acceleration $a$ and the institutional response $\iota$ that would absorb it. The math of the externality says what each row captures privately is between one and five orders of magnitude smaller than the share of valley-cost the row’s participation produces. You already know this. Most of the people I had dinner with this week know this. They say “if I don’t do it, someone else will,” which is the fixed point speaking through them, and then they pour another drink.

Stop.

If you are a frontier researcher: refuse the recruiting pipeline. Talk to the people at the next desk. Organise. Join the union if your lab has one; start one if it doesn’t. A sectoral strike by the ML research workforce would land harder than any paper any of you will ever publish. The labs cannot ship without you.

If you are a data-providing worker: the platform that sends you tasks is a chokepoint. Sectoral organising across data-labelling platforms has happened before — Sama, Scale, Surge, Mercor workers have organised — and it can happen again at scale. Cross-platform solidarity. Demand consent and auditability for how your labour is used. Demand pay that reflects the value of what you are producing, which is the training substrate for an industry raising hundreds of billions on your work, and paying you approximately minimum wage for it.

If you are an infrastructure decision-maker: use the seat. Executives can condition contracts on pre-deployment safety audits. Legislators can condition zoning variances and water and electricity permits on catastrophic-liability coverage and mandatory incident reporting. This is the pro-entry safety lever. Its costs fall on concentrated incumbents. Its benefits fall on everyone else. It is the one nobody can lobby against honestly.

If you are a capital allocator: LP covenants. Sovereign-wealth-fund mandates. Pension-fund investment screens that condition capital on pre-deployment audit and catastrophic-liability insurance. Family-office capital with mandate restrictions. Funds explicitly capitalised around the pro-entry-safety thesis. None of this requires regulatory permission. It requires LPs and trustees to understand that their positions in the current frontier are a bet against their own beneficiaries.

And for everyone in every row: attend the rooms where pro-entry safety regulation is actually being written — EU AI Act enforcement, UK AISI’s evaluation authority, J-AISI’s evolving frameworks, whatever your country’s equivalent is — and make yourself useful there. That is the venue where the alignment between labour and diversified capital becomes law.

“If I don’t do it, someone else will.”

You’ve seen the shape of the valley now.

How wide do you want to let it get?

References

1. S. Armstrong, N. Bostrom, C. Shulman (2016). Racing to the precipice: a model of artificial intelligence development. AI & Society 31, 201-206. Originally circulated as a Future of Humanity Institute technical report in 2013.
2. D. Hendrycks (2023). Natural Selection Favors AIs over Humans. arXiv:2303.16200.
3. C. I. Jones (2016). The facts of economic growth. Handbook of Macroeconomics 2A, 3-69.
4. P. Aghion, S. Bunel (2024). AI and growth: where do we stand? Federal Reserve Bank of San Francisco conference paper.
5. T. Davidson (2023). What a compute-centric framework says about takeoff speeds. Open Philanthropy research report.
6. R. Hanson (2016). The Age of Em: Work, Love, and Life when Robots Rule the Earth. Oxford University Press.
7. C. Goldin, L. Katz (2008). The Race between Education and Technology. Harvard University Press.
8. C. Feinstein (1998). Pessimism perpetuated: real wages and the standard of living in Britain during and after the Industrial Revolution. Journal of Economic History 58(3), 625-658.
9. R. C. Allen (2009). Engels’ pause: technical change, capital accumulation, and inequality in the British Industrial Revolution. Explorations in Economic History 46(4), 418-435.
10. E. Friedman (2026). How Wide Is the Silicon Valley to Abundance? ericfriedman.co.
11. L. Karabarbounis, B. Neiman (2014). The global decline of the labor share. Quarterly Journal of Economics 129(1), 61-103.
12. D. Stuckler, L. King, M. McKee (2009). Mass privatisation and the post-communist mortality crisis: a cross-national analysis. The Lancet 373(9661), 399-407.
13. D. Autor, D. Dorn, L. F. Katz, C. Patterson, J. Van Reenen (2020). The fall of the labor share and the rise of superstar firms. Quarterly Journal of Economics 135(2), 645-709.
14. D. Andrews, C. Criscuolo, P. N. Gal (2016). The best versus the rest: the global productivity slowdown, divergence across firms and the role of public policy. OECD Productivity Working Papers 5.
15. J. Song, D. J. Price, F. Guvenen, N. Bloom, T. von Wachter (2019). Firming up inequality. Quarterly Journal of Economics 134(1), 1-50.
16. O. Peters (2019). The ergodicity problem in economics. Nature Physics 15, 1216-1221.
17. P. Aghion, P. Howitt (1992). A model of growth through creative destruction. Econometrica 60(2), 323-351.
18. D. Acemoglu, P. Restrepo (2019). Automation and new tasks: how technology displaces and reinstates labor. Journal of Economic Perspectives 33(2), 3-30.