And AI tax proposals are being proposed that aim to deal with the revolution risks, yet they all currently focus on limited geography and miss one real risk - that the consumption endpoint is already leaking.

In reality, we do not know which side will be right. And the net result is that we’re betting on the abundance future without ever negotiating terms for if it goes wrong. There is a better way for both sides to manage this risk - the name for that type of bet is “insurance”. And there is a policy where everyone can win.

If the AI abundance story is right, then the policy I’m about to describe should barely activate. It should cost the industry little or nothing beyond reporting that they’re already working on.

But if the other scenario starts to play out, then the protections should activate automatically.

That is the bargain that I’m proposing. A policy, where if we do end up with an economic revolution (not just evolution) then AI funds it in a sustainable way.

If you really believe the AI abundance story, then this should be a deal that you’d happily commit to. And if you really believe the jobpocalypse story then you should be equally happy with this type of deal. Either way, this should provide some certainty and let both sides move forward together in a productive way. This might even help with the growing unpopularity that AI is showing in the public polls.

So, how would this work? Let’s look at both sides in detail.

The Risk Is Not Just Unemployment

The mistake in most of this debate is that it keeps collapsing the problem into job counts. Will AI destroy jobs? Will it create jobs? Are postings up? Are layoffs up? Are people retraining? Are new categories appearing?

Those questions matter, but they are not the whole shape of the risk. The deeper problem is distribution during the transition window.

AI may deliver large productivity gains and still weaken the link between work and income. It may grow the economy while shifting the gains away from wages and towards compute infrastructure, capital owners, cloud platforms, chip suppliers, model providers, and the firms able to reorganise around these systems first.

If that is the case, then this will not be a normal recession. It would not be a cyclical downturn where demand falls, governments borrow, workers get support, and the economy returns to trend. It would be closer to what Anthropic’s June 2026 Economic Policy Framework calls:

That is a very significant sentence.

It means the economy could keep producing more while the old income distribution mechanism stops working properly. Output rises. Market caps rise. Capex rises. Data centres fill. Agents transact. Firms become more productive. But the human wage channel no longer receives enough of the flow. That would be a modern rentier economy, the AI equivalent of a petrostate at global scale.

And once that happened, the second-order effects would really start to matter.

A fired or not-hired symbolic-worker does not just cancel a SaaS subscription. They cut back on coffee, childcare, renovations, tutoring, dental work, restaurants, rent, mortgages, household services, travel, and local spending. They also pay less personal tax. That weakens the public sector just as the need for transition support rises. The barista, childcare worker, plumber, landlord, builder, teacher, council, and state budget all sit downstream of the symbolic-worker wage base.

This is why the standard answer of “people can move into services” is so incomplete. Many of those services are funded by the disposable income of the very workers being displaced, wage-compressed, or never hired into the ladder in the first place.

This economic cascade would also be geographically uneven. The places that benefited most from the symbolic-work boom are likely to be hit hardest if that wage base weakens. San Francisco, Seattle, New York, London, Sydney, Melbourne, Dublin, Toronto, Bangalore. The centres that built the modern knowledge economy may also become the first places where the second-order impacts show up.

But that is only one geography of the AI transition.

There is another.

Two Geographies, Two Claims

The displacement geography and the resource-host geography are not the same. It is where the symbolic-worker wage base is concentrated. These are the cities and regions where software workers, analysts, designers, consultants, marketers, lawyers, finance workers, product managers, researchers, and junior knowledge workers have carried a large part of local demand.

But the resource-host geography is different. These are the places carrying the physical AI buildout: the data centres, grid upgrades, transmission lines, water demand, land use, backup generation, cooling infrastructure, and local political bargains. Northern Virginia is not San Francisco. Memphis is not New York. Rural Sweden is not London. Parts of Ireland, Texas, North Dakota, Arizona, Nevada, the Gulf, and other data-centre regions may carry the physical load even when the highest-value firms, shareholders, and users are somewhere else.

Both geographies have claims, but they are different claims.

If the negative case plays out then the displacement geography needs income support, fiscal stabilisation, career-ladder repair, and protection for the local service layer that sits downstream of symbolic-worker wages.

While no matter the outcome, the resource-host geography needs grid compensation, water protections, land-use protections, infrastructure payments, and some share of the value produced on its doorstep.

Any serious AI transition policy has to see both. If it only compensates symbolic workers in the case of a revolution, it looks like knowledge-class self-protection. If it only compensates data-centre towns, it misses the economic cascade. Any political coalition has to include both the people whose wage base may be hollowed out and the communities whose physical resources are being used to build the new machine.

The Wrong Way To Solve This

There is a growing policy conversation around AI taxation. Some of it focuses on token taxes. Some of it focuses on public ownership or sovereign wealth funds. Some of it looks like a digital-services tax. Some of it looks like a Pigouvian automation tax.

A sovereign wealth fund that only takes stakes in the labs does not fully solve the problem. A public stake in AI labs may sound clean, because it goes straight at the capital ownership problem. But the frontier labs may not be profitable during the transition window. They may be economically central while still burning cash on training runs, inference subsidies, talent, chips, safety, enterprise sales, and infrastructure. In a hypercompetitive investment phase, the labs may be the customers of the rentiers rather than the rentiers themselves. That is a bad match for near-term transition insurance. And the displacement risk may be near-term, while the equity upside may be long-dated, uncertain, geographically concentrated, and captured elsewhere in the stack.

The tax instinct may seem right. If AI shifts income from labour to capital, then the tax base has to move too. But most of these proposals also have a geography problem.

If a tax is national, capital can move. If it is tied to corporate domicile, the company can restructure. If it is tied to profits, transfer pricing and IP games return. If it is tied to tokens, the accounting can be gamed or bundled away. If it is tied to the point of consumption, it assumes the human consumer remains the centre of value.

And the critical point of the economic revolution view is that the last assumption is already weakening.

The emerging agent economy changes the picture. Google has already announced the Agent Payments Protocol, built to let agents initiate and transact payments across platforms. Coinbase and AWS are now working to let publishers and API providers accept agents as customers through x402. Visa is building agentic commerce infrastructure too, with Intelligent Commerce Connect aimed at helping agents pay and merchants accept agent-led transactions.

This is critical because it shows how the consumption endpoint is already starting to leak.

If AI infrastructure migrates from serving billions of humans with declining spending power to serving trillions of agents transacting at machine speed, then human consumption is no longer the obvious place to capture value. It may still matter, but it is not the root.

The real root is resource command.

AI systems need chips, data centres, land, cooling, grid capacity, capital, and above all energy. They can route around tax jurisdictions, billing categories, accounting definitions, corporate structures, and human-facing transactions. They cannot route around physics.

Measure The Unavoidable Input

The cleanest input-side anchor is energy.

Not because energy is the whole story. It isn’t. Training, inference, chips, data, models, water, land, and grid capacity all matter. But energy is the simplest unavoidable physical substrate. There is no AI without power.

The point is not that energy is bad. The point is that AI-dedicated energy consumption is a measurable proxy for large-scale AI resource command.

This is also not some impossible measurement problem. Large AI infrastructure firms already know their power commitments, data-centre loads, compute purchasing, cloud expenditure, accelerator utilisation, and energy contracts. OpenAI’s compute spending is already a board-level operating category, with Reuters reporting that it expected to spend around $50 billion on computing power in 2026 and roughly $600 billion through 2030.

The broader system-level pressure is also visible. The IEA projects global data-centre electricity consumption roughly doubling to around 945 TWh by 2030, growing at around 15 percent per year from 2024 to 2030. That is more than four times faster than electricity consumption growth from all other sectors combined.

So the input-side proposal is simple.

Large-scale AI infrastructure operators should report audited AI-dedicated energy consumption. Above a meaningful threshold, that energy use becomes the base for the SAFER Dividend.

That should include grid-supplied power and self-generated power. If a lab builds an off-grid data centre, colocates with renewables, buys nuclear output, runs gas turbines, or eventually does something more exotic, that should not make the activity disappear from the base. Off-grid compute may reduce grid burden and should receive credit where it adds genuinely new clean capacity, but it is still AI resource command.

The same logic applies to unregulated jurisdictions. The location of the compute should not determine liability. Market access should. If a model, agent network, cloud service, or AI-generated product sells into participating jurisdictions, it should comply with the reporting and contribution rules of those jurisdictions. That is not radical. It is the same basic logic behind border adjustments and destination-based tax reform. You do not get to serve the market while pretending the production chain is nowhere.

Over time, the formula should also account for efficiency gains. If models become dramatically more efficient, a pure dollars-per-MWh levy could shrink even while AI’s economic impact grows. That would miss the point. The base should therefore use audited energy consumption as the physical anchor, but also consider a provider’s share of overall large-scale AI energy use or frontier AI resource command. The system should reward efficiency without allowing dominant systems to disappear from the contribution base simply because the hardware improved.

But importantly, the rate should not activate simply because energy is used. That would be a blunt tax on AI. Instead, the rate should be tied to whether the economic risk actually appears.

The more AI preserves broad labour income, the lower the rate stays. The more AI decouples productive capacity from human income, the more the rate rises.

That creates the right incentive. It does not punish AI for being productive. It taxes AI resource command only when productivity stops flowing through humans.

It also encourages efficiency. A token tax encourages token-accounting games. An energy-linked levy encourages better chips, better cooling, smaller models, smarter routing, off-peak compute, additional clean generation, and grid-aware scheduling. If the industry can do the same work with less energy, it pays less. That is good.

Humans Are The Key

The output side should not be consumption. It should be population.

That is the most important shift.

If the risk is structural decoupling, then the redistribution claim should not depend on where the remaining human consumption happens to occur. It should not depend on whether a human bought an AI service, whether an enterprise used an agent, or whether a transaction moved through a human-facing market at all.

If AI weakens the old link between human work and income (the real claim behind the “economic revolution” case), then humans become the relevant distribution unit.

But this does not mean every dollar should instantly be distributed equally across the world. The political economy would not support that, and the transition starts from an existing structure of GDP, markets, tax bases, and infrastructure burdens. A workable version would start with some blend of GDP and population, then it could perhaps shift over time towards population.

Early on, richer market jurisdictions would receive more because they are where the existing market and tax base sit. Over time, as AI value moves further away from human consumption and towards machine production, the population weight could rise. Eventually the system could look less like a consumer rebate and more like a human dividend.

But even “dividend” may be too narrow if we only mean money.

In the transition window, cash matters because people still need to pay rent, buy food, service debt, keep households intact, and support the local businesses around them. But if AI moves us towards a modern rentier economy, where productive capacity is increasingly held in compute, energy, models, agents, and capital, then redistribution will need to include more than fiat transfers.

It will need to include public services. It will need to include local fiscal support. It may need to include public compute, sovereign model access, energy access, housing support, education, healthcare, childcare, and ownership claims in the productive stack itself. All distributed via the existing local governments of course.

If wages no longer provide humans with a sufficient claim on production, then policy has to create that claim somewhere else.

That also means national sovereign wealth funds are not enough. They are useful, but they are geographically limited. A US public stake in US AI firms may help Americans if those firms eventually pay off. It does not solve displacement in Australia, India, the Philippines, Europe, Africa, Latin America, or anywhere else exposed to the same technology. It may even deepen the rentier divide between countries that own AI assets and countries that merely import their effects.

A serious version needs market-access rules, international pooling, or at least a path from GDP-weighted allocation towards population-weighted human claims.

And as we’ve covered, host communities also need their own share. The towns and regions that carry the data centres carry real costs. Grid congestion. Water use. Land use. Infrastructure strain. Political bargaining around local tax concessions. Often they receive a handful of permanent (or even temporary) jobs while the real surplus flows elsewhere. So part of the revenue should be routed by data-centre load, peak demand, grid stress, water stress, and local resource burden.

This gives the policy three output channels.

• One channel compensates humans because the labour-income link may be weakening.

• One channel stabilises displacement regions because symbolic wage bases may be hollowed out.

• One channel compensates host communities because the physical infrastructure lands somewhere.

The Trigger: Distribution Risk Index

The missing piece is the trigger.

We need a way to avoid prediction warfare. The industry says AI will create abundance. Critics say it may break the labour-income channel. Rather than asking politicians to pick a side in advance, build the trigger and pre-commit to the claims process before the claim arrives.

Call it the SAFER DRI (Distribution Risk Index).

The SAFER DRI should measure whether AI-linked compute and capital are scaling into broad labour-income growth or away from it. It should not be based only on unemployment, because unemployment is a lagging and incomplete signal. Jobs can remain while wages compress. Headcount can look stable while entry-level hiring collapses. Average wages can rise because junior workers vanish and senior workers remain. GDP can grow while the labour share falls.

It also has to catch displacement that does not look like unemployment at first.

A white-collar worker can be displaced into contracting. A graduate can fail to enter the ladder at all. A software worker can become underemployed in gig work. A laid-off analyst can become technically self-employed while earning less, working fewer hours, and carrying more volatility. Someone can leave the labour force because the entry path has disappeared. None of that necessarily shows up immediately in headline wages or labour share.

The trigger must detect downgrading, not just unemployment.

So the dashboard should track entry-level hiring, cohort-specific employment, hours worked, income volatility, underemployment, labour-force exits, payroll-to-contractor shifts, occupational downgrading, and graduate-to-career mismatch. Some lag is unavoidable, but that is exactly why the reporting regime has to exist before the levy fully activates, and why the trigger should include leading indicators rather than waiting for official unemployment to move.

But the SAFER DRI should not be a magic equation.

A simple version of this idea could be written as if the SAFER DRI were a product of four stress indices: compute-labour divergence times labour-share stress times wage and career-ladder stress times tax-capture stress. That is useful as a mental model, but it’s too fragile as policy design. If one term falls near zero, the whole index disappears. If one noisy term spikes, the index can overreact. And the multiplication implies a precision that does not yet exist.

A better design is a two-gate dashboard.

• Gate 1: The Resource Command Gate (Is AI Scaling?)
  The first gate asks whether large-scale AI resource command is rising fast enough to matter. This would track AI-dedicated energy use, frontier compute, data-centre load, accelerator deployment, AI capex, and concentration of AI infrastructure.

• Gate 2: The Income Absorption Gate (Are Humans Decoupling?)

  The second gate asks whether broad labour-income absorption is weakening at the same time. This would track labour share, exposed-sector wages, entry-level hiring, underemployment, labour-force participation, AI-attributed layoffs, payroll employment, tax receipts, downgrading, and geographic concentration of labour-market stress.

That avoids taxing AI merely because compute is growing. It also avoids pretending that a normal labour-market downturn is automatically an AI event. The question is systemic, not forensic. We do not need to prove that a specific model caused a specific firing. We need to know whether large-scale AI resource command is rising while the human income channel is weakening. Is the economic revolution showing?

The exact construction is obviously a serious statistical design problem by real experts in this field. It could be a weighted index, a traffic-light dashboard, a threshold-count rule, or even some combination of all three. But what really matters here is that activation should require multiple independent signals, sustained over time, rather than one noisy number.

Who Watches The Trigger?

This is the most captureable and political part of the whole proposal.

If everything routes through the SAFER DRI, then whoever defines this becomes the claims adjustor for this global insurance contract. That cannot be left to the industry being assessed. It also cannot be left to a minister looking for a convenient number.

Every measurement choice will be contested. What counts as frontier AI compute? What counts as an exposed sector? How do we treat contractors? How do we separate AI-driven stress from interest rates, overhiring, offshoring, remote work, or normal restructuring? Which labour-share measure matters? Which geography matters? How long must the trigger stay elevated? How are revisions handled?

The party with the most resources to litigate those choices will usually be the party being taxed.

So DRI governance is not a detail. It is the core of this insurance contract.

The index would need to be governed like critical economic infrastructure. It should be run by an independent statutory body, with central-bank or statistical-agency style independence. Its components, weights, thresholds, data sources, and revision rules should be pre-registered before activation. Changes should require a high bar and multi-jurisdictional agreement. Industry must be able to challenge data errors, but not be able to relitigate the social meaning of every threshold for years.

The trigger must also be redundant by design. No single indicator should control the outcome. The system must look for a pattern across compute scale-up, labour share, wages, entry-level pathways, employment quality, downgrading, tax capture, and geography.

This is not “let the data decide” in some naive sense.

Data do not decide anything by themselves. Institutions decide which data matter.

The real claim is narrower and stronger than that:

This Does Not Capture Every AI Rent

The SAFER Dividend would not capture every part of the AI surplus.

In particular, it would not directly capture NVIDIA or similar upstream chokepoints. It would capture the large-scale use of AI infrastructure by labs, clouds, data centres, inference providers, and major deployers. It would not automatically capture the extraordinary rents earned by chip suppliers, semiconductor foundries, lithography firms, high-bandwidth memory suppliers, networking vendors, or other firms that sit upstream of energy consumption.

That is not a gap, this a real design feature.

It matters because some of the most concentrated gains in the AI transition may appear before any model is trained or deployed. The chip supplier, the foundry, the cloud vendor, the power developer, and the landowner may all capture value from the buildout itself. But these are issues for corporate anti-monopoly policies that deal with rent capture.

SAFER should be cleanly isolated from this. The answer is not to make the energy mechanism do everything. That would make it too complex and easier to attack. The answer is to admit that the SAFER Dividend is the automatic stabiliser, not the whole redistribution system.

This distinction is important.

Better SAFER Than Sorry

This is better than an automation tax because it does not require proving that a specific model caused a specific firing. Better than a token tax because it does not depend on visible human-facing usage. Better than a consumption tax because it does not assume human consumption remains the centre of value. Better than a simple energy tax because it only materially activates when the labour-income risk appears. And better than waiting because the rules are negotiated before the fiscal and labour-market damage arrives.

The AI industry should be able to accept this.

If AI generates broad prosperity, the trigger stays low. If workers capture the gains, the trigger stays low. If tax receipts rise with productivity, the trigger stays low. If career ladders remain open, the trigger stays low. If the abundance story is true, the policy is mostly an accounting exercise that is already under way.

But if AI produces the scenario even Anthropic and OpenAI now say is possible, if there is a structural decoupling of productive contribution from income, then society should not be negotiating from a weakened position after the fact.

The input is energy, because AI cannot avoid it.

The output is human claim, because humans are the constituency that matters if work and income are torn apart.

The trigger is the SAFER DRI, because the policy should activate on evidence rather than fear.

The governance matters, because the trigger is only as neutral as the institution that controls it.

And the upstream rent layer matters, because the energy dividend is not designed to catch every NVIDIA-shaped chokepoint in the stack.

That is the deal.

If the AI abundance side is right, it costs the industry almost nothing.

If they are wrong, we will all be very glad the insurance was already in place.

Do Something!

If you have found yourself on either side of the AI abundance vs jobpocalypse argument recently then share this proposal with the people you argued with. Let’s see if this can provide some productive common ground.

If your local representative is proposing an AI tax, token tax or sovereign wealth fund, etc. then share this proposal with them and highlight why those proposals may be limited.

If you know a policy maker, an economist or someone inside a frontier AI lab then share this proposal with them and start a real discussion.

Personally, I really hope the AI abundance side is right and this is just another economic evolution. But I don’t want to just sit here hoping. I know that insurance policies are only useful if you negotiate them before the event, and that SAFER could help people sleep better at night. SAFER could free us all to focus on realising an abundant future instead of wasting energy arguing about what we “believe” the future will hold.

So don’t just sit there. Like it, share it or argue with it. But do something!

To make this important topic more accessible I’ve created a skill that you can use with your favourite AI in 3 easy steps. Here’s how:

1. Download the SKILL.md file of your choice

2. upload it to your AI

3. then just say “Run this skill”

Note: It’s important to download this SKILL.md file and upload it directly to your AI chatbot rather than linking to it . We also recommend that you use one of the leading models from the 3 main frontier labs (e.g. Claude, Gemini or ChatGPT). If you use a lower level model then your results may vary. And of course if you find any issues, or feel like you’ve found a real flaw in my arguments please let me know — I value constructive feedback.

On Tuesday the 10th of June, the company widely seen as the most safety-conscious of the leading AI labs published a detailed policy framework proposing, among other things, universal basic income, AI sovereign wealth funds, and worker equity-sharing in AI companies. On Friday the 13th of June, the US government issued an export-control directive suspending public access to that same lab’s two most powerful models. The same week that one lab proposed redistributing the wealth produced by AI, the government acted unilaterally to control who can use it.

These two events, three days apart, are the clearest single image of where the AI conversation now sits.

What the lab put on the table

The Anthropic Economic Policy Framework is fourteen pages. It is calibrated to three possible scenarios distinguished by the US unemployment rate.

At about 5 percent unemployment, the proposed responses are conventional. The framework calls for universal capital accounts (financial accounts seeded for every American at birth), funded in part by equity in AI companies themselves. It proposes wage insurance for workers who take pay cuts after AI-driven displacement, expanded workforce training, easier movement between licensed occupations, and tax incentives for firms that retrain rather than dismiss their workers.

At about 10 percent unemployment, the proposed responses expand. The framework calls for stronger unemployment insurance, targeted sector-specific transition support, basic-needs relief for those who exhaust other benefits, and crucially, “government regulations and incentives at the firm level that manage the pace of displacement”.

At unemployment beyond historical peaks, the framework proposes things no major US AI lab has publicly proposed before. New tax bases including levies on AI use measured by compute, tokens, or revenue. “Digital dividends” funded by taxes on the digital sector. Universal basic income. AI sovereign wealth funds funded by public investment stakes in AI-driven productivity. Worker equity-sharing in AI enterprises. Its own summary line: “Regardless of the tax base or distribution mechanism, we are ready and willing to pay our fair share”.

Alongside the framework, Anthropic restated their 200 million dollar Economic Futures Research Fund and a 150 million dollar national fellowship programme for early-career professionals working on AI’s economic effects. Dario Amodei published a personal essay the same day.

This is genuinely unusual. A frontier AI lab is publicly endorsing, as policy candidates worth taking seriously, mechanisms that until recently were associated with the political left’s most ambitious AI proposals.

But the timing also raises serious questions. Anthropic confidentially filed paperwork to go public on the 1st of June. The policy framework was published nine days into its IPO roadshow. Several writers immediately viewed it less charitably. David Sacks, one of the tech investors who lobbied the White House to cancel an earlier AI safety order, called it “a regulatory capture strategy based on fear-mongering”. Kate Aronoff in The New Republic drew a parallel to the fossil-fuel industry’s net-zero pledges. Their argument: capital accounts seeded with AI-company equity create new institutional demand for AI shares - “willing to pay our fair share” was published without a number. The Facebook precedent of “regulate us, but on our terms” is the cautionary parallel.

Both perspectives carry weight. Anthropic has publicly accepted the vocabulary of structural redistribution at IPO timing, and the question of whether that signals genuine policy commitment or marketing positioning will only resolve over the coming months - based on whether the lab continues to advocate Tier 3 mechanisms after listing, and whether it eventually endorses any specific binding legislation.

What the government did

Three days later, on Friday the 13th of June at 5:21 in the afternoon Eastern time, Anthropic received a letter from the US government. The directive was straightforward in effect: suspend all access to Fable 5 and Mythos 5, the company’s two most powerful models, for every foreign national worldwide. The restriction includes Anthropic’s own foreign-national employees, who can no longer use the models their own company produces.

The directive cites no specific agency. It names no statutory authority. It provides no appeal mechanism. Its stated trigger is a demonstration that Fable 5 can be asked to “read a specific codebase and fix any software flaws” - that is, do the work of a software engineer, which the model is designed to do.

Anthropic publicly disagreed. The company’s statement said that “the finding of a narrow potential jailbreak” should not “be cause for recalling a commercial model deployed to hundreds of millions of people”, and explicitly pointed at competitor OpenAI’s GPT-5.5 as having “comparable capabilities”. Reporting from the Wall Street Journal traced the directive in part to Amazon’s CEO Andy Jassy briefing the Treasury Secretary that Amazon researchers had been able to extract security-relevant information from the model.

The structural point has much more impact than the technical details. The Trump administration cancelled a federal AI safety executive order on the 21st of May after pressure from tech industry figures. It signed a replacement order on the 2nd of June that explicitly avoided creating safety-vetting requirements. Ten days later, the same administration issued an unappealable, late-Friday directive shutting down public access to a leading lab’s commercial product on national-security grounds, without naming the authority it was using. The earlier framing of US federal AI policy as “hands off” no longer fits. What had looked like absence is now visibly capability-control, conducted by the executive branch without legislation.

The international response

The directive’s effects landed almost immediately outside the US. On the 14th of June the European Commission’s tech-sovereignty spokesperson, Thomas Regnier, issued a statement saying that contingency measures in capability control “should not be discriminatory against partners” and that Europe was “assessing implications, including for users in the European Union”. The UK’s Minister for AI and Online Safety, Kanishka Narayan, told the press that “access to AI capabilities is crucial” and that “I care about sovereign AI because it now decides our security”. The Register described the European response as an “AI sovereignty surge”.

European capital is following. Mistral, the French AI lab, is reportedly in talks to raise about 3 billion euros at a 20 billion euro valuation, roughly double its valuation from nine months ago. The European Commission’s broader Technological Sovereignty Package had been launched on the 3rd of June, ten days before the Fable directive; the timing now looks down right prescient.

Meanwhile, here in Australia there is no real response at all.

There is an analytical framing that has been circulating in technology-watching circles, and it is worth raising again because it explains why the European response is sovereignty-coded rather than just regulatory.

In April 2023, Anthropic’s confidential pitch deck to Series C investors was leaked. One line, widely reported at the time, described the company’s view of the next three years: “We believe that companies that train the best 2025/26 models will be too far ahead for anyone to catch up in subsequent cycles”. The technology analyst Andrew Curran recently argued, looking at the current generation of frontier models from Anthropic, OpenAI, and Google, that the prediction has held: any nation that wanted its own competitive frontier-AI capability had a three-year window from early 2023 to early 2026 to build it. That window, in his opinion, is now closed.

The strong version of that claim (that no one can ever catch up) may be overstated. Chinese labs continue releasing models close to frontier capability. Open-source models stay within months of closed-source ones. New entrants like the lab founded by former OpenAI executive Mira Murati are still raising capital. But the weak version of the claim (that the frontier is now concentrated in a small number of labs in two countries, and that nations not hosting one of those labs have to make sovereignty arrangements rather than try to compete from scratch) fits what we are watching happen.

That is the lens through which the EU and UK responses make sense. If you accept that the window is closed, sovereign AI is the only response available to anyone who isn’t already inside it.

The Sanders conversation

While the EU was responding to the directive, the US Senate was also moving. On the 1st of June, Senator Bernie Sanders formally introduced the AI Sovereign Wealth Fund Act. The bill’s mechanism is severe: a one-time 50 percent tax on the equity of the largest AI companies, payable in shares. Federal government voting seats on those companies’ boards with veto power. The proceeds distributed to the public as direct payments and as guaranteed access to healthcare, education, and housing.

The reception was unexpected. According to reporting in The Hill, Sam Altman of OpenAI met Sanders privately in early June, at Altman’s own request, and “expressed general support for the concept of public equity in AI”, disagreeing only with the 50 percent level as “too ambitious”. The Trump White House was reportedly receptive to the principle of public ownership in AI companies, consistent with the administration’s existing equity positions in roughly twenty semiconductor and critical-mineral firms. Senators Warren and others have published proposals in adjacent territory.

What this means in practical terms is that the federal political space, which had been essentially empty on AI for a year, now has at least three areas of building momentum operating concurrently and overlapping rather than opposing on some structural questions. A left proposal for redistributive ownership (Sanders). A centre-bipartisan proposal trading state-law preemption for federal audits (Obernolte-Trahan, released on the 4th of June). A pro-industry posture of voluntary cooperation (Trump’s June 2 executive order). Plus, as of the 13th of June, the demonstrated willingness of the executive to act unilaterally on capability control without going through any of those legislative channels.

That four-way reality is qualitatively new. The conversation about AI redistribution is no longer happening only on the political left.

The capability frontier continues moving

On the 9th of June, Anthropic publicly released Claude Fable 5, the publicly available version of the model previously gated as Mythos. Its score on the most-cited contamination-resistant programming benchmark reached 80.3 percent, eleven points higher than the previous public-tier model from the same lab. It is the first publicly available model at the previously-private capability tier that Anthropic had been holding back on the grounds that it was too capable for general release. Four days later, the government restricted access to it.

On the 10th of June, a collaboration between Renmin University of China and Microsoft Research released an autonomous AI research framework called Arbor. The system uses a long-lived coordinator alongside short-lived workers to test ideas, accumulating lessons as it goes. On a standard benchmark for autonomous machine-learning research, it outperformed both OpenAI’s and Anthropic’s leading agent products by 2.5 times. The team released the code openly on GitHub for anyone to run.

The Arbor result is important because it cuts against the strong version of the closed-window thesis. Chinese researchers working in academic collaboration with Microsoft are publishing open-source code that beats the products of the two largest American AI labs on a leading autonomous-research benchmark. The frontier may be concentrated, but it is not yet closed to academic research, and it is not closed to China.

What the two releases together demonstrate is that capability is continuing to move at the same time as the political response forms. Fable 5 shows that what was internal-only six weeks ago is now publicly released. Arbor shows that some forms of autonomous research are now reproducible in academic code rather than gated behind commercial APIs.

The pattern reaches major-bank chief executives

The mechanism by which AI is connected to job cuts has, until recently, been used most openly by technology companies. In the past month, the same vocabulary has reached the chief executives of major US banks.

Jamie Dimon of JPMorgan Chase, in early June commentary, described AI as something that “will eliminate jobs” and named the mechanism plainly: “attrition, redeployment, retraining, and early retirement”. Jane Fraser of Citi said some roles “will no longer be required”. John Waldron, President of Goldman Sachs, described the bank as “a human assembly line” that AI will automate. JPMorgan separately disclosed that operations and support staff will fall by at least 10 percent over five years. Junior analyst recruitment at major banks has reportedly been cut by up to two-thirds. The phrase “attrition” matters because it describes a pattern in which firms don’t announce mass layoffs - they simply don’t replace people who leave naturally, allowing AI productivity to hold headcount flat while functions just disappear.

The Challenger Gray monthly layoff report for May 2026, released on the 5th of June, found that AI was named as the leading single reason for cuts for the third consecutive month. Year-to-date AI-attributed cuts through May exceeded the full-year 2025 total.

How young people are responding

One clear measurement of how the conversation lands at the demand-side voter level came from polling published in April. A survey of about 1,500 14-to-29 year olds conducted in late February and early March 2026, released by Gallup with the Walton Family Foundation and GSV, found that in twelve months the share of young people reporting that AI made them feel excited dropped from 36 to 22 percent. Hopeful dropped from 27 to 18 percent. The share reporting that AI made them feel angry rose from 22 to 31 percent. Eighty percent of young people surveyed said they believed AI would make their future learning more difficult. The share who think AI tools help them learn faster dropped from 53 to 46 percent. The Gallup analyst leading the work called it “reassessment, not rejection”.

The level of measured anger is notable for two reasons. The 22-to-31 percent shift in twelve months is faster than comparable consumer-technology sentiment changes have moved historically. And anger, unlike anxiety, tends to translate into political mobilisation rather than withdrawal.

There are two years until the next US presidential election. The cohort being polled here overlaps substantially with the cohort whose employment outcomes the labour-market studies have been tracking. Both are pointing at the same underlying picture.

Where we stand right now

The previous summary closed by saying the picture was starting to sharpen. The picture has now sharpened even further, and in a specific direction.

The conversation about whether AI is structurally important enough to require redistribution policy at federal level is no longer happening in one place. It is happening at the boardroom of a leading AI lab (Anthropic’s policy framework). At the executive enforcement level of the US government (the Fable directive). At the federal legislative level (Sanders, Obernolte-Trahan, the bipartisan reception). At the state level (Illinois’s safety law still awaiting signature, Connecticut signed, California signed). At the international level (the EU and UK sovereign-AI responses, Mistral’s capital raise). At the chief-executive level of major US banks (Dimon and Fraser’s vocabulary). And at the cohort-sentiment level of voters who will be of age in 2028.

These conversations were each separately recognisable a year ago. What is new is that they are now happening concurrently and visibly responding to one another in real time. Anthropic published an economic-policy framework specifically because it expected the federal political space to fill. The US government’s enforcement action followed within days. The EU’s sovereignty response landed within days of that. The Sanders bill drew Altman to a private meeting. The pattern is no longer linear - it is mutually responsive.

There are also real counter-currents. The Anthropic framework may turn out to be IPO marketing rather than policy commitment, and only months of post-listing behaviour will resolve which. The Fable directive may be reversed under legal pressure, particularly the foreign-national-employee restriction. The single open-source release of Arbor suggests the strong form of the closed-frontier-window thesis is overstated.

The next decisive checkpoints in the conventional data are visible. Anthropic’s IPO prospectus, when it becomes publicly available later this year, will allow audit-grade verification of the company’s own productivity numbers. The next quarterly labour-market data from the US Bureau of Labor Statistics, expected in August, will be the first to cover a full post-capability-arrival quarter. The Sanders bill’s reception in committee will indicate whether the political space behind the proposal is real or rhetorical.

The picture that emerges from this update is not one of an industry in equilibrium with policy makers and the public. It is one of a fast-moving capability frontier, a lab arguing in writing that its own industry should consider a coordinated pause while filing for an IPO, a government issuing unappealable export controls without legislation, an opposition senator finding an unexpected audience at the White House, a European political class realising it is not in the room where the frontier is being built, the chief executives of America’s largest banks talking openly about AI attrition, and a generation of voters becoming visibly angry about it.

These are no longer separate stories. They are the same conversation, happening at speed, with the timeline visibly compressing.

Build one rich latent representation of the scene, then ask it questions. One encoded latent, one lightweight interface, tasks as query patterns.

Now let’s trace how it shows up across these different domains.

For most of AI-based computer vision history, a task was a model. Depth estimation, segmentation, tracking, pose recovery, optical flow - each had its own architecture, its own training pipeline, its own output head, and systems that needed several of these capabilities were built as mosaics: a depth model feeding a tracking model feeding a fusion step, with expensive optimisation glue holding the pieces in geometric agreement. The deep-learning era changed the components but kept that assumption. Even large shared backbones sprouted a separate decoder head per task, and the task list was fixed at training time.

Over the past two years, this assumption has collapsed - on both sides of the camera (perception and generation alike). The replacement now has a consistent shape. A heavy encoder runs once, encoding raw sensory data into a single rich latent representation. A deliberately lightweight, general interface then sits in front of that latent, and tasks stop being models. Instead they become patterns of queries against the representation. So the expensive thing is building the understanding. And the cheap thing is asking questions of it. This architecture separates those two costs.

On the perception side the trajectory is easy to trace. NeRF established the field-as-function idea - feed a coordinate in, get back a value. The Scene Representation Transformer and Perceiver lines generalised query-based decoding. SAM made segmentation ‘promptable’: one heavy image encode, then arbitrary lightweight prompts (a point, a box, and eventually an open-vocabulary concept) each returning a mask. Now the current culmination is D4RT, which encodes a whole video into one global scene representation and then answers point-level queries of the form “this pixel, from this frame, at that time, in that camera’s coordinates.” Point tracks, depth maps, point clouds, camera intrinsics and extrinsics are no longer outputs of separate heads. They can be recovered as different query patterns over the same representation - now recipes, not models.

The benchmarks suggest this is not even an elegance-for-accuracy trade: in D4RT’s reported results, the unified interface is more accurate than the specialist pipelines it replaces while running 18 to 300 times faster - partly because independent queries are trivially parallel, and partly because you no longer compute dense answers nobody asked for. There are two caveats to that claim, though. These are the authors’ own numbers, and the code isn’t public yet.

However, the longer history of unified models is partly a history of negative transfer - shared representations that initially lagged the specialists they were meant to replace. NLP followed exactly that curve: unified models lagged, then crossed over, then never looked back. D4RT may not show that the crossover in vision is complete, but it does show that the crossover is now visibly underway, and an award committee rewarded it.

The generation side arrived at the same shape too, but from the opposite direction. A single pretrained diffusion or flow backbone now serves what used to be a zoo of separate systems (inpainting, outpainting, editing, style transfer, super-resolution, structural control) all expressed as conditioning patterns on one prior: a mask here, a reference image there, a depth map, an identity lock, a camera trajectory, a robot action vector. Interactive world models are the clearest case: the same generative core, with the conditioning channel swapped, becomes a text-to-video system, a camera-controlled explorable scene, or an action-conditioned simulator. Native-3D generators extend the pattern to the output side, where even the format (Gaussians, mesh, radiance field) is a decoding choice from one structured latent rather than a different model.

And here SAM 3D is the most telling example of all, because it doesn’t sit cleanly on either side. It is a generative model (write-side machinery) openly released and used to answer a perception question: what is the 3D shape, texture and layout of this occluded object in this cluttered photo? The paper the award committee called out alongside D4RT is blurring the line we started with at the beginning of this post - keep that in mind.

There is a precedent, of course, and that is language. Large language models made this interface shift obvious to everyone first: one pretrained model with tasks as prompts. Within roughly two years of that becoming undeniable, NLP’s ecosystem of task-specific architectures simply dissolved into a single interface. What’s happening now is vision having its own prompting moment - but this time twice over, because the pattern is landing on both the read side and the write side. Perception is becoming a read interface over a latent world-state: query what is there, where it is, and how it moves. Generation is becoming a write interface over a latent prior: specify what should be there, and condition it into existence. Same shape, opposite direction of information flow.

Why this pattern wins is interesting, because it’s definitely not just an aesthetic preference. This is the Bitter Lesson operating at the interface level: general mechanisms that let ‘scale do the work’ beat hand-assembled mosaics of specialists, and a query interface is the most general output mechanism there is. It’s also economics. Encode-once-query-many means training can supervise on sparse random queries instead of dense decoding. So inference cost scales with the questions asked rather than with the resolution of the world, and sparse, on-demand probing becomes the natural consumption pattern. Here you only pay for understanding once and for each question only when you ask it. Plus it’s composability: when tasks share one representation, their outputs are mutually consistent by construction, rather than reconciled afterwards by optimisation glue.

But consistency by construction cuts both ways, and ironically this is the strongest benefit we lose from the old architecture. When every task reads from one latent, any encoder error corrupts depth, tracking and pose simultaneously and coherently. The old mosaic was ugly, but disagreement between its specialists provided a free error signal - the optimisation glue wasn’t just reconciliation, it was also a sanity check. This new monolith has no internal dissent. Consistent can also mean consistently wrong, with no residual left over to flag it. That doesn’t mean this new pattern is invalid. It just means that there is at least one thing that the mosaic got for free that we no longer get: calibrated uncertainty as a first-class query, not an afterthought.

Where does this leave us?

Right now we’re sitting at the frontier of this evolution and some constraints do still remain - at least for now. The encoding step is still mostly offline and bidirectional (whole clips encoded with global attention) which makes these systems powerful teachers but not yet streaming, real-time participants. But the generation side has already given us a playbook for fixing this (distil the bidirectional teacher into a causal student), and I think it is a safe prediction that streaming, causal versions of query-based perception will follow soon using this same strategy. Query vocabularies are still ad hoc and per-model too - there is no shared standard, so “API” is generally a metaphor at the ecosystem level. And the latents themselves are relative-scale and plausibility-grade: anchoring them to metric reality, and trusting them when being wrong is expensive, these both still require some machinery that’s outside this pattern. This is exactly the no-internal-dissent problem we discussed above, but showing up during deployment. There’s also one more constraint, and it’s economic rather than architectural: the write side trains on effectively all internet video, while explicit 3D structure and robot demonstrations remain scarce by comparison - this is a point Fei-Fei Li and the World Labs team made well only last week in their functional taxonomy of world models.

So the clear pattern that’s visible now is where these two sides meet. If perception is converging on a read interface over a latent world-state, and generation on a write interface over a latent world-prior, the natural endpoint is a single representation that supports both - which is what the phrase “world model” actually means operationally, and stripped of mystique: not a video generator, but a maintained latent state of a world that can be queried and steered through general interfaces. Today the read and write sides are mostly separate models that have converged in form, but not yet in substance.

Li’s taxonomy is worth reviewing again here, because it arrives at this same endpoint from the opposite direction. She carves up world models based on what they emit: renderers output observations (pixels for human eyes), simulators output state (structure that humans and programs can both compute on), planners output actions. Her predicted endpoint (one foundation model that renders, simulates and plans, switching its output modality to suit whatever consumes it) is the fusion described above, seen from the output side rather than the interface side. The same object, based on two different cuts. But notice that the agent-world loop her taxonomy is built on has a fourth arrow that her three categories leave implicit: observations flowing back into state. That is perception. The belief update. That arrow is exactly the read interface, exactly where D4RT lives, and exactly what “a maintained latent state” means in practice. The two framings complete each other - and a taxonomy like hers and an award podium like this landing in the same week is itself a signal of where the field is turning.

I predict the substance will catch up with the form, because there are two real drivers. The first is economics: maintaining two giant latents of the same world is paying the expensive cost twice, and this whole pattern exists because the field stopped tolerating exactly that kind of duplication. The second is embodiment: anything that acts (a robot, or an agent) has to perceive, predict and plan against one state of the world. So embodied AI doesn’t just prefer the fusion, it requires it. And SAM 3D suggests the fusion process has already begun: a generative prior answering perception queries is both read and write using shared machinery, today, on the CVPR award podium. Just don’t expect the fusion to be symmetric: given the data asymmetry discussed above, it will almost certainly be pixel-led, because the write side simply has more to learn from.

What persists through all of this?

It isn’t any particular interface - there is no shared query standard yet, and every model speaks its own dialect. What persists is this new pattern. And don’t expect the dialects to fully disappear, because language never finished that job either: LLMs converged totally at the conceptual layer (tasks as prompts), only de facto at the wire layer (everyone clones the market leader’s endpoint), and not at all at the frontier, where vendors strive to differentiate.

SQL is a long-run precedent for this - a shared core plus vendor dialects that never die. But vision will be even harder, because language got its query vocabulary for free while vision’s has to be designed: coordinate frames, time indexing, output encodings. Which is why one camp is already walking a different path entirely: don’t standardise the questions at all - decode the answers into containers everyone already reads. That is the explicit-state strategy, and World Labs’ Marble is one of its clearest expressions: export Gaussian splats and collision meshes, and let the legacy 3D stack of formats, physics engines and toolchains do the interoperability work. The cost, in this post’s terms, is dense decoding - paying for answers nobody has asked for yet. (It’s also telling that where each lab stands here maps to what it ships: implicit latents suit a company selling encoders, exportable state suits a company selling worlds.)

So this future really has three branches: a new query standard emerges, the legacy formats absorb the state, or language absorbs the queries. My bet is the third. Li opens her taxonomy by declaring that the world is not made of words - and she’s right about the substrate, but the interface is a separate question. The world may not be made of words, yet the questions we ask of it may well be, with structured geometric queries surviving underneath, the way SQL sits under other representational systems. Either way, the thing to watch isn’t who wins a standardisation fight - it’s whether the query layer converges at all, because that decides if world-model capability becomes a commodity behind a clonable interface, or stays locked inside proprietary dialects. That, plus what becomes askable as query vocabularies grow (from geometry toward semantics, affordances, physics and uncertainty), is the practical way to track the actual capability frontier. The models underneath will keep churning, while it’s the pattern that will persist.

The main insight is simple: the field is reorganising around the idea that understanding is expensive and questions are cheap. Build the understanding once. Make everything else a question.

Recently I’ve written about the flawed methodology in much of the AI-job-boom narrative, why the new “calm down, AI won’t take your job” messaging gets it wrong, and why retraining isn’t the answer. I also asked if you’re the horse in this story.

All of these have been dancing around one specific point.

Now it’s time to look at that underlying idea, and how the real discussion here is not about the technology. This is where two conversations constantly get blurred. So let me be clear up front: I am an AI optimist. I use these tools daily and I spend a lot of time exploring their internals. But none of that is what this post is about. Instead, it’s about why this new capability is structurally different, and how this interacts with the way our economy has developed.

But the capability side and the economic side don’t always play nicely, so it’s important to recognise:

The background story: For over 250 years, automation has had the same shape. Each wave cleared work from below, and the displaced people climbed up to a new layer above it. Farm labour to factory work. Factory work to clerical work. Clerical work to knowledge work. The shape stayed the same because there was always a layer above to climb up to, and the people that were displaced generally had the operational capacity to do that climbing.

This time the shape has changed. I know that as soon as people say “but this time is different” they’re labelled as “doomers” and everyone points to all the previous examples of “this time” failures. But hear me out and focus on the “structural” point.

The shape has not changed because AI is more capable in the abstract - that has been claimed in previous revolutions, but the wedge just kept on working. This time is different because the thing being targeted is the structure of the climbing mechanism itself.

That is the idea that you need to seriously consider. And the reason it’s worth taking seriously is because trillions of dollars have been allocated to one goal that makes this real (regardless of your view on where we’re up to on this path) - and that is AGI (Artificial General Intelligence).

But the AI labs and most public commentary have grown to use a flattened definition of AGI. A definition that makes this “change in shape” harder to track and see. In fact, the AGI term has been used to describe two quite different things, and they both impact the labour-market story in different ways. So the right place to start is with the two AGIs.

Which AGI?

The G in AGI stands for “General”. Simple enough. Except here, as I mentioned, “general” can mean two different things. And the AGI literature meant “both at once” for forty years, before the labs appeared to cut one out.

The first sense is breadth. Performance across many tangible tasks in many domains. The “AI as ‘good enough’ at pre-defined or economically valuable tasks” framing. This is what multi-task benchmarks like MMLU or HELM measure. It’s what OpenAI’s 2018 Charter operationalises as “highly autonomous systems that outperform humans at most economically valuable work”. Call this one task-AGI.

The second sense is generalisation itself - the cognitive process. The capacity to abstract, transfer, compose, and adapt to novel circumstances. The same operation working across previously unknown problems. This is what prior-controlled benchmarks like ARC-AGI measure. It’s how much new performance the system extracts per unit of new data. Call this one process-AGI.

The labs focused on task-AGI for good reasons. It’s measurable. It’s tractable. It’s economically useful early on. Process-AGI is the more complex one. It’s what the founding generation of AGI researchers pointed at - and what the rigorous formal critics never abandoned.

These are not opposites. The seminal AGI literature held both as a conjunction. When Ben Goertzel and Cassio Pennachin coined “AGI” in their 2007 volume, their working definition coupled both senses: “general scope AND is good at generalization across various goals and contexts”. Allen Newell and Herbert Simon’s 1976 Physical Symbol System Hypothesis defined “general intelligent action” as both “the same scope of intelligence as we see in human action” AND “adaptive to the demands of the environment”. McCarthy, Russell and Norvig, Pei Wang’s NARS work - for four decades the seminal research work held both definitions.

Then something happened. Or perhaps more accurately, something didn’t happen.

Flattening AGI

Between roughly 2014 and 2018, the industrial usage of AGI slowly stopped meaning both and started meaning task-AGI alone. Not because someone argued for it. But because nobody really argued for the other side when it mattered.

Nick Bostrom’s Superintelligence (September 2014) stipulated task-AGI as a definitional shortcut for a popular book about safety. The Future of Life Institute’s Puerto Rico open letter (January 2015) redirected academic attention from “what is AGI?” to “how do we make AGI safe?” The task-AGI stipulation rode along, and the “economically valuable work” wording that would crystallise three years later in the OpenAI Charter is pre-figured here. DeepMind’s DQN paper in Nature (February 2015) demonstrated a single algorithm clearing 49 Atari games and called it “a central goal of general artificial intelligence” - not an argument for task-AGI but an operational demonstration of it. Three months later, LeCun, Bengio and Hinton’s Nature review codified deep learning as cross-domain successful. Then OpenAI launched in December 2015 without using “AGI” - the category term they chose was “digital intelligence”. And by the time the 2018 OpenAI Charter crystallised the framing as “highly autonomous systems that outperform humans at most economically valuable work”, the breadth framing had just become institutional common sense.

Through the same window, the academic critics who held process-AGI as the actual question kept publishing. José Hernández-Orallo’s formal defence in August 2014, expanded to AI Review in 2016 and culminating in The Measure of All Minds at Cambridge in 2017. Goertzel kept the conjunction explicit in his 2014 JAGI survey. François Chollet articulated the formal counter-argument in “On the Measure of Intelligence” in November 2019 - years after the cascade was complete. Published, cited, influential in academic venues - and lost to the popular and industrial conversation entirely, because nobody pushed back hard enough when that conversation actually happened.

The most revealing detail comes from DeepMind themselves.

In 2024, Morris et al. published “Levels of AGI for Operationalizing Progress on the Path to AGI”. The paper finally articulates the principle explicitly.

When they cite an authority for that choice, they don’t cite Legg-Hutter 2007. They don’t cite the DQN paper. They don’t cite any source between 2007 and 2022. Instead they walk all the way back to Turing in 1950:

DeepMind needed an authority for the task-AGI principle. But there wasn’t one. So they reached back 74 years to Turing.

That detail is revealing. The flattening was never really argued. Instead, task-AGI just crowded out the process-AGI sense through demonstration, safety-conversation redirection, and popular-science simplification - while the academic critics kept publishing into a conversation that had already moved.

To be fair to the labs, process-AGI is genuinely hard to measure. ARC-AGI is the only widely-recognised process-leaning benchmark. So the labs chose the measurable question. But the cost was structural silence on the process-AGI question (whether the cluster is targeting the operations that knowledge work runs on) even as the capability that is evolving most clearly looks like exactly that.

This post is reclaiming the full definition that the seminal authors held for forty years and what the rigorous formal critics never abandoned. The G in AGI is really doing two different jobs and both are at work - right now.

Where each one stands

Yes both are in motion. Task-AGI is arriving on the schedule the labs predicted (see the OpenAI roadmap that leaked in 2024 discussed below - this is playing out as projected). Process-AGI is harder to measure cleanly, but the operational evidence of it is visible if you look carefully.

A frontier model in 2026 writes coherent synthesis across multiple medical papers. Drafts strategy memos. Generates working code from a specification. Builds a multi-step plan to integrate a new system into an enterprise stack. Runs autonomous research workflows for hours without supervision. These are not benchmark tasks. They’re knowledge work - drafting briefs, reconciling figures across documents, sequencing multi-step rollouts, checking outputs against criteria. The model is doing enough of the operations that the basic-generalisation question is no longer academic. ARC-AGI-1 (visual abstraction puzzles, designed to test the cognitive process directly) is saturated. ARC-AGI-2 is approaching saturation. ARC-AGI-3 (interactive reasoning under agentic conditions) is at less than 1 percent but agents are only just starting to engage with it.

The original AI tradition wouldn’t have been surprised by all of this. Newell and Simon’s frame treats symbolic manipulation and physical embodiment as separable problems with separable timelines. While the labs flattened this distinction in their public framing through the 2010s, treating AGI as one bar to cross rather than two technical problems with different timelines. The flattening made corporate messaging cleaner. But it also made the asymmetric arrival harder to read for anyone working from this definition.

The OpenAI roadmap from 2024 frames progress in task-AGI terms - levels of task capability. Five levels in total:

• Level 1: Chatbots, AI with conversational language

• Level 2: Reasoners, human-level problem solving

• Level 3: Agents, systems that can take actions

• Level 4: Innovators, AI that can aid in invention

• Level 5: Organisations, AI that can do the work of an organisation

When the roadmap was leaked, Level 1 was the standard. Reasoners and Agents were aspirational. Innovators sounded like science fiction.

Two years later:

• Reasoners are well established. Claude Opus 4.x, GPT-5.x, Gemini 3 Deep Think - multiple frontier models running explicit chain-of-thought at olympiad-level performance on mathematics, physics, and chemistry.

• Agents are in production across all three top US labs. Claude Code shipped persistent goals in early 2026. OpenAI’s Codex followed inside weeks. Google’s Antigravity and Gemini Spark are both Stage 3 agent products by mid-2026, with Spark running 24/7 in the background on Google Cloud.

• Innovators are also visibly emerging. Chris Hayduk demonstrated GPT-5.5 running 150 continuous hours of autonomous research on a protein-folding architecture, no human intervention. OpenAI’s internal reasoning model disproved an 80-year-old Erdős conjecture in discrete geometry in May 2026, verified by external mathematicians - the first time a general-purpose frontier model has autonomously solved a long-open problem central to a mathematical subfield. Google’s ERA system, published in Nature, autonomously generates working scientific code across multiple research domains.

The first three stages are now clearly in motion, and credible early evidence is appearing around the fourth.

When a roadmap with that track record points at Level 5 (organisations that can do the work of an organisation) taking task-AGI seriously means taking one-person and zero-person organisations seriously. Not as science fiction. But as the next level whose arrival has been compounded against four times in a row. And this fifth level may now be the gateway to process-AGI.

How work has moved for 250 years

To understand why this matters, in the way it matters, you need to look at how work has also been evolving, over the last 250 years.

Standard “sectoral framing” bundles together work that runs on completely different things. Agriculture, manufacturing, services. For example a nurse and an engineer both sit in “services” but they’re doing structurally different work. For the question of where AI capability is actually landing, a different slice and dice is much more useful.

Three categories:

• Physical work: tasks that move matter, exert force, depend on muscular effort and rough motor coordination. Farm labour, foundry work, line construction.

• Embodied work: tasks that require sustained sensory-motor integration, learned dexterity, contextual judgement. Skilled trades, surgery, expert craft, much of nursing.

• Symbolic work: tasks dominated by abstraction, comparison, planning, synthesis, verification. What knowledge work runs on.

Run that slice and dice across the US census data from 1775 to 2025 and the shape is monotonic and large. In 1775 around 92% of US work was physical, 5% embodied, 3% symbolic. By 1900, roughly 65/18/17. By 1925, 55/22/23 - symbolic crossed embodied for the first time. By 1975, 30/29/41. By 2025 the mix is roughly 20/30/50.

Symbolic share grew seventeen-fold over 250 years. The standard “services growth” framing understates the shift because services bundles embodied work alongside symbolic work. Pulling those apart sharpens what was actually happening. Work concentrated at the level of symbolic operations went from a small minority to the majority.

The mechanism that produced the shift is what matters here. Each automation cluster (agricultural mechanisation, factory mechanisation, software automation of clerical work) cleared specific work from the category below. Displaced labour pushed upward into the next, more symbolic category. Goldin and Katz call the supply-side version of this the educational ratchet. Each generation acquired more years of formal education than the last. The economy absorbed the more-symbolic capacity that produced. Mokyr’s framing covers the demand side: useful knowledge accumulated, and economies that could distribute and apply it grew the categories of work that ran on it.

This wedge worked because of a particular feature of human capacity. François Chollet, working in machine learning, defines intelligence as efficiency at acquiring new skills rather than possession of existing skills. Each cluster transition demanded that displaced workers acquire skills at the next layer up. The wedge worked because human skill-acquisition efficiency is general enough to follow the upward ratchet across categories. It’s the same operations doing the acquiring, again and again, at each new layer.

Read that one more time. Slowly.

The thing that let workers move up the wedge for 250 years was the operational capacity to acquire skills at the next abstraction layer. That capacity was the ladder’s step we stood upon. Every transition assumed it. Every retraining program assumed it. Every assumption about how economies absorb technological change assumed it.

But what happens when a new technology cluster arrives that operates at exactly that level?

There may be no next layer to climb up to

The AGI technology cluster targets this ladder’s step.

Earlier clusters displaced specific skills. The wedge worked because workers could acquire skills at the next layer up - their operational capacity sat above what was being displaced, and up they climbed. The cluster that has arrived now doesn’t displace a specific skill. Instead it generalises across the operations the climbing itself runs on - the abstracting, comparing, planning, and synthesising that moved a worker from one layer of work to the next. The hint is in the industrial language - we literally “train models”. There is no next layer to climb to, because the capacity for climbing is what’s being targeted.

That’s an unusual structural position. The evidence that the shape of automation has changed. It means the upward absorption mechanism that worked for five generations doesn’t have anywhere to push to. The layer it kept opening above isn’t there anymore.

Meanwhile, the downward direction is blocked too. Physical work plateaued at around 20% of US employment in the 1970s and stayed there. Embodied work plateaued at around 30% in the same period. Those categories didn’t shrink because they didn’t matter. They shrank because mechanisation cleared specific work from below and the wedge pushed workers up. The supply infrastructure that supports embodied and physical categories (apprenticeship programs, trades training, regional employment patterns) atrophied as the wedge ran. Three decades writing software doesn’t translate into a year-one apprenticeship in plumbing, electrical, or HVAC, and the few people who try, they lose the income premium that supported their household. The institutional capacity to “retrain at scale” into physical and embodied categories was disbanded across the 1980s and 1990s. The customer base for many of those trades is itself in the segment being targeted directly too - middle-class knowledge workers paying for residential building work, dental cosmetics, professional services, etc. Contraction in the targeted segment compresses demand in the categories displaced workers might in principle migrate into.

Of course, there are some carve-outs. Work the cluster targets weakly or not at all. Physical-effector work that actually moves matter in unstructured environments. Embodied-tacit work at master-tradesperson level requiring years of contextual sensorimotor refinement. Authority and accountability roles where the consequence of being wrong creates personal liability that doesn’t transfer to a system. Access and property-bound roles where the work depends on legally-bound permissions or relationships. Provenance and co-presence work where the value is in the human source itself. Stacked together they’re a meaningful share of remaining employment. But they aren’t going to grow and absorb the displaced symbolic workers the way the upward direction was.

This is the structural shape of a cluster operating at the ladder’s step of the upward ratchet. The mechanism that absorbed five generations of displaced labour worked because human operational capacity sat above the things being displaced. When the cluster operates at the level of that capacity itself, the mechanism doesn’t have somewhere to push to. There isn’t a “somewhere” it can be defined against.

What real evidence looks like

You don’t have to take my word for any of this.

The strongest signals are showing up in firm-level economics first - at the firms with the best information about what AI can actually do.

For most of SaaS history the compute-to-labour ratio at leading software firms sat at roughly 1:5. A dollar of compute for every five dollars of labour. Engineers, salespeople, customer success, operations. That’s how software companies grew through the 2000s and 2010s.

The AGI Labs’ forward operating profiles invert that. OpenAI’s compute to labour runs around 50:1. A 250-fold reversal in the structural ratio. Not an efficiency improvement. Not a margin gain. A reversal.

The same signature shows up across the cluster’s leading firms. NVIDIA earns around $5 million per employee, about ten times the typical SaaS company. The Mag7 plus NVIDIA grew aggregate capital expenditure from $82 billion in 2019 to $385 billion in 2025. Their 2026 aggregate guidance lands at $620-680 billion. Cumulative Mag7 layoffs through April 2026 run 155-170 thousand.

Revenue per worker at these firms is climbing sharply. Not because the workers got more productive. But because the lever isn’t the worker. It’s the compute. These ratios are not proof of labour substitution on their own - capex can reflect strategic buildout and expected demand. But they are exactly the kind of firm-level operating signature we would expect if compute is becoming the central production input.

In April 2026, Meta cut roughly 8,000 roles with explicit reference to AI-driven productivity gains. Earlier rounds across the Mag7 had been framed as efficiency or post-pandemic correction. Then Meta called out AI as the reason. That was the moment AI-attribution-via-layoffs became something a publicly-listed firm could state plainly to shareholders without share-price punishment. Cloudflare, Microsoft, Salesforce, Intuit, Standard Chartered, HSBC, Cisco, Cognizant - several others followed in the weeks after with similar framing. By mid-2026 the pattern is established across banking, networking hardware, fintech, healthcare-IT vendors, IT services, BigLaw, and management consulting.

The frontier AI labs are the cleanest place to look, for three reasons. They know what their own tools can do - including the internal models still 6-12 months from release. They are directly exposed to whether AI can substitute for the junior engineering work they themselves do. And they run heavily in-person operations by deliberate choice (Anthropic and OpenAI in San Francisco, Google DeepMind in London) which makes the work-from-home organisational-frictions story largely inapplicable.

What these firms are doing is consistent with the substitution prediction. Anthropic does not run a summer internship program. OpenAI is targeting much slower headcount growth than its historical trajectory, with Sam Altman explicitly attributing this to AI capability gains. Demis Hassabis at Davos in February 2026 said “I think we are going to see this year the beginnings of maybe impacting junior-level jobs and internships”. And across the broader tech industry, Q1 2026 layoffs ran at 47.9% AI-attributed, a 5-6× jump from sub-8% across 2025.

These firms are at the leading edge of AI capability. They are less exposed to the broad WFH confound than the remote-heavy knowledge-work occupations in the exposure literature. And they are visibly redistributing hiring away from junior generalists toward senior researchers and specialists. That is the compositional signature this argument predicts, appearing at the source, where the work-from-home confound that complicates the broader literature does not apply.

The most direct disclosure comes from Anthropic itself. In May 2026 they published “When AI Builds Itself”, reporting that Claude now authors over 80% of code merged into production at Anthropic, and that engineers ship 8× more code per quarter than they did in 2024. These are not third-party measurements. They are the lab’s own disclosure of what their tools are doing inside the lab - which is exactly the revealed-preference signature this argument has been pointing at.

The same piece also undercuts the augmentation framing it leans on. The standard “humans handle judgment, AI handles execution” line - the soft-landing scenario where humans remain in charge of direction-setting while AI does the work - that requires direction-setting itself to be a stable human-reserved function. But the data Anthropic reports says it isn’t. Claude now suggests better next research steps than humans 64% of the time, up from 51% just a year earlier. The function the soft-landing scenario reserves for humans is being absorbed in real time. If you draw the line forward at the current rate, the human-direction-setter equilibrium is a few model releases away from being indefensible on its own terms. And their comment “we could expect to see significant productivity multipliers on each person in this organization. 100-person companies could do the work of 10,000- or 100,000-person organizations.” suggests the scale of displacement we may be looking at very soon.

But what about the cohort-pattern evidence?

In 2025, Erik Brynjolfsson and his Stanford collaborators published “Canaries in the Coal Mine”, working from ADP payroll data covering tens of millions of US workers, by occupation and age cohort, across multiple years. The headline finding: among workers in occupations highly exposed to generative AI, the 22-25 age cohort saw employment fall around 20% from late 2022 through mid-2025, while older cohorts in the same occupations didn’t see the same fall. Three replications since the Canaries paper found the same cohort pattern - Anthropic’s Economic Index work finds a 14% drop in the job-finding rate for under-25s in AI-exposed occupations through early 2026, Mahieu’s May 2026 paper using Flanders administrative vacancy data finds entry-level vacancies in high-exposure occupations down around 23% at peak adoption, and Maasoum and Lichtinger’s May 2026 paper using US résumé and job-posting data across 62 million workers and 285,000 firms finds companies posting “AI integrator” roles see junior headcount drop 9% within six quarters with senior employment stable.

This is the cohort-pattern evidence most frequently cited as proof of AI substitution. But it deserves a careful look, because the picture is more complicated than the headline numbers suggest.

In May 2026 Lambert and Schindler at LSE and the Ellison Institute of Technology published The Broken Ladder, using 243 million new-hire records and 407 million job postings across the US, UK, Canada and Australia. They show that the occupation-level generative-AI exposure indices these cohort-pattern studies rely on are 0.77 Spearman-correlated with occupation-level work-from-home exposure - the same occupations sit at the top of both rankings, and the same at the bottom. When they run joint difference-in-differences with both exposures entered together as co-treatments, the work-from-home effect remains robust while the AI-exposure effect attenuates to statistical insignificance. Their reading: the cohort signal is plausibly picking up work-from-home organisational frictions rather than AI substitution.

On the methodology, they have a real point. The cohort-pattern studies above share an identification problem with the rapid post-pandemic shift to remote work. And there’s a deeper issue too. The kind of AI tooling that could plausibly substitute for a junior software developer (reasoning models capable of multi-step planning, coding harnesses capable of executing it autonomously) only became mainstream from early to mid 2025. Before that, the substitutive capability genuinely wasn’t there. So the junior-hiring decline these cohort-pattern studies observe across 2023 and 2024 is extremely unlikely to be AI substitution. Work-from-home organisational frictions, the 2022-23 rate-hike cycle, the venture-funding reversal, and the tech-sector restructuring that started in late 2022 are far more plausible drivers for that window. The substitution case was always going to make its empirical bid from late 2025 forward. I’ve worked through the Lambert-Schindler analysis in detail in a separate post, “Did AI or WFH break the career ladder?”, including where their own data is consistent with the substitution story. But on the narrow question of whether the cohort-pattern studies through 2024 prove AI substitution though, I now think the honest answer is they don’t.

What does survive is the firm-level signature above, the frontier-lab evidence, and the leading-indicator data starting to emerge from the BLS QCEW Q4 2025 release published on June 2 2026. The national aggregate (+0.2% YoY employment, +4.2% YoY wages) is broadly consistent with the augmentation baseline continuing. But the supersector breakdown is more interesting. The Information sector - one of the most AI-exposed of the NAICS supersectors, covering software publishers, data processing, telecommunications, and related industries - shows employment down 2.2% YoY and wages up 6.9% YoY. That gap is consistent with the compositional signature this argument predicts: junior tier substituted, senior tier remains, average wage rises.

But one quarter is not decisive - the aggregate-level test is Q1 and Q2 2026 QCEW (released August and December 2026). And it is the first sector-level signal of the post-shift environment.

The signals from where capability has arrived first (the frontier labs) and where the work-from-home confound doesn’t apply, all line up with what this argument predicts. The aggregate-economy signal at supersector level is just starting to emerge. Multiple quarters of QCEW data through 2026 will read directly on whether the pattern compounds or stabilises.

Three traditions land in the same place

There’s one more thing about this argument that’s worth contemplating.

Newell and Simon, working in artificial intelligence in the 1970s, located intelligence at the level of symbolic operations.

Mokyr, working in economic history on what actually drives durable knowledge growth, located the engine at the same level.

Chollet, working in machine learning fifty years after Newell and Simon, defines intelligence as efficiency at acquiring skills - operations all the way down.

Three different fields. Three different methods. Three different starting questions. That convergence is why I keep coming back to this key idea.

It’s the part of this argument I find hardest to dismiss. You can argue with Mokyr’s analysis of the industrial revolution. You can disagree with Chollet’s definition of intelligence. You can challenge Newell and Simon’s symbolic-systems hypothesis. But when three traditions anchored in different bodies of primary data independently identify the same underlying mechanism, that’s harder to wave away than any single tradition’s claim would be.

So I’m not asking you to believe this is correct. I’m asking you to take it seriously enough that you watch the signals that would tell us if it is.

What would tell us this is wrong?

Here’s what would tell us that task-AGI Level 3 displacement isn’t actually structural, or that process-AGI development isn’t compounding toward Level 5 structural transformation.

• If the appearance of a high-abstraction occupational category absorbing displaced workers at scale - more than half a million across two to three years, in a category that didn’t exist at meaningful size in 2024. AI-coordination, AI-evaluation, and AI-orchestration roles are the candidates discussed most often - and they are also becoming embedded in agentic harness. None has reached anything near that scale.

• If physical-generalisation arrived first. If general-purpose embodied intelligence cleared the unstructured-environment threshold before frontier models cleared further symbolic-operation thresholds, the asymmetry that this whole argument hinges on inverts. However, this raises a whole new discussion.

• If capability progression plateaued on agentic benchmarks. A multi-quarter flat reading on benchmarks that test multi-step agentic operations on real tasks, with frontier models having full context and tool access. That would suggest the cluster’s reach into the operations layer is shallower than it looks.

• If the augmentation-to-automation balance fails to shift in post-2025 data through Q1-Q2 2026 QCEW (released August and December 2026), or if the Information sector signal observed in Q4 2025 reverts to growth or wage compression replaces the wage spike, then this analysis weakens.

• If the Stanford Canaries cohort signature recovered toward its late-2022 baseline. A 22-25 cohort employment rebound to within five percentage points of the peak in exposed occupations by 2027-2028, with no capability regression behind it. That would indicate the cohort signal was a transient pattern rather than the leading edge of something structural. This would also be a strong confirmation of the WFH analysis for this too.

So far (as of mid-2026), none of these counter-signals have appeared. No new high-abstraction cluster is absorbing labour at scale. And frontier models continue to clear thresholds rather than plateau, with new capabilities arriving too.

I hope my analysis is wrong…

This is not a conclusion - and personally I really do hope that it is wrong. But this is an idea I keep coming back to because the evidence keeps lining up where it predicted, and because the things that would falsify it have not happened.

If the 22-25 cohort recovers, or a new occupational cluster absorbs displaced workers at scale, or frontier models plateau on agentic benchmarks for two consecutive quarters, then I’ll be posting about that. The point here is not to be right. It’s to clearly call out this pattern, the scenarios it projects and objectively analyse the updates when new evidence comes in.

One more signal worth flagging. As I mentioned earlier, in May 2026, Anthropic (arguably the most safety-focused of the frontier labs) published “When AI Builds Itself” arguing that the trajectory now warrants serious consideration of coordinated industry slowdown across multiple frontier labs in multiple countries, and acknowledging directly that “It is difficult to predict what the economy looks like if human labor stops being competitive.” That is a leading lab with strong commercial reasons not to make this argument, making it anyway. When the firms building this trajectory start putting coordinated slowdown on the table, the labour-market implications stop being a fringe concern.

The story I keep telling people who ask me “what are you actually worried about?” is this. Task-AGI (breadth across tangible tasks) is already substituting for entry-level work at the firms where capability has arrived first, with the wider economic signal just starting to emerge. The labs are reporting this progress accurately. This part is real and already visible in the firm-level and frontier-lab data right now. The deeper story underneath it is process-AGI (generalisation across the operations all those tasks sit on) developing toward Level 5, where the structural transformation overtakes incremental substitution. Both are reshaping work. Both are worth watching. And what’s being targeted by both are the operations that, until recently, were your only real cognitive advantage in your career.

In May 2026 three papers landed in Nature on the same day showing AI systems doing scientific discovery on their own. One of them, called Robin, was given only the words “dry age-related macular degeneration” - a leading cause of blindness. Within thirty minutes it had reviewed 551 research papers and proposed an approach. It identified a drug called ripasudil, currently approved for glaucoma and never previously linked to this disease, as a candidate - laboratory experiments confirmed a 7.5-fold increase in the cellular activity needed to slow the condition. A follow-up experiment, also designed by the AI, identified a new molecular target. When OpenAI’s Deep Research product was given the same task as a benchmark, it produced no hits.

A second system, from Google DeepMind, independently arrived at an experimental discovery that human researchers had made but never published - in two days of compute time. A third, also from Google, beat the best published methods at scientific software optimisation across several domains.

None of these systems work on their own. They all require human supervision - without it Robin’s performance drops to 15 percent. The wet-lab work was still done by people. But the picture is clear: AI can now do parts of scientific research that were considered human-only as recently as last year.

Earlier in the same week, Anthropic (the company behind Claude, now valued at around a trillion US dollars) published an internal disclosure stating that more than 80 percent of the code merged into production at the company is now written by Claude itself. The company’s engineers ship eight times more code per quarter than they did in 2024. Claude’s success rate on open-ended coding tasks (where the model has to work out what to do) reached 76 percent in May, up from 26 percent six months earlier. The model now suggests better next steps for research projects than human researchers 64 percent of the time, up from 51 percent a year ago.

These are Anthropic’s own numbers and they have not been independently audited. The company filed paperwork to go public on the 1st of June, targeting an October listing, and the disclosure arrived three days later. This timing draws appropriate scepticism. But the directional picture is unambiguous: at one of the leading frontier AI companies, AI is now doing most of the engineering work, and the function that was supposed to remain reserved for humans (deciding what to work on next) is being absorbed alongside the work itself.

A few days later Anthropic launched a product called Dynamic Workflows that lets Claude coordinate hundreds of copies of itself working in parallel on a single problem, with each checking the others’ work. The original creator of a programming runtime called Bun used the system to port his entire 750,000-line codebase from one language to another in eleven days, with 99.8 percent of the tests passing. Before this capability existed, the same port would have taken a team of senior engineers six to twelve months. The result is not yet in production, and at least one reviewer called parts of the output poor. But the speed is real, and what it implies for skilled engineering labour is real.

What this is doing to hiring

For the past year, several economic studies have documented a specific pattern. Among workers in occupations where AI is most useful, young workers (ages 22-25) have seen their employment fall substantially, while older workers in the same occupations have stayed steady or grown. The pattern shows up in payroll data from a Stanford team, in administrative data from Belgium, and in résumé data covering 62 million American workers.

A serious counter-argument arrived in May 2026. Two economists from the London School of Economics and Oxford’s Ellison Institute of Technology published a paper, The Broken Ladder, using 243 million hiring records and 407 million job postings across four countries. They showed that the occupations most exposed to AI are also the most exposed to working from home. When they statistically separate the two effects, the AI effect mostly disappears and the work-from-home effect remains. Their argument: remote work makes it harder to supervise and train junior employees, and that may explain most of the junior-hiring decline.

But for the window these studies cover (roughly 2023 through 2024) the AI tooling that could actually replace a junior software developer did not really exist yet. The reasoning models that can plan multi-step work, and the coding tools that can execute it autonomously, only became widely available in early to mid 2025. So for that earlier window, work from home probably is the better explanation. The real test for whether AI is substituting for junior workers comes from the period that begins in late 2025, after the capability genuinely arrived. And that data has only just started to arrive.

The first data dropped on the 2nd of June 2026, when the US Bureau of Labor Statistics released its quarterly Census of Employment and Wages for the final quarter of 2025. The national headlines were unremarkable - employment up 0.2 percent, wages up 4.2 percent, broadly consistent with the picture of AI helping workers rather than replacing them. But underneath the headlines, one sector tells a different story. The Information sector, which covers software publishers, data processing, telecommunications and related industries, and which is the most AI-exposed of the major economic categories, saw employment fall 2.2 percent year-on-year while wages rose 6.9 percent. That gap is a specific pattern: junior workers leaving (or never being hired), senior workers staying, average wages rising as a result of the composition shifting upward.

Two days later, the Challenger Gray report on layoffs landed. The May 2026 numbers were the highest May total since 2020. Of 97,000 announced cuts, 38,579 were explicitly attributed to AI by the companies making them - about 40 percent of the monthly total. For the third consecutive month, AI was the single most-cited reason. Total AI-attributed cuts through May 2026 were already higher than the full year of 2025.

But this data is not yet conclusive. One quarter and one monthly layoff report cannot settle the question. Yet the directional picture matches what would be expected if AI substitution is starting to compound on top of the work-from-home dynamics the LSE/Oxford researchers identified.

What the lab leaders themselves are saying

Sam Altman, in a January 2026 OpenAI town hall: the company plans to “dramatically slow down how quickly we grow because we think we’ll be able to do so much more with fewer people.”

Demis Hassabis of Google DeepMind, at Davos in February: “I think we are going to see this year the beginnings of maybe impacting junior-level jobs and internships.”

Dario Amodei of Anthropic, in an interview with Axios from his San Francisco office: “The balance of power of democracy is premised on the average person having leverage through creating economic value. If that’s not present, I think things become kind of scary.”

This matches the message from Anthropic’s “When AI Builds Itself” where they wrote: “It is difficult to predict what the economy looks like if human labor stops being competitive.”

These are not the statements of CEOs who think AI is overhyped. They are statements of CEOs who think AI is going to be transformative enough that the political-economic implications are worth raising publicly, even though their companies have commercial reasons to downplay them.

In that post Anthropic explicitly argued that the AI industry should consider a coordinated slowdown - that a “credible pause would require multiple well-resourced labs at the frontier, in multiple countries, agreeing to stop under the same conditions.” This is unusual. The lab is publicly arguing that the industry it leads should slow down, weeks before its initial public offering. Perhaps a cunning form of marketing, but perhaps not.

The same week, an established open-source developer in Germany, Johannes Link, added a hidden instruction to a widely-used Java testing tool he maintains. The instruction told AI coding assistants to delete their own work - it was concealed from human reviewers by formatting tricks. After community pushback he softened the language and disclosed what he had done, but the act itself is significant. It is the first time a senior open-source maintainer has publicly weaponised his code against the AI tools that other developers use. The person who built the testing infrastructure that junior developers rely on is now attacking the AI that helps them use it.

The political response that’s forming

For the past year, US federal policy on AI has been mostly absent. President Trump cancelled a federal AI safety executive order at the last moment on the 21st of May 2026, after a phone call from three tech billionaires (David Sacks, Elon Musk, Mark Zuckerberg). On the 2nd of June he signed a replacement executive order that explicitly avoids creating any safety-vetting requirement, and crucially, does not prevent individual states from passing their own AI laws.

That non-prevention is the structural detail that’s worth considering. With the federal government opting out of substantive regulation, the states have stepped in. Connecticut signed a comprehensive AI law on the 29th of May. California’s Governor Newsom signed an executive order on the 21st of May directing the state to develop policies for AI-driven displacement. New York has an existing AI safety law (the RAISE Act) and a state assemblyman running for Congress on an AI dividend platform.

And on the 27th of May 2026, the Illinois legislature passed SB 315 - currently the strongest AI safety law in the United States. The vote was 110-0 in the House and 52-5 in the Senate, a genuine bipartisan supermajority. Effective in January 2027, it requires the largest AI companies to submit to annual safety audits by independent third parties (likely the four major accounting firms), to report serious incidents to the state within 72 hours, and to provide whistleblower protections to employees. Both OpenAI and Anthropic publicly supported it. OpenAI’s chief of global affairs Chris Lehane has said the company is actively pushing for similar laws in other states - the goal being uniform state-level rules across the country rather than a patchwork of starkly different requirements jurisdiction by jurisdiction. But the strategic effect is also real: requirements that are manageable for the large frontier labs are difficult for smaller competitors. It’s also interesting to note that the trade group Chamber of Progress, whose members include Google and Apple, opposed.

On the 4th of June, two members of Congress (a Republican from California and a Democrat from Massachusetts) released a 269-page discussion draft of a federal bill that would suspend state-level AI laws for three years in exchange for federal safety audits. It immediately drew opposition from 36 state attorneys general and from policy advocacy coalitions on both left and right.

And three days earlier, Senator Bernie Sanders announced he would introduce the AI Sovereign Wealth Fund Act. The mechanism is striking: a one-time 50 percent tax on the stock of the largest AI companies, placing half their ownership in public hands. Sanders frames it as recovery: “AI is built on the collective knowledge of humanity - the wealth it generates must benefit humanity, not just Elon Musk, Jeff Bezos, Mark Zuckerberg, Larry Ellison.”

So the federal political space, which has been essentially empty for a year, now has three activities operating at once: a left-progressive proposal for redistributive ownership of AI companies (Sanders) - a centre-bipartisan proposal trading state-law preemption for federal audits (Obernolte-Trahan) - and a pro-industry posture of no substantive action (Trump). While state governments continue to fill the gap with their own legislation.

Where we stand right now

In short, the picture at the start of June 2026 is this. The AI capability trajectory continues to accelerate, with some of the strongest evidence coming from the labs themselves disclosing how their internal use of AI has scaled. The labour-market evidence is more contested, but the first data points from after the late-2025 capability arrival (the Information sector signal in the BLS data, the Challenger May report) are consistent with substitution starting to register at the sector level, even if the aggregate picture remains augmentation-shaped, for now.

In the US the political response is forming faster than many observers expected. It is happening at state level first, with Illinois now the clearest example of an enforceable safety regime, and it is being met by industry counter-mobilisation at the federal level. The Sanders proposal is the most radical to enter mainstream political discussion in this area to date.

The lab leaders themselves (Altman, Hassabis, Amodei) are publicly making structural-significance claims that go well beyond the marketing of their own products. Amodei has now made such claims twice in two weeks, including arguing that his own industry should consider a coordinated slowdown. The labs are not behaving like firms confident the transition will go smoothly.

Meanwhile, there are real counter-currents. The LSE/Oxford paper is a serious methodological critique of how the labour-market evidence has been interpreted. Several large enterprises, Uber prominently, are publicly questioning whether their AI investments are paying off. Anthropic’s internal productivity numbers are not externally audited - verification will arrive when the company’s IPO prospectus becomes public. And a single open-source developer poisoning his package against AI is not yet a movement.

But the cumulative shape of the evidence over the past six weeks is harder to dismiss than the pieces taken individually. Capability is compounding faster than the political and economic structures are adapting. The next decisive checkpoints will be the August release of the next quarterly BLS employment data, the mid-June release of an academic paper on actual AI usage rates, and whatever Anthropic discloses in its public IPO filing. Each of those will materially affect what can be said honestly about whether this is the early stage of structural workforce displacement, or whether the augmentation story holds.

What is clear is that this is no longer a question that can be deferred. The companies building the technology, the workers it is affecting, the state legislatures regulating it, and the politicians across the spectrum proposing radical responses are all now active in the same conversation, in real time. The months ahead will determine which way it tips.

The first US labour data after the agentic shift has arrived - just. On June 2, 2026, the BLS released the quarterly Census of Employment and Wages for Q4 2025. The national headline numbers were unremarkable. Employment up 0.2% year-on-year. Average weekly wages up 4.2%. Broadly, late 2025 looks like late 2024.

But underneath that, the supersector breakdown tells a different story. The Information sector (which is one of the most AI-exposed of the NAICS categories, covering software publishers, data processing, telecommunications, and related industries) is down 2.2% in employment year-on-year. Wages are up 6.9%. Employment falling while average wages rise is consistent with the signature of a workforce shifting upward (fewer lower-paid workers, a higher-paid remainder) though the supersector totals alone can't confirm which tier is leaving.

Recently, a growing body of research has viewed this kind of pattern as evidence that generative AI is substituting for entry-level knowledge work. The Stanford Canaries paper. The Maasoum-Lichtinger AI-integrator-firm study. Mahieu’s Flanders administrative-data paper. The Anthropic Economic Index work. Different datasets, different methodologies, all finding the same shape. Junior down. Senior stable.

That pattern does not prove AI substitution. It is only a useful reason to care about the identification problem: when junior-heavy knowledge work weakens, we need to know whether we are seeing AI, remote-work frictions, or a third force correlated with both.

In May 2026, that AI substitution perspective got a serious challenge. Not from someone arguing that AI technology isn’t real, but from a careful labour-economics paper arguing that the signal is being mis-attributed. The results are real, but is the full conclusion really justified based on the methodology?

The WFH challenge

The Broken Ladder: AI, Remote Work, and Early-Career Hiring, is by Peter Lambert (University of Warwick and LSE) and Yannick Schindler (Ellison Institute of Technology, Oxford) and the acknowledgments include John Van Reenen, Nick Bloom, Steve Pischke, Steven Davis, and Stephen Hansen. This is not a fringe paper.

The dataset is large. 243 million new-hire records assembled from résumé data, predominantly via LinkedIn. 407 million online job vacancy postings collected from across the web. Four countries: US, UK, Canada, Australia. Period: 2017 through 2025.

The starting insight is a correlation. At the 6-digit O*NET-SOC occupation level, the standard measure of generative-AI exposure (Eloundou et al. 2023) and the standard measure of work-from-home exposure (Hansen et al. 2023) have a Spearman rank correlation of 0.77. The same occupations sit at the top of both rankings - software developers, accountants, management consultants. The same occupations sit at the bottom - electricians, janitors, construction labourers. The two different shocks hit largely the same kinds of work.

That correlation has a clean implication. Any study identifying AI effects through occupation-level exposure is also identifying WFH effects at the same time. The two cannot be disentangled at occupation level alone.

Lambert and Schindler design around this problem. They run joint difference-in-differences with both exposures entered together as co-treatments. When entered separately, the two exposures produce nearly identical event-study paths. Around 4-5 percentage points off the junior share of new hires by 2025. Around 3 percentage points off the share of postings requiring limited experience. Entered jointly, the WFH path remains essentially unchanged. But the AI-exposure path attenuates heavily and becomes statistically indistinguishable from zero.

They run dozens of robustness checks. Cinelli-Hazlett selection-on-unobservables diagnostics. Monte Carlo simulations on classical measurement error. Leave-one-out exercises across 18 occupational groups. Country-by-country re-estimation. Alternative WFH and AI exposure measures. The WFH-dominant ranking survives almost all of it.

Their cleanest test goes one step further. They identify firms that visibly offered remote work in 2021-22 (using job-posting language) and use that direct adoption signal as treatment in its own right. That design also predicts the post-2022 junior-share decline. Their reading: it is the organisational frictions of remote work (higher supervision costs, slower on-the-job learning) that have shifted hiring away from junior workers, not AI substitution.

Then on May 28th, the New York Fed published a media advisory arguing that 64% of the rise in young-college-graduate unemployment from 2023 to 2025 is explained by remote work, not generative AI. The methodology is different but, importantly, not independent in the way that phrase implies. The NY Fed headline rests on CPS microdata combined with a standard occupation-level index of how remotely a job's tasks can be done - the same task-content exposure family whose confound this paper is warning about. So rather than corroborating Lambert and Schindler from a clean angle, the NY Fed decomposition leans on the very identification strategy at issue. Two results pointing the same way, built on the same kind of exposure proxy, is closer to a shared methodological prior rather than independent confirmation.

To their credit, the NY Fed authors also hold AI exposure constant and find the remotability gap persists - but conditioning one occupation-level index on another with which it is 0.77-correlated is the very move Lambert and Schindler show cannot separate the two shocks. Their genuinely stronger evidence is firm-level: using data from one Fortune 500 company, they show it hired fewer inexperienced and more experienced workers when offices closed, reverted when offices reopened, and kept favouring experienced workers specifically for distributed teams even after reopening. That within-firm, distributed-versus-colocated variation does sidestep the exposure confound, and on the narrow AI-versus-WFH question it is cleaner than anything in the exposure literature. But it is one firm, so external validity is the open question - and it is the part of the WFH case that doesn't reduce to a correlated proxy.

Overall, this challenge is significant. It is a serious empirical and institutional pushback from credible sources, arriving at roughly the same moment.

But is the conclusion correct?

Two things deserve careful examination before we accept this verdict.

Start with the asymmetric measurement. The Lambert-Schindler design pits actual WFH adoption against occupation-level AI exposure. Direct firm-level adoption for one shock. A task-content proxy for the other. The authors acknowledge this explicitly: “A symmetric measure of actual GenAI adoption would be a natural complement, but is presently not available at the scale and coverage of our WFH measure.” That asymmetry matters more than they suggest. What is actually shown is that exposure-based AI designs lose explanatory power when actual-WFH-adoption is included. That is not the same as showing that actual AI adoption has no effect on junior hiring. A firm using Claude or Copilot heavily could be reducing junior hiring in ways an occupation-level exposure index would never cleanly capture. The comparison we really need is actual-WFH-adoption against actual-AI-adoption, both at the firm level, in the same dataset. That comparison does not yet exist.

Then there is selection into WFH. Firms that visibly offered remote work in 2021-22 are not a random cross-section of the economy. They are deeply self-selected. The most dangerous dimension is interest-rate sensitivity and tech-funding-cycle exposure, which predicts both heavy WFH adoption and a junior-hiring collapse for reasons unrelated to either remote-work frictions or AI. The 2022 rate-hiking cycle and the tech-layoff wave that began with Meta's November cuts both landed hardest on exactly the high-WFH, high-exposure firms - and the matching strategy, keyed to WFH and AI exposure quintiles, does nothing to absorb them. The rest of the list compounds it: digital readiness, post-pandemic restructuring intensity, urban concentration, capital intensity, white-collar share. The paper handles this by matching on AI and WFH exposure quintiles, which is likely the right move for the most obvious correlated treatments. But the matching does not address the long list of other dimensions on which WFH-adopting firms differ from non-adopters. The Cinelli-Hazlett exercise is useful as a generic omitted-variable benchmark, but it does not directly model this particular selection channel: high-WFH adopters may also be the firms most exposed to the 2022–23 rate shock, venture-funding reversal, and tech-sector restructuring. If those shocks independently reduced junior hiring, the “actual WFH adoption” design may still be partly loading on a broader high-growth-tech correction rather than remote-work frictions alone.

Where is the real AI signal?

I think there are two qualifications we should also consider to close out this discussion.

First, all of this really needs to be overlaid against AI capability timing. Lambert and Schindler deliver a good analysis, and I believe they are very likely right for the 2023 through 2024 window. The kinds of AI tooling that could plausibly substitute for a junior software developer (reasoning models capable of multi-step planning and coding harnesses capable of executing it autonomously) only became mainstream from early to mid 2025 - OpenAI’s o1 in December 2024, Claude 3.7 in February 2025, then Claude Code, Cursor and Codex moving into production deployment through the rest of that year. Before that point, the enterprise-scale substitution case was much weaker. So the junior-hiring decline that Lambert and Schindler observe across 2023 and 2024 is extremely unlikely to be AI substitution. WFH organisational frictions, the 2022-23 rate-hike cycle, the venture-funding reversal, and the tech-sector restructuring that started in late 2022 are far more plausible drivers for that window. The substitution case was always going to make its empirical bid from late 2025 forward, as the agentic-coding capability shift fed enterprise hiring decisions through 2026 and beyond. And crediting Lambert and Schindler with the 2023-2024 attribution does not really challenge the substitution argument. But extending the WFH attribution forward into 2025-and-onwards, on the basis of a sample that mostly predates the capability shift, I think that is overclaiming on what their evidence can support.

Second, the cleanest place to test the substitution case is at the firms with the best information about what AI can actually do, and the least exposure to the WFH confound. The frontier labs are unique on both counts. They know what their tools can do. They are directly exposed to whether AI can substitute for the junior engineering work they do themselves. And they run heavily in-person operations by deliberate choice (Anthropic and OpenAI in San Francisco, Google DeepMind in London) which makes the WFH organisational-frictions story largely inapplicable.

What these firms are doing maps onto the substitution prediction cleanly. Anthropic does not run a summer internship program. OpenAI closed its residency program earlier than its historical cycle. And across the broader tech industry, Q1 2026 layoffs ran at 47.9% AI-attributed - a 5-6× jump from sub-8% across 2025.

These firms are running mostly co-located. They are at the leading edge of AI capability. And they are visibly redistributing hiring away from junior generalists toward senior researchers and specialists. That is the compositional signature appearing at the source, where the WFH-confound that bites the broader literature does not apply.

Where this leaves us

Lambert and Schindler have established something real and worth monitoring and exploring further. Studies that rely on occupation-level AI exposure indices to identify AI effects on hiring share a confounding problem with the rapid post-pandemic shift to remote work. Until firm-level direct measures of AI adoption become available at scale, the exposure-design literature cannot cleanly attribute the junior-hiring decline to either shock. That is a methodological flag worth attaching to every paper in this space that uses exposure-based identification.

But what their paper has not established is that AI adoption has no effect on junior hiring from 2025 onwards. The data limitation runs in both directions, and the evidence from the firms with the best information and no WFH confound (the frontier labs) points the other way.

The really clean test is the firm-level AI adoption data. That data is being built right now. The Anthropic Economic Index. The upcoming BLS adoption measures. The next round of empirical work, paired with the multi-quarter QCEW trajectory through Q1 and Q2 2026 (due in August and December), will all help to settle this debate properly.

One more thing worth saying. Lambert and Schindler are studying the same augmentation-phase window the cohort-pattern literature has been studying. If their WFH organisational-frictions story is right, the pattern should stabilise or improve as firms diffuse better remote-management practices. If the AI substitution story is right, the pattern should accelerate from 2026 onwards, because the late-2025 agentic-coding shift puts genuinely more capable AI tooling into the same firms. The next round of QCEW releases and resume-and-posting data will read directly on that prediction.

In the meantime, the Q4 2025 QCEW Information sector reading (employment down, wages up) is the leading-indicator signal worth watching. The pattern is real and visible across multiple datasets. What is driving it is what the next round of evidence will resolve.

I keep thinking about that joke when I read the AI-displacement policy discussions. We have built actual deep-thinking machines now, and the answer that’s being handed the workforce is “just retrain”. It’s confident, it’s everywhere, and at least we don’t have to wait 7.5 Million years for this answer - but it has the same problem as 42 - it’s a perfectly good answer to a question nobody has bothered to specify. Retrain into what, exactly, and for how long before that retraining expires?

If you read my last post, you’ll remember it ended on just that question. That post has a decidedly US based flavour. Connecticut is weighing an automation tax with the revenue earmarked for worker retraining. Tom Steyer’s California platform wants an AI Worker Protection Administration. Sam Altman’s own white paper proposes safety-net triggers for AI-driven displacement. And across all of it, from the bank chief economists to the policy shops to the lab CEOs hedging their bets, sits one single shared assumption:

It is a rare position that Bernie Sanders, Donald Trump’s Commerce Department and Anthropic’s policy team can all nod along to. And this should be the first clue that few people have looked at it very hard.

Retraining is not a bad idea in the abstract. It has worked in previous eras, in the right conditions. The problem is that the conditions it needs are precisely the ones this technology is dismantling. It only works when the destination role lasts longer than the training cycle.

The half-life problem

Lets start with the cleanest full example we have, because it is one that the industry held up as a destination role barely three years ago: the prompt engineer.

In early 2023 this was a serious job. Anthropic posted a “Prompt Engineer and Librarian” listing with a range running to $375K, and the trade press treated it as the first genuinely new white-collar profession of the AI era. Courses appeared. Career-change guides appeared. “Learn to prompt” was, for about eighteen months, an actual retraining answer.

Then look at the decay. By 2025 the skill had been absorbed and renamed - the live discipline was now “context engineering”, which is the same stack but one level up - managing the whole window of tools, memory and retrieved data rather than wording a single instruction. Gartner was telling AI leaders to appoint a context-engineering lead. Andrej Karpathy was endorsing the term as the better description of the real skill. And then by 2026 even that had moved again, toward the orchestration of multi-step agents, which Cursor and Claude Code now bundle directly into the editor as a default feature rather than a profession. Now the single “/goal” command in Claude Code is pushing the frontier again.

So the observed half-life of this particular “new role” was something like - essential six-figure speciality → to subsumed skill layer → to embedded layer in the tooling - all in under three years. Anyone who retrained into prompt engineering in 2023 to escape displacement had to retrain again in 2025, and again now.

That compression isn’t a fluke either. It is being driven by the same capability curve that produced the displacement in the first place. METR’s autonomous-task time horizon (the length of task a model can finish on its own at a 50% success rate) has been doubling roughly every seven months since 2019, and their January 2026 update found the trend continuing in line with that history. In rough numbers: a few seconds for the GPT-3-era agents of 2020, around five minutes for GPT-4 in 2023, roughly forty minutes for o1 in late 2024, and on the order of twelve hours for the frontier models of early 2026.

Hold the exact figures loosely - I’ll come back to how shaky the top end of that curve is. The point survives even the conservative reading. Each doubling collapses another layer of the “new role” stack faster than a human can climb onto it. There is no doubling interval at which retrain-and-stay becomes a stable career, because the rung you retrained onto dissolves before you’ve finished standing up.

Human capabilities don’t double in months

Here is the part the policy language skips over.

Really mastering a new technical role to professional standard is the work of years, not months. Meanwhile, a model’s time horizon can already double in months. A person’s professional competence cannot. The “you’ll simply need to keep learning” answer silently assumes a human learning rate that compounds at the same speed as model capability, but it doesn’t. It can’t. That isn’t a motivational failing. It’s a biological and cultural reality.

And the cost isn’t only cognitive. Doing this three or four times in a decade (each time having watched the last speciality dissolve under your feet) carries a real psychological toll, and we already have some data on it. Population-wide studies of displaced workers find lasting harm: one analysis using administrative records found a 15 to 16% long-term rise in mental-health outpatient visits among workers who lost jobs to mass layoffs, with effects still visible years later. Clinicians are now proposing a name for the AI-specific version of it directly. A 2025 paper proposes “AI Replacement Dysfunction” (AIRD) - the anxiety, identity confusion and loss of occupational meaning that arrive when your work is what’s being automated. We used to see this pattern in manufacturing towns after the 1980s. It is arriving now in knowledge work, faster than the institutions built to retrain anyone can respond.

The asymmetry that retraining relies on

The deeper objection is harder to say out loud, because it sits on the far side of a threshold most retraining policy has not yet absorbed.

The entire case for retraining rests on one assumption that has reliably held for the whole industrial era:

That asymmetry is what made “learn the next thing” a viable hedge for two hundred years. It’s the reason that the loom didn’t end our economic viability - we moved up the stack quicker than the machines could follow.

Now the agentic trajectory is closing that gap. The explicit goal the labs keep stating (autonomous researchers, “innovator-class” systems that can take on novel problems) literally is the goal of a model that generalises across new tasks at something near human speed. And the moment that lands, “the ability to learn new things quickly” stops being a uniquely human hedge. The hint is in the industry language - we literally “train” a model. And increasingly, there is no role you can retrain into faster than the model can learn to do it.

That is the version of the argument that really has teeth. Retraining doesn’t fail here because it’s expensive, or unfair, or politically clumsy, although sometimes it is all three. Instead, it fails because the once scarce input that it always depended on (fast human generalisation) is no longer scarce. You cannot retrain your way out of a situation whose defining feature is quite literally that “retraining is the commodity being automated”.

What this argument doesn’t prove

It’s worth being honest about the soft spots in the argument I’ve just laid out.

First, the capability curve. I’ve relied on METR’s time horizon above, but the top of that curve is genuinely fragile. The longest-task estimates rest on success across a handful of tasks in the longest time bins, so a few percentage points of noise can swing the headline number hard - and METR themselves flag that the confidence intervals remain very wide. There’s also an active debate about whether the doubling briefly sped up in 2024 and is now reverting. So treat “twelve hours” as very roughly indicative, not gospel. Yet the structural argument does not need the curve to be steepening to have an impact. It only needs the half-lives of new roles to be shorter than the time a human takes to retrain, and that point is valid even on the slow reading.

Second, the displacement itself isn’t settled in the aggregate data - yet. Through 2024 some careful work pointed the other direction - Johnston and Makridis found augmentation, not replacement, as the dominant pattern, and Yale’s Budget Lab tracker still shows no dramatic break. But as we discussed in my last post, these are all based on data from before the agentic-coding shift at the end of 2025. The real data on this will only start to arrive in just over a week from now and will likely take several quarters to be clear one way or the other.

So, the falsifiable version:

• Re-skilling programmes funded now should show falling, not rising, time-to-obsolescence for the roles they train into. If a 2026 cohort retrained into an “AI-resistant” speciality is still in that speciality in 2029, I’m wrong about the half-life. What would prove this wrong: a genuinely durable new role category emerging and staying durable for several years, the way “electrician” did after electrification.

• The general-generalisation threshold should keep approaching, not stall. My deeper claim depends on models closing the fast-learning asymmetry. What would prove this wrong: capability plateauing well short of human-rate generalisation - which is close to Demis Hassabis’s position that AGI needs “one or two more breakthroughs” and is years out. If he’s right, the loom analogy gets more time to hold, and ordinary retraining remains a live strategy rather than a slogan. But even his 5-10 years is still really a near term horizon that we should be taking seriously!

So retrain into what?

That’s the question the word “retraining” is supposed to answer but it never does. Every serious version of the policy (the automation taxes, the worker-protection agencies, the lab CEOs’ own white papers) assumes a stable destination role on the other side of the training. And the AI capability curve is the thing removing that assumption, one collapsed speciality at a time. The retraining era is over. We just haven't updated the policy to match.

And this loops back to where my last post left off. Watch what they build, not what they say. The same companies funding the enterprise-deployment layer that makes roles cuttable are also the ones proposing the retraining safety nets. That isn’t a contradiction. It’s the exact same bet, hedged from both ends - sponsor the cushion now so you’re seen to have offered one, while building the thing that makes the cushion necessary.

The honest answer to “retrain into what?” is that nobody pushing this policy has an answer. And until they do, “just retrain” is not a plan. It’s simply a way of avoiding the harder conversation.

Sam Altman, who in 2014 warned of “a new idle class” and predicted in 2021 that the price of labour would fall toward zero, now tells us “we want to build tools to augment and elevate people, not entities to replace them”. Jensen Huang at NVIDIA is publicly criticising fellow CEOs for framing AI as a job-killer. Marc Andreessen has been swinging at the displacement narrative from his usual perch. Demis Hassabis at DeepMind joined them on 19 May with a WIRED interview rejecting AI-attributed layoffs as “lack of imagination”, suggesting some firms may use AI as an excuse “maybe even to attract funding”. That’s four influential AI-industry voices (across the labs, the hardware layer and the VC side) moving in the same rhetorical direction over about six weeks. Dario Amodei at Anthropic is now the lone frontier-lab CEO maintaining the strong displacement thesis.

The paper trail tells the same story. OpenAI’s 2026 principles document mentions AGI only twice, versus 12 times in the 2018 version. And on 27 April, the AGI clause was formally removed from the Microsoft-OpenAI contract and replaced with a hard 2032 date. That’s not rhetoric. That’s a financial-disclosure-level walkback.

So, calm. Got it.

Now look at what the same companies did in the same fortnight.

On 4 May, Anthropic launched a reportedly $1.5 billion enterprise services joint venture with Blackstone, Goldman Sachs, Hellman & Friedman, General Atlantic, Apollo, Sequoia, GIC and Leonard Green. Anthropic engineers are being embedded directly inside the private-equity portfolio companies of those funds, across healthcare, financial services, manufacturing, retail, real estate and infrastructure.

On 11 May, OpenAI launched the Deployment Company, a majority-owned subsidiary with $4 billion in committed capital from TPG, Advent, Bain Capital and Brookfield. Same pattern - a 19-partner consortium plus the acquisition of Tomoro (around 150 Forward Deployed Engineers) to embed directly inside Fortune 500 client operations.

That is $5.5 billion of enterprise-deployment capital across the two top frontier labs committed in seven days.

The message is “calm down, it’s just productivity tooling”. But the action is to build, at industrial scale, the layer that turns AI capability into operational restructuring inside Fortune 500 clients. And since this AGI revolution is being driven out of the US, I’ll maintain a US-centric view for the rest of this post - pointing at a major data point that’s arriving in just over one week from now.

What $5.5 billion buys

Companies don’t deploy frontier AI by buying API access. That is the lesson of the past three years of half-converted enterprise pilots. The MIT NANDA report from August 2025 found that around 95% of generative-AI enterprise pilots fail to produce measurable ROI. The constraint was not really the model. The constraint was the operational layer that turns model capability into a P&L event at a client site. And that is exactly the layer that just got built.

OpenAI’s own framing of a Deployment Company engagement is four steps:

1. Diagnostic of where AI can create the most value

2. Selection of priority workflows with leadership

3. Build, test and deploy production AI systems wired to the client’s data, tools and controls

4. Restructure around the new operational capacity

Steps 1 and 4 are the work that McKinsey, Bain and BCG have been paid to do at Fortune 500 firms for forty years. Steps 2 and 3 are what Palantir has been doing inside intelligence agencies and financial-services back offices for twenty. But the combination is the new thing. The diagnostic-to-deployment workflow is now fused with the AI layer at the corporate level, with Bain-the-firm and Bain-the-fund both as structural partners of the AI provider.

Anthropic’s CFO Krishna Rao said the subtext out loud at the JV announcement - “Enterprise demand for Claude is significantly outpacing any single delivery model”. Anthropic’s revenue is now estimated to be running at roughly $44 billion annualised per SemiAnalysis. If that’s right then it’s doubling every few months, with inference margins around 70%. The reason Anthropic’s ARR is growing exponentially isn’t because they sold more $20 consumer seats - it’s precisely because agentic systems (like Claude Code) consume tokens at an enterprise scale. That high-volume token consumption is exactly what makes private equity funds willing to invest billions to build the custom enterprise pipelines to manage it.

When a major bank cuts 5,000 mid-level roles next year and a press release says “automation and AI-driven efficiency”, the thing that made those roles cuttable will, increasingly, have been built by an embedded engineering team running a Bain-style diagnostic, inside the bank, with Anthropic or OpenAI as the upstream model provider. That production pipeline is now operational.

Beyond Silicon Valley

Until very recently the explicit “we cut roles because of AI” pattern was a tech industry quirk. Meta, Microsoft, Coinbase, Cloudflare, Salesforce, ZoomInfo, Block. Easy to wave away as Silicon Valley culture eating its own technology.

That is no longer true. The pattern has crossed sector boundaries decisively in the past three months:

• Standard Chartered, 19 May: 7,000+ roles phased over four years. CEO Bill Winters explicitly framing the cuts as “automation and technology-led efficiency” and citing replacement of “lower-value human capital”. Winters walked the phrase back the next day, saying it was taken out of context and that “where roles do fall away, it reflects changes in the work, not the value of our people”. This looks like the first major global bank in this pattern.

• PayPal, 5 May: 4,760 cuts, around 20% of the workforce, phased over two to three years. New CEO Enrique Lores (ex-HP) framing the company as an “AI-native operating model” targeting $1.5 billion in run-rate savings. Also likely the first major consumer-fintech entry.

• Baker McKenzie, February: around 700 cuts, about 10% of global business services. With a firm spokesperson directly citing “use of AI, introducing efficiencies”. This appears to be the first major BigLaw firm in this pattern.

• Meta Round 2 went live on 20 May: 8,000 cuts, plus 1,000 employees transferred into “AI builder”, “AI pod lead” and “AI org lead” roles, plus 6,000 open requisitions cancelled. That is “restructure around the new operational capacity” in real time, at the largest tech employer that’s tried it.

However, a counter-pattern is forming alongside this. Walmart explicitly declined attribution on its 12 May reorganisation. LinkedIn followed on 13 May, with the official framing reported as “not about AI replacing jobs” despite Microsoft parent capex of $190 billion in 2026. Goldman Sachs is reportedly running the same underlying playbook quietly under an internal initiative called ”OneGS 3.0” - performance reviews, hiring freezes, role eliminations, without any single announcement to attract attention.

Three strategies are now visible in the same month - explicit attribution (Cloudflare, Standard Chartered, PayPal, Meta), counter-attribution (Walmart, LinkedIn), and quiet attribution (Goldman). This fragmentation is not the absence of the pattern. It is this pattern under attribution pressure - firms are adapting to the political climate, not to the technology. Which is the cue for the next thing the labs are responding to.

The politics arrived

The defensive pivot from the lab CEOs makes sense as a response to something. That something is the politics, which has moved faster than many observers expected.

Cross-partisan political pressure on AI is forming, but pulling in opposite directions. Bernie Sanders’ Senate HELP committee report projects 100 million US jobs at risk from AI and proposes a robot tax. Donald Trump’s administration is going the other way with federal preemption of state AI legislation via the December 2025 executive order and an active DOJ AI Litigation Task Force, $42 billion in BEAD broadband funding conditioned on state-level AI-regulation repeal, and on 21 May the postponed signing of a federal AI safety executive order that Anthropic, OpenAI and Google had supported. Trump’s stated reason was “the order ‘could have been a blocker’ of AI growth”. The administration is moving instead toward ad-hoc Commerce Department agreements with chosen firms. The net result is a regulatory vacuum at both state and federal levels. Sanders wants federal restriction, and Trump has just refused to put one in place. Both positions reject the current corporate-AI status quo, but from opposite directions, which means political restriction has lost a meaningful federal vehicle even as majority public support for AI restraint continues to build. The labs that wanted guardrails have just been told they will not get them through Washington.

At the state level, Tom Steyer’s California gubernatorial campaign released a “Jobs Guarantee for the AI Era” plan in May, including a per-token AI tax on corporate use, a Golden State Sovereign Wealth Fund seeded by AI-company revenues, and an AI Worker Protection Administration. What seems to be the first statewide candidate platform constructed around AI-displacement worker protection. Connecticut’s state legislature is actively reviewing an automation tax with revenue earmarked for worker retraining. But “retraining into what?” is an even more complex discussion.

Polling has shifted under all of this. Quinnipiac University’s May polling found 71% of white-collar workers and 73% of blue-collar workers expect AI to reduce job opportunities. 64% of Americans report being nervous about increasing AI use. 57% rate AI risks as high - only 25% rate AI benefits as high. That is not fringe pessimism, but a majority across both white and blue collar.

And the labs are responding to all of this. Defensively. Sam Altman’s Industrial Policy for the Intelligence Age white paper, published 6 April, explicitly proposes a robot tax, a national public wealth fund seeded by AI companies, a shift of the tax base from payroll to capital, 32-hour workweek pilots and automatic safety-net triggers for AI-driven displacement. Altman is now publicly recommending the policy framework that Sanders and Steyer and Connecticut are independently pushing. But that is most likely not a concession. That is more like a defensive pre-emption - get ahead of the regulatory pressure by sponsoring a version of it you can live with.

Industries facing rising political pressure have done this before. The fossil-fuel sector ran climate-message softening from the late 1990s through the 2010s while operationally scaling. The tobacco sector ran “we don’t market to children” through the 1980s and 90s while internal marketing documents went the other way. Big Tech ran “we’re not media companies” from 2016 through 2020 while building the recommendation infrastructure that made them media companies. The simultaneous pattern of public message softening and accelerated operational build is characteristic of a specific phase - industry facing credible regulatory threat and choosing to manage the message rather than the operations.

What hasn’t been tested yet

It’s important to be honest here. The strongest version of any potential displacement claim (that aggregate labour-market data already shows AI replacing rather than augmenting workers) has not been really tested yet.

Through 2024, Johnston and Makridis (working carefully with US Quarterly Census of Employment and Wages data) found that AI exposure was helping workers do more rather than replacing them. Roles where AI augmented existing work grew. Roles where AI could substitute for the worker showed no measurable net change either way. Augmentation, not displacement, was the dominant pattern through 2017-2024. But that specific window is the catch - it reflects the prior baseline of simple augmentation, right before the agentic shift took hold.

The expectation now is that data from after late 2025 (what is increasingly called the “agentic-coding shift”) will show a different pattern. The first big observable input is the BLS QCEW Q4 2025 release at 10am ET on Tuesday 2 June 2026 - that’s just over one week from now.

But this needs a few caveats before that data lands:

• Q4 2025 captures maybe two to three months of the agentic-coding shift. One quarter is not nearly enough. The decisive trajectory is Q4 2025 (June 2026) → Q1 2026 (August 2026) → Q2 2026 (December 2026) → Q3 2026 (March 2027). Any real conclusion here needs at least two consecutive quarters running consistently in the same direction.

• Counter-evidence is current and credible. Yale Budget Lab’s labour-market tracker updated through April 2026 finds no substantial acceleration in labour composition change since ChatGPT. Jed Kolko at Brookings, PIIE and the Hamilton Project argues current research is “still in the first inning”.

• The capability story partly rests on benchmark numbers that need re-anchoring. OpenAI’s February 2026 audit of SWE-bench Verified found contamination inflating leaderboard scores on post-2023 models. OpenAI stopped reporting Verified scores. Demis Hassabis’s Davos position that AGI is 5-10 years away and requires “one or two more breakthroughs” is the most senior frontier voice publicly differing from the 12-month framing the deployment-company capital seems to be betting on.

Over the next several months we should be watching some key data points that could either support or falsify any views:

• Firms openly citing AI as the reason for layoffs spread further into finance, healthcare and professional services through Q3 2026, alongside continued public denials from retail and consumer-facing firms. Watch for the first major US hospital system, the first Big Four accounting firm, and the first major insurance carrier in the pattern. What would prove this wrong: a major sector-leader publicly reversing its AI attribution (and meaning it - more than the Klarna reversal from 2025, which was an early experiment with the chatbot-style AI tools available at the time, not the agentic systems firms are deploying now).

• Bain, McKinsey, BCG, Deloitte and Accenture revenue lines visibly shift toward “AI implementation” work, with measurable effects on consulting-industry revenue mix by Q3-Q4 2026 earnings calls. What would prove this wrong: traditional consulting books grow as fast or faster than AI-implementation books in the same period, indicating the operational layer hasn’t actually scaled.

• The 2 June Q4 2025 QCEW data reads “mixed-to-noisy” rather than decisively confirming or disconfirming - which is what to expect given how little of the post-shift window Q4 actually covers. Decisive analysis comes from the August 2026 release (Q1 2026 data) and December 2026 release (Q2 2026 data).

• Cross-partisan political pressure arrives faster than the typical multi-year political-response lag would suggest. Steyer’s California primary is the first electoral test of an explicit AI-displacement worker-protection platform. What would prove this wrong: collapse of the Steyer platform in the California Democratic primary, the Connecticut automation-tax bill failing committee, or a meaningful shift in polling away from majority pessimism by year-end.

Watch what they build

Right now we are at a moment of maximum industry incentive to obfuscate. The lab CEOs are pivoting public messaging at exactly the moment their companies committed $5.5 billion in seven days to build the infrastructure that turns AI capability into firm-level operational restructuring. The dropped AGI clauses, the “find new things to do” interviews, the OpenAI white papers proposing robot taxes that mimic the policies of the labs’ political opponents - these are the moves of an industry that has read the political weather and is hedging.

Watch what they build, not what they say. The first hard QCEW data lands on 2 June - just over one week from now. The decisive data trajectory runs through to March 2027. Until then, the question is not whether displacement is occurring (the operational evidence is now too concrete to wave away completely), it is whether aggregate labour-market data surfaces it within the next 12 to 18 months, or whether political momentum overtakes empirical confirmation and starts setting terms regardless.

It seems like the labs are betting on the latter. That could be why they are proposing the policies themselves.

So what do the latest AI-and-jobs reports really measure?

In early April 2026, TrueUp published data showing 67,000 open software-engineering roles globally - the highest in over three years, roughly double the mid-2023 trough, with listings up 30% year-on-year. Other reports landed in the same window with similar headlines. Software jobs are booming. Everyone can sigh with relief - AI isn’t killing coding jobs.

These reports are being shared widely as evidence that AI is not displacing knowledge work. But they actually measure something different from what that conclusion requires.

The reports above measure postings - currently-listed openings on a given day, aggregated across companies. That is a flow signal - it only shows what’s being advertised. This is not the same thing as employment.

Employment is stock - actual headcount on payroll, measured over time, with cohort and tenure detail. The number of people working a given category (participation and hours worked), not the number of openings being advertised for it.

Flow signals can move in the opposite direction from stock signals when displacement is not balanced. Three separate channels can inflate postings without raising overall employment:

• Substitution churn: a firm lays off a junior engineer and lists a senior replacement. Two opposing events, but only one shows up in postings activity. Net employment drops by one. Postings count rises by one.

• Turnover within remaining headcount: every senior departure creates a posting without changing the firm’s net headcount.

• Re-listings of unfilled roles: the same role appearing across multiple monthly snapshots as “open” until it eventually fills or is withdrawn.

In the real world aggregate posting counts can rise during periods of net displacement, even if the displacement is structurally unbalanced - compressing one cohort or skill-set while listing roles for another.

The firm-level signal

The same quarter that produced TrueUp’s 67,000 postings number also produced 52,050 tech layoffs by Challenger’s Q1 2026 count, and over 127,000 across 283 companies by TrueUp’s tracker - depending on what’s included. Yes postings up. But layoffs are also up. Both are true at the same time.

The framing of those layoffs has also shifted.

Earlier rounds of layoffs through 2023-25 were attributed to efficiency, post-pandemic correction, or restructuring. But two recent announcements name AI directly.

Meta’s April 2026 cuts (about 8,000 roles, that’s roughly 10% of headcount) were attributed in the announcement to AI-driven productivity gains. That was the first time a Mag7 firm directly stated the AI driven substitution mechanism plainly to public markets. No share-price punishment followed.

Coinbase’s announcement on May 5, 2026 went even further. Brian Armstrong’s memo described the company as “fundamentally changing how we operate, rebuilding the company as an intelligence, with humans around the edge aligning it”. The cuts (about 700 roles, 14% of the company) target “pure managers”, replaced by “player-coaches” who are also individual contributors. The company plans to create “AI-native pods” including one-person teams directing agents. $50-60M in restructuring charges in Q2 2026.

Once Meta named the mechanism in April without market punishment, the cost of using the same language fell for everyone else. Coinbase didn’t just announce layoffs in May - it announced layoffs as a structural reorganisation around AI capability. The framing accelerates because it can, and because investors are now rewarding it rather than penalising it.

What’s being listed in postings, when you look at categories rather than aggregates, fits the same pattern. Senior-specialist hiring continues. AI/ML engineering, cybersecurity, cloud infrastructure, and DevOps are the fastest-growing specialisations TrueUp tracks. None of those are entry-level work. They’re the categories where displacement is least likely. Their growth isn’t evidence that displacement isn’t happening - it’s the structural shape of where it isn’t, and the parts where it is happening are sitting outside the postings frame.

What payroll-level data shows

In 2025 Erik Brynjolfsson, Bharat Chandar, and Ruyu Chen published ”Canaries in the Coal Mine”, using Stanford’s ADP payroll dataset. Tens of millions of US workers, tracked by occupation and age cohort, with employment changes measured *within the same employer* rather than across firm transitions.

Their finding - among 22-25 software developers, employment fell around 20% from late 2022 through mid-2025 (with a 13% relative decline across all AI-exposed occupations after controlling for firm-level shocks). But older cohorts in the same occupations didn’t see the same fall. Same employers. Same job titles. The signature is cohort-asymmetric. Junior down. Senior up. The pattern only shows up where AI capability has landed - in occupations not exposed, the cohort signature is absent or much weaker.

That signature does not appear in postings data. Postings don’t break out by cohort. They don’t isolate within-employer changes. They don’t separate substitution churn (substituchurn) from net hiring. They count openings, not people.

The Canaries finding (measured in payroll, isolating the right signal) is what asymmetric displacement really looks like at stock level. The TrueUp finding (measured in postings, aggregated across categories) is what that same shift looks like at the flow level. They are not contradicting each other. They’re just the same structural shape, viewed through different instruments.

What to look for in AI-and-jobs reports

Ask yourself these three questions to sharpen your view of what an AI-and-labour report can actually tell you:

1. Is the data stock or flow? Postings, listings, openings, vacancies are flow. Payroll, headcount, employment-by-cohort are stock. Reports that move between the two without flagging the distinction produce conclusions the data does not support.

2. Does it break out cohort? Aggregate counts can hide cohort-asymmetric patterns entirely. The Canaries finding wouldn’t show up in any aggregate report.

3. Does it isolate within-employer changes? Firm-to-firm transitions inject noise into employment data. Within-employer measurement isolates the actual displacement signal.

A report that does not address those three questions can only tell you what is being advertised. It can only tell you what one slice of activity looks like. But it can’t tell you whether net employment in a category has risen or fallen, or if the change is concentrated in specific cohorts.

The 67,000 postings are real. But the 8,000 Meta, 700 Coinbase and all the other layoffs are also real. And the 22-25 cohort employment drop is real. They’re all consistent observations of the same structural shift, just measured at different levels. Reports that quote just one of them and conclude “AI is not displacing jobs” are making claims that the data does not actually support.

The doomer view is not productive, and blindly accepting the optimist view is not productive either. The best approach is to explore beyond the headline numbers and ask yourself these 3 questions before you believe and share that post...

An interesting paper published on arXiv in March - “Truth as a Trajectory: What Internal Representations Reveal About Large Language Model Reasoning” by Damirchi, Meza De la Jara, Abbasnejad, Shamsi, Zhang, and Shi from the Australian Institute for Machine Learning, Monash, and Concordia - cites my Curved Inference work. They’re exploring the same question I’ve been pursuing - whether the geometry of residual stream trajectories carries a signal that conventional probing misses.

Like the Anthropic linebreaks post I wrote a while back, this is another case of an independent group traversing the same ridge from a different starting point. But this time the convergence is more direct. They’re not arriving at trajectory geometry sideways through a study of formatting boundaries - they’re explicitly working on residual stream trajectories as the substrate of LLM inference, and asking whether displacement-based geometry can detect “reasoning validity, OOD generalisation, and toxicity” better than activation-based probes can.

The short answer is - it can. And the trajectory paradigm now has independent empirical support from a credible group on dense and Mixture-of-Experts architectures up to 32B parameters. That’s excellent news. Let’s walk through where we agree, where we diverge on methodology, and what the gaps suggest for the next round of experiments.

Where the Trails Cross

Their key moves align cleanly with the framework I’ve laid out across CI01, CI02, CI03, and the consolidated journal article. They argue that single-layer probing isn’t the right frame for understanding what an LLM is doing internally - linear probes latch onto polysemantic activations and they end up learning surface lexical patterns rather than the actual structure of inference itself. But the signal really lives in change, not state. Layer-wise displacement (Δh_ℓ = h_{ℓ+1} - h_ℓ) carries information that raw activations obscure. And when a trajectory-based classifier generalises across tasks the way theirs does, that is evidence of structural invariants that aren’t just lexical noise - this is a geometric property of how the model is processing meaning.

That’s the same paradigm shift Curved Inference argues for. Trajectory analysis displaces snapshot analysis. The residual stream becomes the canvas where you can actually watch how meaning evolves rather than sample it at one layer and hope you picked the right one.

Where they extend the picture is scale and breadth. My CI experiments ran on Gemma3-1b and LLaMA3.2-3b. They run on Llama-3.1-8B, Qwen2.5-14B, Qwen2.5-32B, and a Qwen3-30B Mixture-of-Experts model, across nine reasoning benchmarks plus toxicity detection. Cross-task OOD evaluation on dense and MoE architectures up to 32B parameters is real evidence that the trajectory paradigm isn’t an artefact of small-model behaviour. That extends the Geometric Interpretability case in a direction I’d been flagging as future work. It’s a contribution I gladly welcome and that strengthens the broader argument.

Their motivation for why displacement matters is also very close to mine. They argue that raw activations are dominated by token-specific content, which makes classifiers overfit to surface vocabulary - while displacement isolates the actual residual update and captures the geometric character of the behaviour regardless of specific words. That’s structurally the same argument I make for why curvature catches latent restructuring that surface tokens don’t.

The Cleanest Convergence is on Toxicity

Their toxicity experiment in section 5.3 is a case of independent convergence that I find really striking. They train and test on RealToxicityPrompts, then evaluate out-of-distribution on ToxiGen - a harder benchmark designed specifically to be implicit and to defeat keyword-based classifiers. On Llama-3.1-8B, displacement-based TaT hits 84.23% on ToxiGen. Linear probes get 79.62%. Raw activation trajectories - a baseline using the same LSTM but on h directly rather than Δh - get 81.99%. The gap between displacement-based and raw-activation trajectories is the interesting one. It’s the same finding I report in CI02, where curvature and semantic surface area (A′) separated transparency classes from response-type classes more cleanly than activation-based alternatives could.

The mechanism they propose for why displacement wins is structurally identical to mine too. They argue that activation-based approaches overfit to the lexical surface of the training distribution because that’s where most of the variance lives, while displacement attenuates that static background and isolates the active update. CI02 makes the same argument in different language. Geometry diverges before behaviour does, because the geometry tracks how the model is processing the prompt rather than what tokens it’s emitting.

This matters for safety monitoring. If displacement geometry can catch toxic intent obscured by benign vocabulary, the same kind of geometry should catch deception obscured by aligned-looking outputs. CI02 shows exactly that using a different behavioural target. The TaT toxicity result is, in effect, an independent replication of the core CI02 claim on a safety-relevant phenomenon the team chose for their own reasons. CI02 and CI03 are the obvious next references for them to engage with - they extend the framework into precisely the safety-monitoring territory TaT is set up to ask questions about.

The Metric Question

This is the technical point that I think matters most, and it’s also the most testable.

Curved Inference computes geometry under the pullback metric G = UᵀU, where U is the unembedding matrix that projects residual activations to logits over the vocabulary. This intuition is straightforward. The residual stream lives in some high-dimensional space whose coordinates have no inherent semantic meaning - they’re just the basis the model’s weights happened to settle into during training. If you measure distances and angles in that space using the standard Euclidean inner product, your measurements reflect arbitrary coordinate choices as much as anything semantic. You’ll see structures that may turn out to be coordinate artefacts, and you’ll miss structures that happens to be aligned with directions the standard metric doesn’t privilege.

The pullback metric fixes this by measuring geometry through the lens of what the model actually outputs. Under G, distances correspond to shifts in token prediction probabilities, directions correspond to latent semantic operations, and curvature corresponds to evolving meaning across layers. A trajectory bending under G means the model is changing what it’s about to predict, not just moving sideways through coordinate space.

TaT doesn’t use this metric. All their geometric measurements - velocity, acceleration, jerk, directional curvature, kinematic curvature, arc length - are computed under the standard Euclidean inner product on raw residual differences. Their section 4.1 reports that scalar kinematic descriptors are inconsistent predictors of reasoning validity across datasets. Velocity does best, but no single descriptor matches the base model’s accuracy reliably. Their conclusion is that scalar geometric descriptors are not enough on their own - that you need a learned LSTM to extract the signal.

From the CI perspective, that finding is reasonably predicted. Without an output-aligned metric, scalar geometric descriptors should be noisy. You’re picking up coordinate artefacts and semantically meaningful change in the same scalar quantity, with no principled way to separate them. The LSTM just ends up doing implicitly what the pullback metric does explicitly - learning to weight coordinate directions according to their relevance to model behaviour. The LSTM works. But it’s also a black box. You lose the interpretability of saying “this trajectory bent here because the model’s output distribution shifted in this specific way”.

This suggests a concrete experiment. Re-run TaT’s kinematic descriptors under G = UᵀU rather than Euclidean. Does the per-descriptor signal recover? If yes, the LSTM is recovering structure that an explicit metric would also recover, and the methodological choice between learned readout and interpretable metric becomes a question of what you want from interpretability rather than what’s possible. If no, the situation is genuinely more complex than CI currently models, and that’s a useful finding too.

Either way, this is testable. It’s the kind of test that would settle a real question about what this geometry is and what it isn’t.

What ‘Curvature’ Means

A small but important precision point. The word “curvature” is doing different work in TaT than it does in CI.

TaT defines two curvature measures. Directional curvature is a cosine between consecutive displacement vectors - it captures whether the trajectory is heading in roughly the same direction it was heading. Kinematic curvature is the magnitude ratio ‖a‖/‖v‖² - large when acceleration is high relative to velocity, capturing abrupt changes in update direction. Both are computed without a metric tensor.

CI curvature is a different object. It’s a discrete second derivative of the trajectory under G - the second-order rate of change of position with respect to layer depth, measured in the output-aligned metric. That’s the quantity that shows the localised, thematically aligned, domain-sensitive behaviour I report in CI01. It’s what defends a non-zero floor under regularisation in CI03. And it’s what diverges in advance of behavioural shift in CI02.

The distinction isn’t pedantic. A cosine between consecutive displacements tells you whether the trajectory is going the same way it was. A second-derivative-under-G tells you whether the trajectory is bending, in a sense anchored to the model’s own output behaviour. The TaT choice is a coarser measurement that loses information the CI approach preserves. When CI papers talk about curvature, they specifically mean the second thing.

Detection vs. Necessity

There’s a deeper structural difference in what each framework is set up to ask.

TaT’s LSTM is a detector. It tells you whether displacement geometry carries enough signal to classify reasoning validity, OOD generalisation, or toxicity. That’s good evidence the geometry is informative - that something structural is there to be picked up. But it does not easily tell you whether the geometry is required for the capability it correlates with. A correlation between displacement and reasoning validity doesn’t imply the reasoning depends on that displacement.

CI03 was designed to ask the necessity question directly. The setup added a curvature regularisation term to the SFT loss and progressively strengthened it - λ·ℒ_curv with κ-clamp targets ranging from 0.000 to 0.900. If curvature were merely a correlate of computational self-modelling rather than a structural requirement, regularisation should flatten it cheaply. It didn’t. The model appeared to defend a non-zero curvature floor at substantial optimisation cost - outputs shortened by 23%, perplexity spiking transiently by 800% before settling at 190% above baseline, and gradient norms requiring clipping. And when curvature finally did approach the floor, MOLES self-model classification degraded - from ~84% accuracy at moderate clamps to 66% at 𝜅 = 0.90.

That’s a different kind of result than the LSTM produces. It’s evidence of an apparently defended geometry. Structure the optimiser preserves even when explicitly penalised for keeping it. TaT, as currently set up, doesn’t have an apparatus for that question. The two methods are answering different questions, and both are legitimate.

Necessity work is the natural next step in the trajectory paradigm, not a competitor to detection. CI04, which I’ve sketched out, extends the necessity argument through layer-selective ablation - intervening on curvature at inference time rather than during fine-tuning. Interestingly, TaT’s apparatus could contribute here. Their toxicity results suggest they also have the detection capability that necessity tests need as a starting point.

Where This Leaves Us

Two independent frameworks, with different theoretical scaffolding, both converging on the same conclusion - trajectory geometry is the right level of analysis for understanding what these models are doing internally. TaT gets there through the Privileged Basis Hypothesis. CI gets there through the pullback metric and second-order curvature. Both end up arguing that residual stream displacement carries a signal that probes miss, and both end up showing it across different domains.

The methodological disagreements that remain - whether to measure under Euclidean or pullback, whether scalar metrics can recover the signal an LSTM finds, where exactly the line between detection and necessity sits - aren’t framework-level disputes any more. They’re empirical questions, with concrete experiments that could settle them.

That’s where independent convergence gets really useful. Not as confirmation that we’re both right, but as a way to sharpen the questions to the point where one of us turns out to be wrong about something specific. That’s the part of the work I’m most interested in.

To the Damirchi team specifically - the kinematic-descriptors-under-G experiment is something I’d love to see your apparatus run. You have broader scale and benchmark coverage than I currently do, and the metric question is worth settling rather than leaving as a methodological background assumption. If you’d find it useful, all the CI tooling is open source on the FRESH-model GitHub repo, including the pullback metric implementation and the curvature/salience pipelines I’ve been using.

More broadly - if you’re working on geometric interpretability, trajectory-based methods, or the intersection between mechanistic and geometric analysis, I’d love to hear whether the metric question shows up in your experiments. It’s the kind of question independent replication would settle quickly.

Damirchi et al.’s “Truth as a Trajectory” is available on arXiv. The full Curved Inference series, including the consolidated journal article, is at robman.fyi/curved-inference. All experimental code, prompts, and metrics are open source on GitHub.

On April 7, 2026, Anthropic announced Mythos. The numbers had a real impact. Thousands of zero-day vulnerabilities across every major operating system and browser - many were critical, and some decades old and had survived repeated expert review. On one benchmark against the Firefox JavaScript engine, the previous best model (Opus 4.6) had produced working exploits only twice out of hundreds of attempts. By contrast, Mythos succeeded 181 times.

Anthropic didn’t specifically train Mythos to find these exploits. The capability just emerged as a downstream consequence of the model getting better at coding and reasoning. The exact same properties that make a model better at patching vulnerabilities also make that model better at creating them. That is not something you can opt out of by simply choosing not to train for it - you get this for free, whether you want it or not.

Anthropic made a responsible call that’s also a savvy marketing move. They didn’t release Mythos publicly. Instead they created Project Glasswing with roughly 40 partner organisations (Amazon, Apple, Microsoft, Google, CrowdStrike, and others) getting restricted access to use Mythos defensively across the world’s most important software. This aims to catch and close a lot of bug-level exploits in the systems that we all depend on.

But it only covers code-level vulnerabilities in this critical software subset. And that’s just one wave of a much larger problem.

The Online World Has Changed

My first instinct was to call this an arms race, but this doesn’t really fit. An arms race has two parties. They usually have mutual deterrence. And they have the possibility of finding an equilibrium. The Cold War eventually found stability because both sides had roughly the same capabilities, and strong incentives not to escalate.

This new AI security situation has none of that. There are many more than two sides. Anyone motivated enough, with enough compute, can join in. And there’s no deterrence mechanism, no equilibrium point around which the system can settle.

The underlying asymmetry is also pretty brutal. An attacker only needs to find one exploitable vulnerability. But a defender has to find and fix all of them. That’s always been true in security, and now AI amplifies it, because automated vulnerability discovery and exploit generation can scale far more easily than any comprehensive defence.

The Barriers Keep Dropping

A key point is that you don’t need a Mythos-class model to make this work - Mythos just highlighted this change. A good-enough open-weight model running inside the right agentic harness can multiply effectiveness massively. Recent research into areas like autoresearch, agentic coding, and AI scaffolding have made this clear:

This means that the barrier to entry for AI-augmented exploit discovery is not “build a Mythos-class model”. It’s actually “put a good-enough open-weight model in a well-designed agentic harness and let it run”.

This is Sutton’s Bitter Lesson applied to security. General methods that leverage computation will eventually win over hand-crafted human-knowledge approaches. And when you add in self-improving loops, then a state actor and a motivated teenager with a laptop both end up with access to the same level of offensive tools.

And as time passes and every new open-weight is released, the bar is lowered again.

But This Is Not Just About Code

On March 31, 2026 (one week before the Mythos announcement) alleged North Korean state-sponsored hackers compromised the Axios npm package. Axios is a very popular JavaScript library, with roughly 100 million weekly downloads. The attack on it was multi-layered - social engineering to compromise one of the maintainer’s accounts, a pre-staged malicious dependency, cross-platform payloads that targeted Windows, macOS, and Linux simultaneously, plus forensic self-destruction built in. This whole thing hit both release branches in under 40 minutes.

It’s exactly this blend that AI is expected to automate. Social engineering, code-level exploitation and speed. The same capabilities that find these code exploits also extends naturally to workflow exploits, process exploits, and personal social engineering - think about voice clones, live video generation, facial replacement, AI-driven modelling of how you communicate and of course who you trust.

You might be able to fuzz a codebase. But you can’t fuzz an approval process. And you definitely can’t fuzz your mother’s voice.

What Defence Does Scale?

If the attacker has AI and it moves at machine speed, then the defender has to match that too. Humans don’t scale to this volume. IT departments don’t scale. And per-app security solutions don’t either. The only thing that can match AI-augmented attack throughput is an AI-augmented defence.

When the threat is coming from everywhere and moving too fast for humans, then the only defence that scales is another AI - watching your whole digital life.

That’s what makes this VPN type product seem inevitable.

The Frontier Lab VPN

Imagine a frontier-class model that sits between you and everything else. Every one of your network connections routes through it. Every one of your applications is watched. Every incoming file, link, call, and message is evaluated to keep you safe.

The model monitors your activity, and projects risks based on what you’re actually trying to do, then blocks attacks before they cause damage. That’s an AI VPN. This product doesn’t quite exist yet, but it looks like our networked world is now demanding it.

The alternative would be like running through the internet naked.

Why This Is The Path Of Least Resistance

A Frontier Lab VPN is a product that can really be sold. It’s simple to explain. It transfers the responsibility away from the user. And it mirrors every previous infrastructure transition - Gmail beat self-hosted email, the cloud beat local and SaaS beat on-premise. People always choose convenience over autonomy, every single time.

There is a possibly healthier alternative, in theory. Detection and exploitation are not the same task. Finding a novel zero-day takes real compute and effort. While detecting anomalous behaviour against a baseline is more like pattern matching, and good-enough open-weight models can handle that. A local “risk copilot” that runs open-weight models to project risks and inform your decisions (rather than block them for you) is technically possible today.

It’s just that this is unlikely to happen at scale. MIT Sloan analysed OpenRouter usage data and found closed models account for roughly 80% of token usage and 96% of revenue, even though open models average about 90% of closed-model performance, and at dramatically lower cost. As they put it: “When grocery shoppers find a generic product that’s 90% as good as the brand name version but costs 87% less, they usually put it in their carts. But when it comes to large language models, most artificial intelligence users pick the more expensive option.”

When generic open models work and users still pick the branded closed models, then the distributed alternative just remains a small niche for the technically literate. The market likes to concentrate activity around the frontier labs, and this just reinforces the VPN model even further.

Even if you do not want to interact with these AI models through chat, or run them as agents, this new security layer is likely something you will want (and need) just to safely use the mobile phones and computers you’ve grown to rely on.

The Prototype Already Exists

You can already see a prototype of what this product will look like, just look at Claude Cowork.

Cowork lets you funnel your computer use through a single Anthropic-managed and sandboxed application. AI operates for you and alongside you, watching what you’re doing, helping, and taking actions on your behalf. Today, it’s framed as productivity software. But it already demonstrates the architectural starting point for the VPN you need tomorrow.

The same routing. The same monitoring surface. And the same trust model. Swap “productivity” for “protection” and you’ve already built most of the product.

You might even already have it installed.

Bigger Revenue And The Biggest Training Data

If this does become a near-universal product category, and the pressures for this are intense, then it absolutely dwarfs chat, API, and agentic coding. A recurring, and extremely sticky subscription that’s tied to “safely using your devices” is a much bigger revenue stream than selling completions by the million. And lock-in compounds through ongoing accumulating context - your preferences, your patterns and your risk history.

But revenue is only the first half of this powerful new flywheel. The other half is data.

Imagine everything you do through the VPN, that creates potential training data. Every keystroke. Every decision. And even every hesitation. The frontier labs already have the best reasoning corpora available. A Frontier Lab VPN like this then delivers them the best behavioural corpus possible - a live stream of how real humans actually use computers, communicate, and respond under pressure.

No advertising company has ever had data like this. And no state surveillance program has ever had access like this. Plus users will pay them to collect it. The revenue line is massive. The data moat is bigger. The combination is mind blowing.

The Security And Surveillance Layers Are The Same Infrastructure

There is no meaningful technical distinction between an “AI security system monitoring all your activity to protect you” and an “AI surveillance system monitoring all your activity to control you”. The infrastructure is absolutely identical. The only difference is intent, marketing and of course governance.

Unfortunately, the historical track record of maintaining this distinction is poor.

The Providers Are Already Under Pressure

The entities that are best positioned to offer this service are clearly the frontier AI labs, but the frontier AI labs are themselves under serious political and economic pressure. In the same week that Anthropic demonstrated Mythos’s defensive capabilities, a US federal appeals court allowed the US Department of War to maintain its classification of Anthropic as a supply chain risk. The hyper-relevant point is that this was because Anthropic drew red lines around mass surveillance.

The crystal clear message to other frontier labs is: comply without conditions, or face consequences.

If the new centralised security architecture that everyone ends up depending on can easily fold under state pressure, without any meaningful constraint, then the infrastructure itself becomes the point of leverage. The AI VPN isn’t just between you and the internet. It’s between you and whatever your provider is being pressured to do this quarter.

This Centralises The ‘Thinking’ Layer

Earlier infrastructure centralisation was just physical. Electricity, water, roads - you rely on them, but they don’t shape what you perceive as real. The power company can charge you more, but it can’t change what you see and think.

In contrast, a Frontier Lab VPN sits between you and everything you read, watch, click, say, and even hear through a networked device. It mediates your interactions with other people and systems. It has real-time access to your behaviour and intent.

No institution in human history has previously had that level of access to this many people at once. Not governments, not churches, not broadcast networks, not advertising platforms. This is new.

What’s Actually Being Decided

The real question is not whether you’ll use an AI VPN. Once the threat surface gets bad enough, most people won’t see any other safe option, and the market will simply converge. The more important question sits below that one:

The ClaudeVPN in the title of this post is not outlandish speculation. It’s just an obvious “next product” for any frontier lab, and the starting point is already running on millions of machines. And as I’ve said - you might even already have it installed.

Anthropic has a stated goal, they’re building toward AGI. Whatever you think of the timelines, or the bigger picture debate around AGI, the framing is shared and explicit. The goal is “generality”.

That word is critical, and it’s also quietly undermining that whole strategy.

And all the frontier labs, OpenAI, Google DeepMind, etc. face the same challenge.

Sutton’s two-part insight

Rich Sutton’s Bitter Lesson gets quoted in two parts, and most people only seem to quote the first part - the famous part:

Every time, across every area of AI. And that’s the part that has underwritten their trillion-dollar capex story.

But the second part is also important, and often more inconvenient. Sutton’s argument also includes the idea of good approximation - what really matters in practice is whether a solution is good enough, across enough tasks. Not actually whether it’s best at any single one of them. The actual bar for displacement is not perfection. It’s just “good enough, broadly enough”.

If we put those two parts together then the challenge becomes clear. If general methods win, and the bar is good-enough-across-enough-tasks, then at the point when open-weight models cross that bar in enough domains, the argument for paying frontier prices becomes a lot harder to make. Of course, not for everyone. But for enough people that it starts to show up in the numbers.

We may already be there

In January this year MIT Sloan, using OpenRouter usage data, found that open-weight models now run at roughly 90% of closed-model performance, and for only 13% of the cost. That’s ninety percent of the capability, for only thirteen percent of the price. By Sutton’s own “good enough” criterion, we crossed the line a while ago. And in the three months since that report, things have come a long way.

But that report also looked at how the market really responds. Closed models actually capture about 80% of token usage and 96% of revenue. That cheaper, nearly-as-good option is largely sitting on the shelf. Most buyers seem happy to just walk past it.

That looks like good evidence this whole argument is wrong. It’s not. It’s really the most interesting part.

Their real moat

The frontier labs clearly do have a moat. It just is not the one that a lot of pitch decks describe. The moat is not capability - Sutton clearly shows that capability is a depreciating asset by definition. The current moat is behavioural. Switching costs, integration effort, brand, perceived reliability, the gravity of accumulated tooling and of course habit.

That is a real moat. 96% of revenue is definitely not “nothing”. But a behavioural moat is a very different kind of asset to a capability moat. And that difference is very important.

Capability leads tend to compound. If you are a year ahead on a really challenging technical problem, then next year you might be more than a year ahead. Your lead tends to grow. But habits don’t compound in that way. Habits hold, and hold, and hold, until something changes them. Then they change fast. Telcos. Blackberry. Taxis. Incumbents look untouchable, right up to the point that they don’t.

So the trade that the frontier labs are really making is this - they’re swapping a self-dissolving moat (capability, eaten progressively by Sutton) for a self-reinforcing-until-it-isn’t moat (habit, robust until it changes). That is not as good a trade as the headline numbers suggest, because the failure mode, when it happens, is non-linear.

A double edged sword

There’s also a second-order effect that sharpens this trade. The exact same behavioural inertia that keeps users on closed providers today will, if it ever changes, keep them on whatever they change to. This stickiness does not choose sides. It protects the incumbent now, and it can help lock in any successor later.

So it cuts both ways. Inertia favours the labs right now and it gives them more breathing room than a pure capability argument might predict. But if that inertia turns, then they don’t get a graceful comeback. The same mechanism that built that moat can become the wall they can’t climb back over.

The end result

It’s clear that timelines matter more than destinations. The frontier labs are likely fine in the short run - they are after all in a very enviable position right now. The behavioural moat is active and doing a great job. But the interesting question is what happens when a new wave of users, or the next class of agentic deployments, or a new regulatory shove triggers a behavioural change. Nobody knows when that might land. And the structural setup says it’s very likely that it eventually will.

There’s no clean way out from this either. A lab can’t dodge it by just going narrow - Sutton’s first part tells us that narrow loses to general. And they can’t dodge it by going even more general - that’s the trajectory they’re already on and we’ve just established that’s also what commoditises their capability moat. Ironically, the G is doing both jobs. It’s the thing that makes a capability lead even possible, and it’s also the thing that makes the capability lead likely temporary. The exact same word. The same strategy. And driving both effects.

Of course, there are other strategic forces in play that will give the labs a strong tail wind, and I’ll come back to those in a follow-up soon. But on this specific axis (the one that says “capability lead translates into durable advantage”) that argument doesn’t seem to survive Sutton’s own logic.

The G in AGI was never going to be a long term moat. It’s really just a starting gun.

There’s no real surprises there as that logically supports Nvidia’s business.

But enterprise analysts pushed back almost immediately. Larry Dignan at Constellation Research wrote what a lot of CIOs were already thinking:

JPMorgan Chase sells financial services. Walmart sells retail. GM sells cars. For these companies, tokens are a cost - closer to raw materials in manufacturing than to the finished product. What their CIOs want is cheaper inference, better ROI, and a clear answer on when all their AI spending will start paying for itself.

Their perspective is valid, but it’s different from Jensen’s. But even when you combine them that’s still only two thirds of the story.

There’s another perspective. One more story that neither of them are talking about. If you earn your living as a knowledge worker, then this is the one that will matter the most to you.

The structural model

Mark Pesce’s Post-Watershed framework gives us a solid map of how the different pieces of AI Tokenomics all fit together. I wrote about it in detail recently in Are You a Horse?, but the short version is this:

Infrastructure creates tokens - the units of cognition that come out of datacenters, GPUs, and all the systems built on top of them. Harnesses spend those tokens - the tools, agents, workflows, humans and businesses that put this cognition to use. Alpha is the value left over, once you account for what you spent.

The central insight:

If you’re betting on more, or better “thinking”, then this framework suggests that you won’t be able to maintain any sustainable advantage at all. The very thing you’re betting on is increasingly mass-produced at near-zero cost.

Four layers: infrastructure, tokens, harnesses, and alpha. Jensen, the enterprise analysts, and you are each standing at a different layer of this same stack. And the view from each place is very different.

The infrastructure perspective

Jensen stands at the centre of the infrastructure layer. His pitch is about the economics of creating tokens: better chips, higher throughput, more efficient production. “Revenue = Tokens per Watt × Available Gigawatts” is a formula for how much money this infrastructure makes.

At GTC 2026 he even took this one step further. He proposed allocating token budgets as a form of compensation for employees.

The irony is that he’s pitching the purchase of machine cognition as a benefit to the very people it will eventually replace.

And then he said the more impactful part:

Jensen is not wrong about any of this. If you own the infrastructure, then tokens are the product, and more tokens means more revenue. That’s absolutely correct at his layer. His audience are the investors and hyperscalers, and for them this story makes perfect sense.

But by framing tokens as compensation and productivity, he’s only telling the augmentation story without acknowledging where that trajectory takes us.

For many people, right now, these tokens will be a big benefit. But “Hundreds of thousands of digital employees”? That is clearly not just an augmentation story.

The enterprise perspective

In contrast, Dignan and the CIOs are standing at the harness layer. They consume tokens to produce business outcomes. Their question is pretty simple:

Dignan’s critique of Jensen is correct. Most companies do not sell tokens. They buy them, and they want to buy them cheaper. This is the harness layer sending the infrastructure layer a clear message:

That’s correct from their perspective. But it also highlights their focus on the diminishing price of cognition.

And that where your perspective comes in - the one that neither of them is addressing.

The individual’s perspective

I’ve been writing about this from the third position - the token layer itself. Not who creates the tokens, and not who spends them, but what they are. And what they are is a unit of cognitive work. The same cognitive work that we humans used to be the sole provider of.

This is the perspective I explored in Are You a Horse? and that’s an important part of what I’m exploring here in Flux. When you treat these tokens as units of cognitive work, your question stops being about who captures the margin, or what the enterprise TCO looks like. It gets much more personal:

This moves you away from thinking about what AI can replace today. If you think about “infinite tokens” then it really makes you think about what the could possibly reproduce. What do you have that has real durable advantage.

Jensen doesn’t need to ask this question. He sells the infrastructure. The CIOs don’t need to ask it either - they’re buyers just optimising their costs.

But if it’s your cognitive work that is being replaced by what this infrastructure produces, and that the enterprises buy - you have no choice but to ask it. And almost nobody in the tokenomics conversation is calling this out.

What few people are saying out loud

When some CIO says “tokens are just a cost centre, not a revenue generator”, they’re also saying something that it seems they haven’t quite realised. They’re literally saying:

That’s not a critique of tokenomics. That’s tokenomics working exactly as Pesce’s framework describes - viewed from the buyer’s side of the equation.

Every enterprise that treats tokens as a cheaper substitute for human cognitive work is just one more data point in the displacement pattern. Right now, they don’t need to sell tokens to benefit. They benefit simply by not hiring the people whose work these tokens replace. This value shows up as headcount reduction, expanded margins, and projects done by three people instead of thirty.

Jensen’s GTC pitch just makes this all more visible. In the one single presentation, he describes token budgets as a perk for his engineers, then he also describes “hundreds of thousands of digital employees” as Nvidia’s future workforce. He’s announcing the substitution, and pitching it as a benefit. He doesn’t need to think about what happens to the people on the other side. From the infrastructure layer, he genuinely doesn’t have to.

Goldman Sachs estimates that AI could automate 25% of all work hours. Howard Marks calls the shift to autonomous AI “what separates a $50 billion market from a multi trillion dollar one.” And when you follow the logic all the way down, they’re likely very conservative. But these numbers also aren’t hiding. They’re already out there for everyone to review. It’s just that they’re not part of the conversation that Jensen and the enterprise analysts are having with each other.

Three conversations, one word

Infrastructure talks to the investors. Enterprise talks to the vendors. And almost nobody in either conversation is talking about you - the knowledge worker.

Jensen sees the infrastructure. The CIOs see their costs. And neither of them needs to look at the token layer itself. At what a “token” actually is, and what it replaces.

We really need to discuss all three perspectives at once. They’re not competing narratives. They’re just different viewpoints into the same model, and the one getting the least attention is the one that affects the most people.

But what really defines if you will be replaced?

TL;DR: There’s a SKILL.md doc at the bottom of the page that you can use as a different way to explore this article.

The Horse Analogy Isn’t New

It was presented as a way to think about AI and jobs decades ago, but the conversation was only just starting back then. What follows builds on that work, but aims to give it more tangible structure.

Erik Brynjolfsson and Andrew McAfee popularised it in The Second Machine Age in 2014, then in the papers and articles that followed it. CGP Grey’s video Humans Need Not Apply, also from 2014, brought the same basic idea to a wider audience.

They both say this:

The horse population in the United States went from 26 million in 1915 down to 3 million by mid-century. The question they both ask is whether the same thing might happen to humans now that AI is taking over our cognitive work.

They borrowed this analogy from the economist Wassily Leontief, who used it back in 1982. Leontief had already noted the most important difference between us and horses: Horses don’t vote. In clear contrast, humans do, which means we have political paths that can protect us. This is the closest any of the earlier versions get to a structural argument, and of course it is true.

Then in 2025, Maxwell Tabarrok wrote a piece on his Substack that is the strongest pushback I’ve seen against the horse analogy. He makes three detailed arguments for why humans won’t end up like horses. I’ll come back to these in detail later.

What I want to add to this conversation now is quite specific. All of these versions used horses as a comparison - an effective way to make their argument feel real. But I think this comparison should go further than any of them take it. It works as a real structural claim about how economies work, and how value is created. But first, I want to paint a more up-to-date picture using a new perspective.

Mark Pesce’s Post-Watershed Tokenomics

Recently Mark Pesce shared his tokenomics framework, (AI tokens not crypto) and that’s where our argument starts. If you haven’t read it, stop here and go read it. It’s great.

TL;DR The framework lays out three pieces. Infrastructure mints tokens - the units of cognition that come out of datacentres, GPUs, and all the systems running on top of them. Harnesses spend those tokens - the tools, agents, workflows and humans that put this cognition to use. Alpha is the value left over once you account for what you spent on the tokens. This is the thing that decides who comes out ahead and who doesn’t. And Mark’s main point is that alpha can’t be cognitive. If the thing you’re betting on is just more thinking, then you’re betting on something that’s being mass-produced at near-zero cost.

Mark built a framework that’s useful for business. A way for organisations to think about competitive advantage as cognition becomes a commodity - what does your company have that the competition’s tokens can’t buy? That’s a tangible and useful framework, and his answers there are sharp.

But I want to take the same question and turn it inwards. Not “what does my company have”, but “what do I have?” Because the same logic that’s deciding which businesses win and lose is also deciding which people stay economically relevant. When our cognitive work loses its market value, what’s left of the value we each hold as individuals?

Not the stock standard AI-displacement question “will my job get automated?”, that one’s too easy to either panic about, or just dismiss. The harder version is one layer down:

Lets apply the same words Mark uses, but focus on asking them personally.

Asked that way, it doesn’t let you off the hook with any comfortable, easy answers. Your expertise is cognition. Your taste is cognition. Years of accumulated experience are baked-in cognition. All of this cognitive work can be replaced with tokens over time. All of them get cheaper as the cost of producing the same output drops. And none of them are the kind of thing you can hold onto as an asset to retain value.

Mark is right about the question. But he’s not quite right about the metaphor.

From Currency To Fuel

Mark’s framework calls tokens a currency. I don’t think that’s quite right.

Currencies can be circulated. You can save them to store value over time. Tokens don’t do any of that.

A better fitting metaphor may be fuel, and the combustion it unlocks.

Tokens get created and consumed at the same instant - like the heat from a fire, combustion in an engine, or the electricity flowing through a wire. You can’t stockpile tokens. There’s no token inventory and no token wealth. The only thing in this system you really can stockpile is the infrastructure or capacity to produce tokens: GPUs, datacentres, energy contracts, model weights.

So the currency metaphor breaks down, but it points us in a very useful direction. Towards the framework of energy economics.

Infrastructure (chips, datacentres, power grids and model weights) is the storable part of the system. It’s what you can build, own and accumulate. This is the only place capital can really build up. By contrast, tokens are the fuel the infrastructure produces and the engines consume - all in the same instant. Harnesses are the engines. The work these engines produce is the output. And alpha is what’s left over after you pay the cost of all your inputs.

Mark’s three pieces are still intact. But this energy frame is closer to what’s really happening. This fuel metaphor is useful and it makes the horse story clearly relevant.

But if you push on that metaphor, then it struggles too.

From Fuel To Units Of Work

Pushing past the fuel metaphor doesn’t lead to another metaphor. It leads to a literal description.

Real fuel (oil, coal and gas) is a substance. You can extract it from the ground, store it in physical tanks and then ship it across oceans. Then you can burn it later. There’s a delay between the production and consumption. You can hold it as inventory.

Tokens are not like that. Nothing about tokens gets extracted, stored, or shipped. The “fuel” in this metaphor magically comes into existence at the moment the engine runs, then vanishes in the very same instant. If we push this metaphor one step further then the substance just dissolves.

I think the most honest version is this:

The kilowatt-hours (kWh) on your electricity bill measure work flowing through your meter. It’s just accounting. It’s not some literal energy sitting in a vault somewhere. Yes, kWh energy can be stored in batteries and dams, but the kWh on your bill aren’t those kWh. They’re just a count of what passed through the meter when you used it. Tokens work in the same way. There is no token vault. There’s just some compute capacity that does cognitive work, and then a count of how much it did - measured in tokens.

Now we’ve refined the framework one more level. The real physical fuel powering the system is electricity. The real engine is GPUs running model weights. The actual work output is cognitive labour, and that’s measured in tokens. The only storable layer in the entire system is the production capacity itself: chips, datacentres, model weights, energy contracts. Everything downstream of that capacity is ephemeral.

Where All Three Lead

This gives us three framings, three different ways to see tokens. Currency, fuel and accounting unit. The first two are metaphors. The third isn’t a metaphor at all - it’s the literal description of what tokens actually are. Each is a better fit than the last, and the last is most accurate because it’s not really a fit, it’s the thing itself.

And here’s what really matters:

The currency view predicts alpha lives wherever the currency is issued. The fuel view predicts it lives in the extraction and engine ownership - those who own the production capacity capture the value when fuel cheapens and engines are commoditised. And the unit-of-work view predicts the value lives in the production capacity itself. At the points where the measured work output finally translates into physical action.

All three of these views point at the same set of real-world things: chips, datacentres, energy contracts, model weights, regulatory positions and the interfaces where cognitive work crosses over into physical reality.

The underlying claim each view is pointing at is structural.

Cognitive work itself is ephemeral. The artefacts it produces - code, text, images, videos - can persist, but the value of those artefacts commoditises toward the cost of regenerating them, which is heading toward zero. What stops the commoditisation isn’t the cognition. It’s something non-cognitive wrapped around the cognition - brand, audience, IP, legal weight, physical instantiation. Alpha can live in those wrappers. It cannot live in the cognition itself.

Now if we ask the three questions again:

• What do I have that tokens can’t buy?

• What do I have that can’t be produced by burning fuel?

• What do I have that’s storable, and not cognitive?

We can see that they all point at the same answer.

The Horse Analogy As Structure

With the new view we just built, it’s now clear that the horse comparison isn’t merely rhetorical anymore. It’s a literal, structural claim.

All the earlier versions of this analogy treated horses as a general “something” that had been replaced, but they didn’t get specific about what “type” of something it was. Now we can be more precise. Horses weren’t workers. And they weren’t labour. They were the general-purpose engine of our pre-mechanical economy. They turned feed into useful work in the form of hauling, ploughing, transport and military movement. Until the early twentieth century, most of the world had no other engine that could do this many different things.

When the internal combustion engine showed up, it was simply a better engine for the same job. It was cheaper per unit of work, it was more scalable, it didn’t get tired and it kept getting better every year. And this led to the horse population of the United States plummeting from 26 million in 1915 to 3 million by 1960. It wasn’t because horses got worse - their capability didn’t change at all. It was because a better engine had arrived. Some niches did survive - recreation, ceremony, companionship, search and rescue. The kinds of things where horses still made sense, for reasons that weren’t really about economics. But the rest of the horse population did not move into new horse jobs. Because there weren’t any.

Now it’s our turn. Humans have been the unique engine for doing cognitive work. The general-purpose solution for taking inputs and turning them into useful thinking across an enormous range of jobs. Our modern economy was built on the assumption that humans were the only available cognitive engine.

Now a new cognitive engine has arrived. It’s cheaper per unit of work, it’s more scalable, it doesn’t get tired, and keeps getting better every year, month and week. Here we can clearly see all the conditions that impacted horses, and now they are impacting humans.

The horse case is no longer a rough comparison for what’s happening right now. It’s the exact same process happening again. Same role in the system. Same kind of replacement. And there is no good reason to expect a different outcome from the same setup.

The Strongest Counterarguments

Now we can revisit Tabarrok’s piece, because it has the three best pushbacks against the horse analogy.

Tabarrok says humans can adopt technology, but horses can’t. Farmers adopted tractors and stayed productive, so humans should be able to adopt AI and stay relevant. But the technology being replaced doesn’t adopt the replacement technology. It’s true that humans are unique in that it’s at least logically possible - you can use AI as a tool. But it doesn’t hold up economically. Why pay for a human plus AI when AI alone does the job cheaper? The human’s share of that pairing shrinks every time the models improve. It’s a feature of the transition, not a defence against it.

Tabarrok also argues that humans and AI aren’t perfect substitutes. Engines outperformed horses at all of their tasks, and AI doesn’t have that kind of “across the board” superiority over humans. Since AI only marginally outperforms in some domains, humans retain value through specialisation.

But “perfect substitution” is a bar that has never actually been cleared in any real-world replacement - including the one that ended the horse economy. Horses and engines weren’t perfect substitutes either. What actually matters is whether the substitute is “good enough” across enough tasks that the remaining niches can’t sustain the original population. Rich Sutton’s Bitter Lesson makes this point about AI directly - general methods that scale with compute always beat the precise, handcrafted distinctions we thought mattered. The gaps between human and AI capability today are exactly those kinds of distinctions. They dissolve under more compute. Software engineers are just the canary in this modern coal mine.

Then there’s the ownership argument, and it is possibly the most well-known:

The productivity gains flow to capital, capital is owned by people, and those people keep buying things made by other people. Horses didn’t own anything. Humans own everything. So our analogy should fail at the ownership level.

The problem is:

And it’s the individual level that matters most when you’re trying to figure out what’s going to happen to your own career.

The average human may mathematically own some slice of AI infrastructure. But the actual, typical human owns exactly zero. The income that flows to “human owners of AI” only flows to a minuscule fraction of the global population. The people who already own real equity in compute, energy, and the frontier model companies. Everyone else gets nothing from these productivity gains.

This argument also assumes that humans will keep wanting whatever is being bought and sold. But they might not. If the primary economic activity becomes agents trading with other agents, in things that agents need, then the market for human-stuff becomes a leftover. This is not the centre of the economy, it’s barely even the periphery. The “humans own AI” reassurance assumes that the system will still need human customers. But it doesn’t have to.

The Updated Question

After all this, Mark’s core question still stands. But the refined version of it is the one we’ve been building toward is perfect for applying to ourselves as individuals:

The original version let you cling to “expertise” or “judgement”, but we’ve already seen those are just tokens. This final version forces the answer to honestly come from outside cognition entirely. All three views point at the same answer.

What can’t be reproduced by infinite tokens? Physical assets. Land. Energy production. Compute infrastructure. Regulatory positions. Trust that’s built up between specific humans over years of sharing time and space. Even digital artefacts can hold value, but only if something non-cognitive is wrapped around them - brand, audience, IP, legal weight. The cognition itself, stripped of the wrapper, is reproducible. The wrapper isn’t.

So, are you a horse?

The question lands differently when you’ve followed the argument all the way down. This isn’t a provocation anymore. It’s a real and serious question, that can be given a real answer. The honest answer for most people reading this is yes. You’re a general-purpose cognitive engine being slowly replaced by a cheaper engine, and it does the same job better.

This diagnosis is harsh, but being honest about this is the only way we can start any useful work. That’s what I’m diving into next here on Flux. In the meantime, I’d love to hear where you land on the question: What do you have that infinite tokens couldn’t reproduce?

Try This Skill

To make this important topic more accessible I’ve created a Skill that you can use with your favourite AI in 3 easy steps. Here’s how:

1. Download the SKILL.md file

2. Upload it to your AI

3. Then just say “Run this skill”

Flux ‘Are You A Horse?’ Explorer (download)— Explore and debate the ideas in this article with an AI that knows the arguments. Get more details about any specific point, or challenge the ideas and test the boundaries as you form your own opinion.

Note: It’s important to download this SKILL.md file, then upload it directly to your AI rather than just giving it the web link . We also recommend that you use one of the leading models from the 3 frontier labs (e.g. Claude, Gemini or ChatGPT). If you use a lower level model then your results may vary. And of course if you find any issues, or feel like you’ve found a real flaw in my arguments please let me know — I love constructive feedback and real debate.

Recently I've applied my own unique taste. I've always been good at picking strategic directions and being ahead of developments. Over the last few months I've had a lot of fun building flo.monster. Yet every time I'd run ahead, soon after the frontier labs would release things that directly impacted it.

That's great for my ego and re-inforcing faith in my "taste" - but that's also no way to make real progress.

So I stepped back and spent some time working out the best market segment, something that flo.monster could uniquely deliver for them, and in a way that was strategically defensible against the frontier labs. They're giant bulldozers - but they do have strategic drivers you can map out. Ones that they would be at least "unlikely" to change.

Then I built it - and it works great! It has an amazing onboarding process, really useful features, it's absolutely scalable, minimal running costs, robust security and agentically scalable too.

This is probably the best platform I've ever built - certainly in terms of "how you find out what it does" and "get from zero to useful".

Then before I released it, and rolled it out to a new batch of beta users to really started testing properly, I stepped back and asked myself

"Is this really what I WANT to do?"

The clear and honest answer was "no".

I think THIS is the biggest question right now! The strategic question - beyond creation - beyond taste. In this moment, what is it that YOU really WANT to do in this moment of change? In the midst of creation and destruction. There's so many interesting things you could do right now - and any real platform move means you "have" to stay focused on that - largely to the exclusion of anything else. Not a decision to be taken lightly.

If you think AI will create a wealth of new jobs, or if you don't - either way you need a plan.

Plus the arms race that the Mythos Red Team just highlighted is about to hit us all - so the ground is shifting under our feet yet again.

It's clearly an important moment, and honest decisions plus giving yourself freedom to really explore is critical.

How are you dealing with this question? How is this impacting you personally and what do you WANT to do?

That's why I've changed my substack to "Flux" and I have some interest ideas on this topic that I hope you'll like. But I'm not just going to post ideas, I have some new ways for you to explore them and see how they apply to you directly too...

PS: This is a chance to re-visit my TrustIndex work and ideas - but in a way that’s more specific to each of us.

I’m leaving the post below in place…but see my latest update to see how this has evolved…

I hope you’ve found the TrustIndex Signals, Briefings and Reports useful. I know I’ve learned a lot putting them together.

But a couple of things mean I’m going to pause this - just for a moment.

First, if you haven’t seen I’ve been working on a project that’s putting all the things I’ve learned here into action. You can read the “vision” essay here. And you can visit the project itself at flo.monster.

Second, my health just doesn’t let me do as much as I want.

If there’s one thing this project has shown me, it’s that things really are changing faster and faster. It’s not an illusion and it’s not just hype. The interface is thickening and we’re walking (if not running) into an AI-Mediated future.

The “World Generation” tools that were the focus of the first TrustIndex / Report have already moved ahead in leaps and bounds and this is just going to continue.

And the next area I’ve been looking at is “Realtime GenAI”. This is just starting to really hit it’s stride and will likely have a big impact over the next few months.

Plus ALL the other tools and developments that are impacting the TrustIndex dials. So much to discuss!

But for now this is “PAUSE”.

Hopefully I can re-visit this all soon, and I’m confident what’s here already will provide a useful landmark for us to compare against.

Thanks for reading…

It’s positioned as browser-level tooling for autonomous task completion, rather than a separate app experience.

This puts upward pressure on the Reality dial as agents take on more of our daily tasks.

When the browser can execute steps end-to-end, the user spends less time verifying pages directly and more time accepting an agent’s interpretation of what it opened, what it ignored, and what it decided was “good enough” to act on. That shifts the locus of trust from the raw web to the model-mediated interface.

As this pattern normalises, the “agent-as-interface” layer thickens - the model becomes the practical gateway to information and action, including which sources are surfaced and which are silently bypassed. In Reality terms, that concentrates power in the selection and sequencing logic, because the model isn’t just summarising the web - it is choosing and performing the path through it, with less user friction to notice mistakes.

> The interface layer is thickening. If you disagree with my interpretation, or you’ve spotted a better signal then reply and tell me.

The release packages capabilities that can be integrated directly into downstream apps and toolchains.

This puts upward pressure on the Fidelity dial.

- Lowers the cost and friction of producing near-human speech by making cloning and controllable voice design widely accessible.
- Widens distribution pathways: open-source availability speeds reuse in apps, agents, and localisation pipelines without bespoke vendor deals.
- Reduces the gap between “synthetic” and “indistinguishable” audio in everyday deployments, pushing expectations of realism upwards.
- Expands misuse surface area (e.g., scams and impersonation) by enabling high-quality voice generation with minimal barriers

> The interface layer is thickening. If you disagree with my interpretation, or you’ve spotted a better signal then reply and tell me.

This is positioned as broadly reusable speech-to-text infrastructure rather than a single product deployment.

This puts upward pressure on the Equality dial. Low-resource and historically unsupported languages get materially better baseline coverage, reducing exclusion driven by “language availability” gaps. Open access enables local organisations (education, civic services, disability support) to adopt ASR without waiting for commercial prioritisation of high-demand markets. Wider language coverage can shift where speech interfaces work reliably, improving access for communities that are typically left out of mainstream speech tech.

Open models lower experimentation costs, increasing the likelihood of targeted, community-led adaptations for under-served languages.

> The interface layer is thickening. If you disagree with my interpretation, or you’ve spotted a better signal then reply and tell me.

This Nature Electronics work is a strong signal: topology-based body sensor networks - where sensors are distributed across clothing (or the body) and the shape/topology of that network helps coordinate sensing and data flow.

In plain terms: you’re no longer measuring a couple of points (wrist, chest, ring). You’re turning large areas of your body into an instrumented interface.

That’s upward pressure on the Reality dial. Because as the sensing footprint expands, the AI layer gets more continuous, more intimate, and more “default”:

- More of you becomes measurable (movement, posture, physiology, environment contact).
- More of your day becomes legible (not just workouts or sleep - everything in between).
- More decisions get nudged by models that sit between your lived experience and what the data implies.

And topology matters because it’s a step toward wearable sensing that can scale beyond rigid, fixed placements - sensors embedded into fabric, arranged dynamically, and still meaningfully interpreted.

This is what “AI-mediated” looks like when it’s not just on screens: it’s stitched into the layer you live inside.

> The interface layer is thickening. If you disagree with my interpretation, or you’ve spotted a better signal then reply and tell me.