Who Needs Humans, Anyway?

In 2025, Keytruda and its newer formulation, Qlex, brought Merck about $31.7 billion in sales. Anthropic's revenue that year was around $4.6 billion. Almost seven times as much for one drug brand. This is a comparison of revenue, not profit: the figures appear in Merck's results and Anthropic's IPO filing.

The comparison is already going out of date. Anthropic reported $16.3 billion in revenue for the first six months of 2026. What interests me is the business taking shape behind those numbers. We talk about how much money you can make selling access to intelligence. Owning what that intelligence invents may turn out to be far more valuable.

On 27 October, Anthropic announced that it had begun drug research of its own and acquired Mithril Medicine, a one-person company. Much of the work described is still preclinical. It doesn't have a new Keytruda. In its IPO filing, the company explicitly describes an opportunity to capture a share of the new economic activity AI creates.

I'd expect large AI companies to try different ways of collecting that share. Rights to drugs, materials and manufacturing processes look more attractive than charging for model use alone. That's my hypothesis about the incentives, not an announced strategy to own every invention. If intellectual work is getting cheaper, a business owner would like to find something everyone else still has to pay for.

A good idea, however, hasn't yet passed trials, obtained approvals or become a manufacturing operation. A patent protects a particular invention in particular territories for a limited period, usually twenty years from filing. Obtaining one requires technical disclosure. A closed model and a patent therefore do different things: you can conceal how a model works, while a patent legally restricts the use of an invention you have disclosed. Owning a clever chatbot doesn't let you patent the whole of medicine. WIPO explains the limits.

Still, the direction is interesting. In Dario Amodei's description, powerful AI is a "country of geniuses in a datacenter". I'd like to know who owns the work produced in that country, and what rights are left to those outside it.

In the previous essay, I used Escort Economy to describe a situation where people increasingly need to appeal to the owners of resources. Now another question bothers me: what if those owners eventually have no reason to choose most of us at all?

I use Claude, ChatGPT and agents myself. It benefits me that work which once required more people can now be done by a smaller team. I'm not going to pretend I reject the technology to preserve the old order. But my gain doesn't establish that everyone else will find a place in the new one.

Market power rarely begins with a declaration of feudalism. Someone makes a good product. People arrive. Convenient infrastructure grows around it, and other businesses build on top. Leaving gets harder: you have to move data, change processes, retrain staff and find customers again. The platform owner gains the ability to change terms that other people's livelihoods now depend on.

AI adds computing, energy, laboratories, capital and market access to the arrangement. With large initial costs, a big player can spread them across many customers. Integrations and habits hold users in place. If that player then starts making its own products in its customers' industries, it becomes both a supplier of an essential tool and a competitor.

In January 2025, the FTC described risks in partnerships between cloud companies and AI developers: restricted resource access, switching costs and unequal access to sensitive information. That isn't a verdict against every partnership. Dependency can grow out of contracts that seemed worthwhile to each party at the time.

There is a catch. Competition doesn't have to end in a single corporation. AI lowers entry costs and helps small teams. But when scale confers strong advantages and leaving is expensive, a market can settle around several large owners who compete with each other while setting the terms for everyone else. Having three landlords doesn't make a tenant an owner.

This dependency is what I mean by feudalisation. You're legally free. You can start a business. The tools are available. But the conditions of your existence increasingly depend on property you don't own, and find it harder to acquire through your work.

The strangest part is that the environment in which a person tries to work out who they are is changing too.

Language exists because people speak, negotiate, argue and pass on experience. In that sense, it is something we produce together. But none of us arrives in an empty world and invents it from scratch. Along with language, we inherit other people's distinctions between dignity and shame, success and failure, a normal life and one that urgently needs fixing.

A child hears what a successful person is supposed to look like. An adult reads which skills are in demand. Later, they discover that even the conversation inside their head uses words they once heard from others. Some help them understand what's happening. Others make them spend years pursuing something they didn't particularly want.

This doesn't mean language determines all thought. Fedorenko, Piantadosi and Gibson's review of the research explains why thinking cannot be reduced to language. But speech and writing give us knowledge of how the world works, and where we might fit into it.

Producing those texts is now a machine service. A model can help write a proposal, answer it, assess a candidate and explain to the candidate how to become more appealing. We can imagine AI taking part in almost every such transition: education, finding work, choosing products, discussing politics, trying to understand our own desires.

Suppose you come to a service for advice on finding yourself. It helps define the person you ought to become, then offers tools to help you fit that image. You get the criteria for success and your self-improvement plan from the same place. You can keep paying for improvements forever. The moment when you've done enough is particularly inconvenient for that business model.

I think this is a significant difference in the present technological revolution. Automation reaches both work and the means by which we learn our next job, make sense of events and find one another. More than the list of available professions is changing. The terms of our participation in the world are changing with it.

It's tempting to say that we are language models ourselves, just biological ones. There is a useful analogy. We learn from others' examples, anticipate the next part of a sentence and adapt answers to the person listening. Goldstein and colleagues did find shared computational principles between an autoregressive model and human brains processing speech. The study involved nine people listening to a story.

That doesn't establish that the whole brain works like an LLM. People have bodies, pain, needs, relationships they remember and consequences to live with. Similarity between particular mechanisms doesn't prove identity. Even if a machine reproduces more of our abilities tomorrow, what it can do won't determine the value of our lives.

For me, the analogy is a reminder of how much outside material goes into forming an opinion I call my own. If a few companies produce a substantial part of that environment, they gain influence over the questions we ask in the first place. They don't need to issue direct orders. Making some choices familiar and convenient, and others hard to reach, can be enough.

There may be more apparent variety at the same time. Everyone gets a personal feed, a personal conversational partner and a preferred way of having things explained. Behind a million personalised worlds, one company may still set the terms of access. The number of voices doesn't tell us how many independent sources of power there are.

A similar tension exists in religions and other large systems for explaining the world. A shared teaching helps strangers trust one another, accept obligations and cooperate beyond the family. But when an institution claims exclusive authority over truth and acquires coercive power, an alternative explanation threatens its position. It can be suppressed, declared an error or absorbed on subordinate terms.

This isn't a universal law of religion. Religious life includes coexistence, schisms and new movements. Faith can give people grounds to resist economic power. What interests me is the condition under which absorption becomes possible: cooperation is worthwhile, or even permitted, only within one order, while the possibility of existing outside it shrinks.

A company uses different means, but it too may have an interest in bringing more of other people's activity inside its infrastructure. Buy a competitor. Offer an independent developer a convenient platform that becomes difficult to leave. Use an open tool as the entrance to a paid service. Users still have plenty to do, but fewer ways of doing it on their own terms.

At this point, it's easy to reach for a biological story about a species that defeats every outside rival and starts eating itself. It's a vivid story. It isn't a law of nature.

Cooperation and rivalry coexist. Cooperation needn't arise for the good of the entire species: kinship, reciprocity, reputation and the structure of relationships can sustain it. Martin Nowak examines those mechanisms in his work on the evolution of cooperation. Similar needs can bring individuals into conflict over resources long before any supposed peak. The gains from cooperation and the temptation to keep those gains for yourself can grow together.

For economics, the more useful question is who receives the results of collective work. A company may have excellent cooperation internally while its output belongs to a small group of owners. Cooperation hasn't disappeared. It simply hasn't guaranteed each participant equal rights to what they produce.

Language, science, culture and technology were also created by many people. That shared history doesn't automatically create a shared claim on the income of every product built with them. Between collective contribution and the distribution of ownership sit specific laws, contracts and political decisions. That's where I'd look for the mechanism before invoking biological inevitability.

Now return to the person being told, once again, to reinvent themselves.

We have a comforting story: the peasant went to the factory, the factory worker moved to the office, and the office worker went off to build a personal brand. In reality, these transitions were uneven, painful and far from universally available. But the story reassures us. There was always another use for people. There will be one this time too.

Earlier transitions did have mechanisms that restored demand for labour. As some tasks were automated, others emerged. In Acemoglu and Restrepo's model, these are distinct processes that can reinforce or offset one another. Higher productivity alone doesn't promise that new demand for people will make up for displacement.

What worries me now is the possibility that a new task will almost immediately become available to the same system that removed the old one. Learning something may get easier. Selling what you've learned may get harder. This still isn't a description of the whole economy: there is a physical world, machines cost money, new needs emerge, and people retain advantages in particular tasks. But "just retrain" now contains an assumption we need to test.

A social revolution in these conditions would look different too. Workers have power while their participation is necessary: refusing to work can halt production. If the share of necessary human labour shrinks, that form of pressure may weaken. An AI owner may also influence both production and the information systems people use to organise collective action.

That doesn't mean resistance is impossible or politics is over. People remain citizens. Governments make laws. Infrastructure has physical dependencies. But replacing the people at the top may not be enough. If access to production and communication is still controlled by a few owners, the new leadership inherits the old power.

I'd look for the substance of a new revolution in the ability to remain independent as demand for our labour falls. That means ownership, rights to infrastructure and a say in its rules. "Make yourself even more useful" leaves all three questions to the owner.

Open intelligence has a practical meaning here. Can I run a system independently of its supplier, move my memory and data elsewhere, choose different rules? Publishing model weights alone doesn't deliver all that. You also need computing, knowledge and access to the means through which the system acts. If all of it requires a single owner's permission, even a good personal AI remains a way of renting someone else's power.

Then comes the question people usually try to move past. Suppose human contributions no longer pay for themselves in one field. Then in many fields. On what basis will people receive the means to live?

The answer from a buyer of labour may be unpleasant. They were buying work they needed. That transaction contains no obligation to support a former supplier. You can explain that the owner needs people as consumers. But consumers' purchasing power depends on how they receive an income. We're back at distribution, having taken the scenic route.

This is where "how will a person earn the right to live?" starts to look like a trap. We've already conceded that the right is granted by whoever the person must be useful to. All that remains is to prepare a convincing presentation of our own existence.

I don't want to have to interview for permission to live.

In the Universal Declaration of Human Rights, dignity, life and an adequate standard of living don't depend on a worker's profitability. We can argue about how to secure those rights. Abandoning them for economic efficiency would still be a specific political choice. No model makes it for us.

Saying that is easier than arranging it. Food, housing, energy and care require resources. If the productive system can manage with fewer workers, we need to establish how everyone else gains a claim on part of its output. Public funds, employee ownership, widely held assets, taxes and transfers, accessible services: different combinations are possible. Each has costs, opportunities for abuse and political opponents.

But the question has changed. We're asking how to distribute what is produced and the power over its production. We aren't asking an owner to find us an endearing enough occupation that they'll agree to feed us.

Payments alone don't guarantee independence. If housing, access to knowledge and participation in public life depend on one supplier's decisions, even a financially supported person can remain a dependent. Material support has to leave room to refuse and to organise your own life.

That's also why the promise that we'll all move into social media doesn't satisfy me. Attention is limited, while content can be generated automatically. Even where human presence is valued for its own sake, its income isn't distributed among everyone who wants it. A court may need considerably fewer courtiers than there are former programmers.

It's easy to fall into the opposite mistake: if production no longer needs me, my life must be meaningless. Again, we've accepted someone else's system of valuation as our own.

A child doesn't need the most efficient parent on the global market. They need someone who cares for them. Friendship doesn't lose its value because a more interesting conversational partner exists. You can draw even if a machine draws better. Curiosity, involvement in another person's life and the ability to make a commitment and live with its consequences don't have to win a productivity contest to matter.

I can't promise that love, care or responsibility will earn everyone a wage. Selling them as the new irreplaceable professions would mean trying to sell the right to live all over again. Someone can have reasons to live even when the market won't pay for them. Society would then have to secure the material conditions for that life.

A person also belongs to their environment as someone whose wellbeing matters. When technology cures a disease, part of its value is that someone gets to live their life. Drug revenue doesn't account for all of it. Participating isn't exhausted by passing information from one process to another. We can decide which changes are needed, whom they help and what price we're willing to pay. We'll have to keep those decisions even with a more capable adviser at hand.

For me, reinventing myself now means understanding both my own aims and the dependencies through which I try to reach them. I can use AI while asking which decisions I keep, what I control and whom I can cooperate with without a platform's permission. That requires personal work and political institutions. Leave either out, and it is easy to end up with helpless outrage or another course in successful adaptation.

And it isn't true that nobody talks about this. Amodei himself, in Machines of Loving Grace, discusses the meaning of human activity and the possibility that our familiar economic organisation will stop working. He acknowledges the lack of a clear answer. Recognising a risk doesn't, however, oblige the owner of new infrastructure to share either income or power.

You can start production, sign a contract, buy an asset. The discussion about fair distribution can carry on. These processes run at different speeds. An entrepreneur needs a decision today, an employee needs the rent paid, a politician needs to answer the next crisis. The future gets whatever is left after the urgent matters have been dealt with. Moloch from the previous essay managed perfectly well without hatred. A payment calendar will do.

Carpe diem is understandable in these circumstances. Sometimes a person just needs to have a good day. But it's dangerous for a society to make that its only relationship with the future. While we try to become more employable, someone else is already defining how employability will be measured.

Perhaps AI will preserve much more work than we fear losing. Perhaps it will open routes to independent income we can't yet see. I want that future, and I use tools that may help bring it about. But I'm not prepared to make the right to live and take part in a shared life depend on an optimistic labour-market forecast.

If our civilisation learns to manage without most human labour, it will have to decide what it needs people for. I'd like us to answer while we still have the right to decide what counts as an answer.

Because standing before your own civilisation with a CV, explaining why the fuck it should keep you alive, doesn't sound much like the liberation of humanity.