Welcome to the Greatest Hallucination

Wake up in the morning, brush your teeth, and open LinkedIn. Tell me, what do you see?

A stream of groundbreaking announcements. Experts absolutely thrilled to share. Everyone is founding something, shipping something, or explaining why the rest of us have eighteen months left to remain employable.

And you? You have a backlog.

Linear algebra, optimization, deep learning, LLMs, agents, MLOps. Papers to read, tools to learn, people to catch up to. You were halfway through understanding yesterday's future when somebody released tomorrow's.

You know what success looks like. A very smart person writes complex math, attaches an appendix full of Greek letters, and draws an architecture diagram that requires its own architecture diagram. We've learned to recognize the shape.

But how did that shape become the thing to pursue?

Let me take you to IKEA. You walk through rooms nobody lives in, admire the life arranged inside them, then go home and move your furniture. The staged room becomes the template for the real one.

Something similar happens with careers. The public performance of success becomes a specification for how to live. You choose the skills, the job title, even the problems that will make your life look like the picture.

Companies do it too. A board asks what a company fit for the future should look like. The answer comes from other companies' announcements. A roadmap follows, then a team, then a budget. By the next funding round, those commitments can be presented as evidence that the future is arriving.

Jean Baudrillard described how representations can become models reality is made to follow. His simulacrum helps us see what is at stake here. A story about the future can acquire authority over what gets built in the present.

The showroom starts issuing instructions.

In AI, this can happen on a remarkable scale. Expectations attract investment. Investment buys infrastructure and produces revenue for suppliers. That revenue becomes evidence for the expectations that attracted the investment.

A useful product can survive and still earn far less than the business built around it needs.

The Programmer Behind the App

Before the industry, start with one person.

Imagine someone builds their first serious application by vibe coding. They describe what they want, ask a model for changes, and keep going until it runs. There is a login screen, a database, a payment flow. They have shipped something people can use.

Then an order appears twice. They ask the model to fix it. It produces a patch. The duplicate disappears, but they cannot explain what caused it or why the patch works.

We might call this forgetting how to code. In this example, there was nothing to forget. They never worked through the implementation decisions we assumed the finished application demonstrated.

The application is real. So is the achievement of getting it built. They may have learned useful skills in directing and testing an AI system. But the result alone cannot tell us whether they can reason about the code when the model's next answer is wrong.

Once that app becomes a portfolio, it can stand in for a history of understanding its maker never acquired. An employer reads it that way. Its maker may start reading it that way too.

That is the interesting hallucination: the working product becomes evidence for a particular picture of the person behind it. The code is available to inspect; the competence we attributed to its author has barely been examined.

How the Story Takes Over

Baudrillard described four successive phases of the image, each further from the thing it once represented. Use them to watch what happens to the link between an announcement and what it promises. In practice the phases overlap. A single company can produce excellent research, exaggerate a product, and sell a future it has no idea how to deliver.

The Sign Reflects Reality

Start with an experiment. A model does better at a task. Other people inspect the method, compare results, and try to reproduce it. The claim has somewhere to answer for itself.

In 2011, IBM's Watson defeated two Jeopardy! champions. The challenge had rules, opponents, and a result people could watch. It was a real achievement in answering questions expressed in natural language.

The distance between the announcement and the result was short enough to walk.

The Sign Distorts Reality

Now imagine a team trying dozens of architectures and publishing the winner. The result can be real while the presentation hides how much searching it took, which comparisons were convenient, or how badly the method behaves outside the benchmark.

The distortion grows as the result travels. A paper demonstrates a capability under particular conditions. A presentation drops the conditions. A sales deck turns the capability into a product. By the time the story reaches the customer, the machine has acquired several skills its authors never tested.

There is no need for a central liar. Each person makes the story a little easier to tell.

The Sign Masks Absence

The promise expands from a demonstrated task to an anticipated capability.

A system can write code. The pitch becomes autonomous software engineering. It can answer questions about documents. The pitch becomes an employee who understands your business. The distance between them contains research problems, integration work, and sometimes a human quietly keeping the demo alive.

This is where AGI becomes commercially useful before anyone agrees how to recognize it. It gives very different investments a shared destination.

In January 2025, Stargate announced an intention to invest USD 500 billion in AI infrastructure over four years. Half a trillion dollars. Before any of it had been spent, the sheer size of the number could already work as an argument: surely nobody would contemplate spending that much unless the destination were close.

A proposed expenditure starts doing the persuasive work of a result.

The Signs Become Their Own Reference

Eventually, announcements begin referring to each other for authority.

A startup must have something because a major company invested. The investment must be wise because demand for compute is growing. Compute demand must demonstrate customer value because startups keep buying it.

Here is a simplified version of how that can be wired. Big Tech Giant invests USD 500 million in an AI startup. The agreement requires the startup to spend some of that money on Big Tech Giant's cloud. The startup announces its funding. The provider records revenue as it delivers the computing services. Both have something impressive to report.

The number is illustrative. The mechanism is documented: the FTC found that major cloud and AI partnerships included commitments to spend substantial investment funds on the investor's cloud services.

The provider delivered real services and earned real revenue. Investment is supposed to finance work before customers can pay for the result. The mistake comes when we count the funding announcement and the resulting cloud revenue as two independent confirmations that customers will eventually support the business. Some of the apparent corroboration comes from the same financing decision.

If those customers arrive and pay enough, the loop has helped build something productive. If they don't, another funding round may have to keep the apparent demand alive.

The same circularity shows up inside companies. Ask why a company needs an AI strategy and the answer may be that every serious company has one. Ask how we know serious companies need one and out come the strategies.

The loop has begun setting its own admission criteria. A project that fits the image of the future can look sensible before anyone identifies whose problem it solves. A useful project outside that image has to explain why it belongs.

Watson Goes to the Clinic

Two years after the televised victory, Watson was being prepared for a very different setting.

In November 2013, MD Anderson told readers that its Watson-powered Oncology Expert Advisor would soon be used by doctors and researchers, starting with leukemia. The ambition extended to all its clinics and to making specialist knowledge available beyond Houston. Later in the same article, the hospital explained that the system was still being tested and that the timeline for patients to receive it was unknown.

The university's later records put institutional decisions beside that public future. Its fiscal 2017 audit report summarized a procurement review: MD Anderson had adequate standard procedures, but they had not been consistently applied to the project under review.

The 2018 audited financial statements recorded a USD 26.5 million impairment on part of the Oncology Expert Advisor project in 2017. This reduced the asset's recorded value; it was not the project's total spending. The note still described the project as construction in progress.

An ambitious project can fail honorably. These records establish a financial disappointment and a failure to apply ordinary controls; they do not establish that the rhetoric caused either. They give us somewhere concrete to ask how much authority a promised future should have over the next deliverable and the next payment.

The same system, two years apart: first a sign that reflected a result, then a sign standing in for one. The showroom had become a project. It still needed to pass inspection.

The Nature of the Simulacrum

No shadow council needs to arrange any of this. The incentives can assemble it from ordinary people trying to get their next decision approved.

In Baudrillard's terms, the product carries sign value as well as use value. It can signal that a company is innovative while contributing very little to its operations. A complicated architecture can make a team look sophisticated while making the service harder to maintain.

The two values can coexist. Trouble begins when the rewards for the appearance become strong enough to weaken the checks on the substance.

Why Awareness Changes So Little

You know all this already. You roll your eyes at the announcements. You recognize the demo with a person behind the curtain.

Then your manager asks about the AI roadmap.

Your competitors have one. The board expects one. An engineer wants experience that will still look current on a CV next year. A founder needs to close a round before payroll. Each has a reason to participate, even if each privately doubts the performance.

This is the part I find hardest to dismiss. The system does not require universal belief. It can run on people who think they understand the bullshit and can manage their exposure to it. Peter Sloterdijk called this enlightened false consciousness. Žižek's shorter version: they know very well what they are doing, but still, they are doing it.

Cynicism becomes comfortable accommodation. We laugh, update our profiles, and carry on.

Somewhere inside the performance, people are also doing excellent work. The incentives rarely give us the courtesy of separating everyone into honest builders and dishonest narrators. The same person can solve a difficult problem in the morning and oversell it in the afternoon.

The Glitches

The useful check comes after the demo, when someone has to do the job with the tool. Some tools pass it impressively.

In a field study of 5,172 customer support agents, AI assistance increased issues resolved per hour by 15% on average, with larger benefits for less experienced workers. That is worth getting excited about.

Now imagine turning that finding into a forecast for replacing knowledge work. The forecast helps justify a valuation, the valuation attracts investment, and the investment is cited as confirmation that the forecast is sensible.

The measured improvement has acquired a much larger family of claims. It cannot support them all.

The glitches worth watching are closer to the business itself. A customer signs a pilot but never expands it. Time saved drafting reappears as time spent checking. Usage rises while the provider loses more money serving it. A company announces impressive revenue but depends heavily on customers funded by its own investors.

Watch what happens when results disappoint. Does the explanation get more specific, producing a better test or a different product? Or does the payoff simply move further into the future?

Where the Loop Can Break

A lab misses its revenue target and needs more funding to keep buying compute. If the next round fails, it reduces its cloud spending. The provider loses some of the growth that helped justify its investment in the lab.

The feedback can run backwards.

The adjustment could come through slower construction, acquisitions, or returns that disappoint investors while the products keep improving. A sudden crash is one possible outcome. A long, expensive disappointment would also send invoices.

Three pressure points decide which one we get.

Paying demand. Customers have to capture enough benefit to pay for the service. Providers have to retain enough of that payment to cover the full cost of delivering it and financing the infrastructure. Competition could make powerful AI cheap and widely useful while leaving investors with disappointing returns. Customers would quite reasonably call that progress. The shareholders might choose another word.

Distribution. If savings accumulate with owners while displaced workers lose purchasing power, demand can suffer. Lower prices, new work, or redistribution could offset that loss. The outcome depends on how the gains reach people, a question no company's plan to reduce its own wage bill can settle.

Research. A company can approve a bigger training run. It cannot approve the discovery that run is expected to produce. A business can commit money and people to the next attempt. Its financing becomes fragile when it also needs the attempt to deliver a particular breakthrough by a particular date. What happens when the next leap takes longer than the funding cycle? The research problem gets a deadline, and the researcher gets a performance review.

What Would Change My Mind

Growing numbers of customers renewing at prices that cover the full service cost would strengthen the case for durable demand. So would adoption that survives the end of subsidies, with measured benefits that hold up in normal operation. If those customers generate enough cash to support the infrastructure, the next funding round becomes less decisive.

The opposite pattern would deepen my concern: impressive usage maintained through discounts, pilots that never become routine purchases, and repeated refinancing of obligations whose payoff keeps moving away.

A theory that can absorb every outcome becomes another way to protect its narrator. Calling everything a simulacrum should not give me that privilege.

The Reload

Suppose some of the financial story fails. Useful technology can survive the correction. People who built it can still lose their jobs. Infrastructure may pass to new owners, sometimes owners with more power than before.

A crash does not award the machines to the most deserving engineer.

The dot-com crash did not erase the usefulness of the internet. It did expose the distance between a technology becoming important and particular investments paying off. An industry can eventually transform the world and still bankrupt people who correctly saw that transformation coming.

A financial reset also leaves plenty of room for the social machinery to continue. The appetite for prestige, the fear of missing out, and the demand for an impressive future will find new material.

There is no pristine normal waiting underneath the market chart. There can still be better ways to build.

Surviving Simulacra

The Matrix offered a wonderfully convenient product: one pill, permanent clarity. Baudrillard himself said the film misunderstood him, which seems about right. I would still like a prescription.

We have careers, bills, and work worth doing inside these systems. We need ways to make a promising story answer to the work, before it acquires a department that depends on keeping the story alive.

When choosing a project, follow a customer's actual task. What became easier? What still requires human checking? Include integration, review, and failure in the cost. The customer who renews from an operating budget tells you something a conference applause meter cannot.

When choosing a technical foundation, examine what you can inspect, reproduce, and replace. Open code and available model weights can give more people that ability. Work out the remaining dependencies too: chips, compute, expertise, and maintenance. Independence needs an actual budget.

At work, give uncertainty somewhere useful to go. A team should be able to report that a pilot disappointed, or that a simpler method worked better, without making its own survival depend on a more exciting explanation. Otherwise, everyone learns which results are welcome.

And keep a record of the original promise. What was supposed to improve, by how much, and under which conditions? Decide what would make you stop before the quarterly review gives you reasons to continue.

You can enjoy the technology. You can build ambitious things with it. You can even be thrilled to share.

Then, a few months later, share what happened.