The Anticipatory Disruption Trap

Why the AI layoff wave will reverse and why that’s not reassuring enough

Photo by Enecta Cannabis extracts on Unsplash

There is a particular kind of panic that spreads faster than evidence. It travels through earnings calls, LinkedIn posts, and executive memos until it becomes received wisdom until companies act on it not because they’ve verified it but because not acting feels riskier than acting. That panic is now running through the technology industry, and it is reshaping careers, degree choices, and company structures in ways that the underlying technology does not yet justify.

The headline version: AI agents are here, coders are done, junior engineers are redundant, and anyone studying computer science right now is training for a job that won’t exist. Companies like Meta, Accenture, and dozens of others have begun restructuring headcount around this premise. It is presented as a forward-thinking strategy. It is mostly a reaction.

This essay argues something simpler and, I think, more defensible: the disruption is real, but the timeline is wrong, and the companies acting on it now are making a costly mistake they will quietly reverse.

I. The Practitioner’s Report

When evaluating whether a tool is ready for a job, the most useful testimony comes not from the person selling the tool but from the person who actually tried to use it and had the technical depth to know when it failed.

George Hotz is that person here. The founder of comma.ai, author of tinygrad, the teenager who first jailbroke the iPhone and unlocked the PS3, he is not a skeptic of AI. He is a builder who ran the experiment honestly.

His verdict, published in May 2026, is worth quoting at length because the texture of the failure matters:

“I’ve now used AI agents on two non-trivial tasks — writing parts of tinygrad, and reversing a USB↔PCIe chip. Each time, I suspect I could have done it better and faster manually. The agent frontloads all the progress, then gives you a slot machine lever to pull hoping it gets the polish done. It never quite gets there.”

The slot machine metaphor is precise. The agent produces something that looks like 80-90% of a solution, plausible-looking and increasingly hard to distinguish from correct, and then spins indefinitely on the last 10%. For a toy project, that’s tolerable. For production code at a company that just laid off the engineers who would have caught the failure, it’s a slow-burning liability.

Hotz goes further on the organisational dimension:

“Agents will end up hurting large organisations more than high-performing individuals or small orgs. Large orgs have much slower feedback loops, much less alignment — and the bottom performers won’t have the self-check that high performers apply instinctively. They’re the ones producing 10x output with agents. What happens to the average output of that organisation?”

This is the structural point that the layoff headlines miss entirely. The productivity gains from agents are real but unevenly distributed; they amplify the already-good and mask the mediocre until the debt surfaces. Organisations that cut their human quality bar first are precisely the ones that will be slowest to catch it.

And then, in the same post, Hotz makes a statement that positions him clearly in an ongoing technical debate:

“I’m now in the LeCun/Marcus camp on LLMs. I don’t think models like this will ever be able to program. Real programming agents will need world models, not some RLVR shit that comments out the failing test.”

II. The Theory Behind the Practice

Hotz is the empirical signal. Yann LeCun is the structural explanation.

LeCun’s critique of large language models is not vibes-based contrarianism — it is a specific architectural argument that he has been making consistently since at least 2022 and that he has now put a billion dollars behind by leaving Meta and founding AMI (Advanced Machine Intelligence) in late 2025 to pursue world models instead.

The argument clusters around four missing capabilities:

1. No grounding in physical reality. LLMs are trained on text — a thin, discrete shadow of the world. They have never observed forces, objects, or causality directly. They model the distribution of language about the world, which is not the same as modelling the world itself.

2. No persistent memory. Each inference is stateless. An engineer working on a codebase builds up months of contextual understanding about why decisions were made, what was tried and failed, and where the bodies are buried. A model resets every call.

3. No genuine reasoning. Current models can produce outputs that resemble reasoning, but the process is not compositional inference from first principles. It is pattern-matching over a vast training distribution. When the pattern breaks — at the edge cases that define hard engineering work — the model fails in ways that are hard to predict and diagnose.

4. Language is an interface, not the engine. Perhaps LeCun’s sharpest point: in a truly intelligent system, language would be like Broca’s area — a narrow module that converts internal representations into speech. Designing intelligence around the language interface is like designing a car around the horn. LLMs are, in this framing, a very impressive horn.

LeCun is careful to acknowledge utility: “LLMs are becoming really good at generating code, and they are probably going to become even more useful in a wide area of applications.” Useful, yes. Sufficient for the end-to-end engineering work companies are betting on? No. His billion-dollar bet is precisely that the path to general intelligence runs through world models, not through scaling LLMs further.

When Hotz says he is now in the LeCun camp, he is saying, ‘I tried the thing; it failed in exactly the way the theory predicts, and the failure is architectural, not a matter of more parameters or better RLHF.

III. The Historical Template

We have been here before. Not metaphorically, almost exactly here.

In 2016, Geoffrey Hinton, the godfather of deep learning and the man whose students Google poached to automate diagnostic imaging, made the following statement:

“I think if you work as a radiologist, you are like the coyote that’s already over the edge of the cliff but hasn’t yet looked down. People should stop training radiologists now. It’s just completely obvious within five years deep learning is going to do better than radiologists.”

This became the canonical example of AI job displacement. It was cited in newspapers, boardrooms, and policy documents. Medical school applicants weighed it when choosing specialisations.

Here is what actually happened. Jensen Huang, in a conversation with Joe Rogan that has since become widely referenced, laid it out plainly:

“Radiology, for example, has largely been converted to AI-driven radiology. The surprising thing is: the prediction that all radiologists would be the first jobs to go was exactly the opposite. The trend shows that there are more radiologists being hired now as a result of AI. You have to go back to what is the purpose of a job. The purpose of a radiologist is to diagnose disease, not to study the image. The image studying is simply a task in service of diagnosing the disease.”

The mechanism Huang describes is important: AI made image analysis faster and more precise, so radiologists could read more scans per day, which meant hospitals could serve more patients, which improved economics, which meant hospitals hired more radiologists. The American College of Radiology, in a study published in February 2025, predicted radiologist numbers in America would grow by up to 40% between 2023 and 2055.

But the most instructive part of this story is not the data. It is what Hinton himself said afterwards. He acknowledged he had spoken too broadly — that he was only referring to image analysis, not the full scope of the role. He revised his prediction: most medical image interpretation would be performed by a combination of AI and a radiologist, with AI making radiologists more efficient and improving accuracy rather than replacing them.

The man who predicted the extinction of radiologists now predicts their augmentation. The field, he said, was walking off a cliff and is now understaffed and growing. This is the complete arc, i.e., prophet, debunker, and prophet’s retraction, and it lives in a single case study.

IV. Task vs. Purpose

Jensen Huang’s framing — task versus purpose — is the conceptual key that connects all three threads.

AI, in its current form, can automate tasks. Reading an image. Writing a function. Summarising a document. It does these things impressively and at scale. What it cannot yet do is hold the purpose — the ‘why’ behind the task, the judgement about which tasks matter, and the contextual understanding of what failure looks like and how to avoid it.

A radiologist’s purpose is to diagnose disease. Scan analysis is one task among many that serve that purpose. When AI took over the task, the purpose expanded: radiologists could diagnose more, catch more, and treat more. Demand grew.

A software engineer’s purpose is to solve problems — to understand a system, identify what is wrong or missing, and produce a change that makes it better without breaking what already works. Writing code is one task in service of that purpose. When Hotz used agents to write parts of tinygrad, the code was produced — the task was partially completed — but the purpose required him to review, catch, and fix what the agent missed. The slot machine never got to the polish. The purpose was not served.

LeCun’s architectural critique says the same thing differently: language models operate at the task layer. The purpose layer requires world models, persistent memory, causal reasoning — capabilities that current architectures do not have and that cannot be obtained by scaling what we already have.

This is why Hotz concludes that real programming agents will need world models. Not because the task is beyond them — the task output is sometimes impressive. But because the purpose requires something more fundamental than next-token prediction over code.

V. The Temporal Mismatch

Here is the actual argument, stated plainly.

The layoffs happening right now, the junior engineers cut, the CS students fleeing the field, and the headcounts restructured around agent productivity are based on tools that a practitioner of Hotz’s calibre rates as not close to the bar at any serious organisation. They are based on an architectural premise that LeCun, having thought about it longer and harder than almost anyone alive, has concluded is wrong — wrong enough to leave a trillion-dollar company and raise a billion dollars to pursue a different path. They are based on a historical analogy – radiologists – that, when you actually look at what happened, runs in the opposite direction from what the narrative assumes.

The disruption is real. The tools are genuinely useful. The productivity gains at the task layer are not imaginary. But the leap from “useful at tasks” to “replaces engineers end-to-end” is not supported by the evidence from the people who have actually tried it, the theory from the people who have thought about it most carefully, or the history of every previous time we made this prediction.

By the time agents can genuinely handle end-to-end software engineering — with the world models, persistent memory, causal reasoning, and error-correction that the purpose requires — two things will have happened. First, the companies that over-cut will have quietly rebuilt their human capacity, having learned the hard way what Hotz documented. Second, new job categories will have emerged around the complexity that didn’t get automated: the roles that supervise agents, validate their outputs, design the systems they operate within, and make the judgments about purpose that the tools cannot make.

That is not a guess. It is the structural pattern of every previous automation wave. ATMs and bank tellers. Spreadsheets and accountants. CAD and architects. The tool expands the envelope of what is possible; the envelope expansion creates demand; demand creates new roles.

VI. The One Real Risk

None of this means the disruption narrative is harmless. Its worst damage is not the layoffs themselves — companies over-hire and over-cut in every cycle, and the rehiring will come.

The real risk is the knowledge pipeline.

Students are already fleeing computer science. If the “entry-level jobs are gone” narrative calcifies into received wisdom, if it shapes degree choices, career counselling, and hiring pipelines for the next five years, then the industry will arrive at the moment when agents become genuinely powerful and purpose-layer humans are most needed and find a generation-scale shortage of engineers who understand what they are doing.

That is the actual cliff. Not the one Hinton described in 2016 — that one turned out to be a gentle slope with more radiologists at the bottom. This one: a pipeline disrupted at exactly the wrong moment, producing scarcity in the thing that becomes most valuable precisely when it is gone.

The fundamentals still matter. The knowledge pipeline still matters. And the companies disrupting themselves based on benchmarks rather than production evidence are, as Hotz put it, living in a golden era for slop and a dark age for quality without yet knowing which they are producing.

References

• Lila Shroff, “There’s Never Been a Better Time to Study Computer Science,” The Atlantic, 2026

• George Hotz, “The Eternal Sloptember,” geohot.github.io, May 24, 2026

• Jensen Huang, Joe Rogan Experience, November 2025 — radiologist example; also Davos / WEF, January 22, 2026, NVIDIA Blog

• Jensen Huang, quoted in CNBC: “Jensen Huang cited radiologists to dispute AI jobs impact,” December 4, 2025

• American College of Radiology, study on radiologist workforce growth (40% increase 2023–2055), February 2025

• Geoffrey Hinton, original 2016 statement on radiologists; subsequent retraction acknowledging combination model

• Yann LeCun, multiple public statements on LLM limitations and world models, 2022–2026; AMI founding, late 2025