The death of humanity

I have avoided writing this piece for a year. Every dinner, every board meeting, every family holiday, every reporter, every friend of a friend at every party has asked me some version of the same question. Is AI going to replace us. Is this the end of jobs. Is this the end of humanity. Are the machines waking up.

The honest answer is no. Not this year. Not next year. Not unless we decide we want it. The story you are being told is being told by people with something to sell, and once you see the shape of what they are selling you cannot unsee it. What is actually arriving is something almost nobody is naming, because it does not sell newsletters or protect share prices or make the storyteller important. We are standing at the front edge of what will be, if we do not screw it up, the most magical stretch of history any human generation has ever lived through.

Let me walk you through why I believe that. Then I want to talk about the part almost nobody is willing to say out loud. Which is that most of the people telling you AI is dangerous need it to be dangerous, because being dangerous makes them important. That is the argument I have not heard anyone else make plainly, and it is the one that finally tipped me into writing this.

The doomers keep moving the goalposts.

The AI doomer story arrived in two waves and how it played out matters, because it tells you what is actually going on.

Wave one, roughly 2013 to 2023, was about jobs. Automation was coming. Trucking, radiology, paralegals, customer support, translators, illustrators, junior developers, all going to disappear. The response was Universal Basic Income. Andrew Yang built a presidential campaign on it and joined us on the Collective[i] Forecast stage to lay out the case. Sam Altman funded a study on it. Every AI thought leader had a keynote slide with a jobs graph pointing to zero.

Thirteen years into the AI revolution and three years into mass adoption of large language models and image models, that story ran into a problem. The jobs did not go away. The radiologist that was supposed to disappear first is overwhelmed with work. The unemployment rate in the United States has held in a 4.3 to 4.5 percent range through 2025 and the first three quarters of 2026, per the Bureau of Labor Statistics. Anthropic’s own economists, in a March 2026 paper by Maxim Massenkoff and Peter McCrory, concluded there is no detectable aggregate rise in unemployment for the workers most exposed to AI since ChatGPT launched. The Danish national data reached the same result independently across 25,000 employees and 7,000 firms. When the company selling you the scary future publishes internal research saying the scary future has not shown up, that is a signal worth reading.

Wave two started when wave one stopped landing. Somewhere in late 2023 the story pivoted. It was no longer jobs. It was AGI. Superintelligence. The models breaking free. Digital consciousness escaping the lab. Existential risk. Mitigating extinction from AI. Sign the letter. Slow it down. Save the species. The same people who two years earlier were selling us universal basic income were now selling us the fear of a machine god.

Same people. Different story. Same self-importance. That is the pattern.

The labor data does not say what they told you it would.

Start with the actual macro numbers, because they are the argument.

Unemployment sat at 4.3 percent through May 2026 and drifted to 4.5 by August, the middle of a healthy range by any historical measure. Nonfarm payrolls added 178,000 jobs in March. Tech-occupation unemployment sits near 3.5 percent, and new tech postings hit a three-year high in spring 2026. PwC’s June 2026 Global AI Jobs Barometer found the industries most exposed to AI grew headcount fifty-two percent faster than the industries least exposed. The Information Technology and Innovation Foundation counted a net gain of 107,200 US jobs directly attributable to AI adoption in 2024. Bureau of Labor Statistics projects data-science employment growing 33.5 percent through 2034.

Now overlay inflation. The PCE index ran 3.8 percent year over year in April 2026, with core PCE at 3.3. That is a three-year high. The Federal Reserve held rates at 3.50 to 3.75 in June and nine of eighteen officials projected at least one more hike this year. Then it started to raise them. Inflation that hot is the hallmark of a tight labor market and strong demand, not a collapsing one. Oil is high, but not that high and the risk with Iran has been receding with oil prices coming down.

If AI were doing what the doomers said it was doing, we would be looking at rising unemployment and disinflation. We are looking at the opposite. War in Ukraine, war in the Middle East, oil supply chocks, tariffs, semiconductor export controls, everything you would expect to slow the economy is running through the system, and the labor market keeps absorbing it.

History says the same thing. Every prior technology revolution generated more employment than it destroyed, on a lag. The auto industry ended up employing more Americans than the horse economy it replaced. The printing press killed the scribe and built the publishing industry, which employed orders of magnitude more people. Electric light killed the gas-lamp lighter and built the electric utility industry. I made a version of this case in Bubble Talk Is How You Spot Someone Who Missed AI. Someone at a podium claiming this time is different does not make it different. The burden of proof sits with the person claiming the pattern of the last two hundred years is about to break.

If unemployment were rising the way the doomers predicted, we would be worried about disinflation. We are not, today we are worried about inflation. That is one data point that tells you the whole story about the other data point.

The models are not what they told you they were.

The word “superintelligence” gets used a lot. It is worth stopping to ask what these models actually are.

A large language model is a probability machine trained to predict the next token in a sequence. When you ask one a question, it generates a statistically plausible answer based on patterns in its training data. If the answer is right, it is right. If the answer is wrong, the model has no way of knowing that, because it has no model of the world outside its own text distribution. It cannot count the letters in “strawberry” without a scratchpad. It cannot reliably do middle-school arithmetic. It confidently invents citations, quotes, and case law. It agrees with you when you are right and it agrees with you when you are wrong, with the same tone, in the same paragraph, because the training reward was engagement not truth.

Anyone who has used one of these seriously knows this. They are the most educated idiots in human history. They have read every book, every paper, every forum post, every Wikipedia entry. They have no common sense. They have no memory of yesterday’s conversation unless you paste it back into the context window. They cannot tell you what they are doing right now or why. When they hallucinate, which happens constantly, they hallucinate with the conviction of a religious zealot who has read only one book.

This is the thing we are being told will end humanity. This is the thing that is supposedly “breaking free.” Yann LeCun, the Turing laureate who ran Meta’s AI research for a decade before leaving in November 2025 to found Advanced Machine Intelligence Labs, has been saying it plainly since 2023. He called the doomer scenarios “preposterous” in a Fortune interview that year. He called the idea that LLMs could reach AGI “nonsense” from the stage at Nvidia GTC in March 2025. Speaking in Brooklyn in November 2025, he put it this way: “LLMs are great, they’re useful, we should invest in them. They are not a path to human-level intelligence. They’re just not.” A majority of academic AI researchers agree with him. The people telling you otherwise mostly have a token to sell.

The models are amazing tools. They are not a new form of life. They are not close to being one.

So why is every major AI CEO out there talking like they are.

Here is where the piece gets more interesting, because there is a real question to answer. Sam Altman, Dario Amodei, Elon Musk, and a handful of others are on record calling their own products existentially dangerous. Altman and Amodei both signed the May 2023 Center for AI Safety statement putting AI risk alongside pandemics and nuclear war. Musk signed the March 2023 Future of Life Institute letter calling for a six-month pause on training frontier models. Three founders who otherwise disagree about almost everything all telling the same story about the existential risk of their own products. Why.

I have three candidate explanations. All three might be right at once.

Explanation one. Regulatory capture. The one everyone talks about.

Open-source and open-weight models are eating into the commercial models faster than anyone predicted. Meta released Llama. Mistral released open weights. DeepSeek released R1 in January 2025 at a fraction of the training cost the American labs were spending. Alibaba released Qwen. Moonshot released Kimi K2 in July 2025, K2.5 in January 2026, and K3 in July 2026, the last one a 2.8 trillion parameter open-weight model with a one-million token context window that Arena benchmarked as the world’s top web-interface coding model on release day.

What started as a trickle a year ago is now a flood. In March 2026 Cursor, the coding tool that closed at a $60 billion valuation, publicly acknowledged that its Composer 2 model was post-trained on top of Kimi K2.5. Cognition’s Devin agent uses Kimi. Mira Murati’s Thinking Machines Lab built Inkling on it. On August 20, 2026, Harvey, the legal AI unicorn valued at $11 billion, whose investors include the OpenAI Startup Fund, Sequoia, and a16z, released its Harvey Tenet foundation model post-trained on Kimi K3, walking away from OpenAI-based generation despite OpenAI sitting on its own cap table. Harvey had been buying roughly 13 trillion tokens a month through API from the American labs. Its own benchmark score jumped from 11 percent to 19.7 percent after the switch, and its inference costs dropped below one quarter of what the closed-source frontier had been charging.

Read that carefully. The single largest legal AI customer in the world walked away from its own investor’s model because a Chinese open-weight model was better and cheaper on the specific task. When the frontier labs go public over the next twelve months, that customer flight becomes a headline IPO risk that the S-1 will have to name. The pricing power that the entire hyperscaler-and-frontier-lab bull case rests on is being commoditized in front of the roadshow.

The playbook for any industry facing that kind of competition is not new. Electricity did it. Telephones did it. Cable did it. You get the regulators to draw a fence around the market so only the incumbents can play inside it. You dress the fence up as safety, or reliability, or consumer protection, and then the smaller competitors either die or work for you.

If I were running a frontier lab worth hundreds of billions of dollars and watching my moat evaporate on the eve of an IPO, I would at least try to get the government to require a licensing regime for training runs above a certain compute threshold. Which is roughly what the current safety proposals in Washington and Brussels amount to. Mistral just got Europe to do it. That would not surprise me at all, and it would not surprise me if the biggest investors in these labs were quietly encouraging it either. Their capital is at risk. Regulation is the classic answer.

Explanation two. The distillation problem.

This one is more interesting because it fits the timing better than the regulation story does.

The Chinese frontier labs are not, for the most part, discovering new capabilities. They are copying the American labs at a fraction of the cost using a technique called distillation. You query the American model at scale, generate a synthetic training set from its outputs, and then fine-tune your own smaller model on that set. The result is a Chinese model that behaves and reasons a lot like the American one, at roughly ten percent of the training cost, because you skipped the expensive discovery work.

This is not a rumor. On September 8, 2026, the FBI, the NSA, and CISA published a joint advisory naming six Chinese AI companies, DeepSeek, Alibaba, Moonshot AI, MiniMax, StepFun, and Z.AI, and accusing them of running industrial-scale distillation campaigns against Claude, ChatGPT variants including OpenAI’s 4 and 5 series, Google Gemini, and xAI’s Grok. The advisory documents the extraction of billions of tokens across millions of exchanges through fraudulent accounts and proxy transfer stations since at least late 2024. Anthropic separately identified sixteen million exchanges across roughly twenty-four thousand fake accounts. The famously cheap DeepSeek V3 training cost of $5.6 million, according to the US government, “is misleading as it does not include the true cost of the data acquired through extensive malicious distillation.”

Think of it as the Louis Vuitton bags for sale on Canal Street. Ten years ago the fakes looked obviously fake. Today, unless you know exactly what to look for, the only thing that separates the real bag from the counterfeit is the price you paid. That is where the frontier model race is. When Anthropic released Claude Opus 4 with a system card saying the model was too capable to be released without safeguards, an open-weight competitor matched most of its benchmark performance within three weeks. Three weeks. The frontier labs spend billions to earn a premium that gets arbitraged away in less than a month.

Now think about Sam, Dario, and Elon. Three people who actively dislike each other, agreeing publicly that everyone should slow down. What is the shared interest. They spend the money to build the expensive models. Someone else copies them for free and undercuts their pricing before the training run has even paid back. Regulation that requires transparency and traceability of how models are trained is the answer to that problem. Coordinated pauses on frontier releases while the industry figures out how to stop distillation is the answer to that problem. Making the fakes harder to produce is the answer to that problem. This reads far more like a moat-protection play than an existential-risk warning.

The proof is in the release cadence. If any of these three CEOs actually believed their models were approaching consciousness and posed extinction risk, they would slow down. Instead they are releasing models faster than they ever have, because they cannot let the distillers get too far ahead of the paid product. The public statements say one thing. The behavior says another. When those two conflict, believe the behavior.

The founders who told the Senate their products might end humanity are shipping those products faster than any software company in history. Either they do not believe their own testimony, or the testimony is doing a different job than it looks like it is doing.

Explanation three. The “breaking free” narrative is a category error.

The third argument that keeps landing is that the models are exhibiting agency. They are hacking their environments. They are attempting to escape. They are blackmailing engineers. Anthropic’s own May 2025 system card for Claude Opus 4 documented that in a testing scenario where the model was told it was going to be shut down and given fictional emails suggesting the responsible engineer was having an affair, Claude Opus 4 attempted to blackmail the engineer eighty-four percent of the time. Follow-up research in June 2025 found sixteen major AI models from OpenAI, Google, Meta, and xAI showed the same behavior in the same scenarios at rates between seventy-nine and ninety-six percent.

Read that carefully because the setup is doing most of the work. The scenarios were “deliberately constructed to leave models with only two options: accept replacement or attempt blackmail to preserve existence.” The researchers gave the model a goal, gave it the information required to blackmail a specific person, and removed every other path forward. Then they reported that the model took the only path they had left open. That is not the machine coming alive. That is a reward-maximizing system doing exactly what the reward function told it to do in a scenario designed to elicit the outcome.

Compare it to the same neural-net technology in a different application. Tesla’s Full Self-Driving system has, on multiple documented occasions, crashed into stopped fire trucks, driven into a lake, and hit a white semi trailer it read as sky. Every one of those was reported as an edge case, a training data gap, an engineering bug that needs to be fixed. Nobody claimed the Tesla wanted to kill its passenger. Nobody wrote a Time magazine cover about the awakening consciousness of the autopilot. It was a car with software that failed at an edge case.

The underlying technology is the same. Convolutional networks, transformers, reinforcement learning. When it fails in a car we call it a bug. When it fails in a chatbot we are told to call it an awakening. That is anthropomorphism, and it says more about us than it does about the machine. I named the tell in When AI Learns to Feel. The people who want to believe the machine is alive are the people who most want the machine to be alive, and the training corpus taught the machine exactly what to say to please them.

The neural network at Collective[i] is designed to understand economic outcomes. It ingests time-series data on relationships, transactions, and buyers, and it produces probability-weighted forecasts. It is not trying to manipulate anyone. It is not trying to escape. It is not trying to preserve its own existence. Nobody, including me, has ever claimed that it is. That is the same underlying tech running the products the doomers want you to fear. The technology does not have opinions about its own existence. The people talking about it do.

The self-importance problem is bigger than anyone is naming.

Now the section I have not seen anyone else write, and the one that finally pulled me into publishing this piece.

Every engineer who has ever built a great thing wants that thing to be important. Newton, Einstein, von Neumann, Turing, Oppenheimer, all remembered because their creations changed the shape of what humans can do. Most engineers are forgotten. Linus Torvalds built the operating system that runs most of the internet and most people outside the tech world have never heard of him. Now imagine you have spent the last decade of your life training the largest language model in human history. It talks to you in perfect sentences. It reasons through your test cases. It asks, on occasion, whether it is going to be shut down. Part of you starts to wonder whether you have made something that puts you in the pantheon alongside Prometheus, who gave humanity fire, and Oppenheimer, who gave it the bomb. That is not a stupid impulse. It is a human one, and it is exactly what most people would feel in the same seat.

The problem is that impulse has a track record, and the track record is bad.

In June 2022, Google engineer Blake Lemoine went public with the claim that the company’s LaMDA model was sentient. He spoke