They Killed an Elephant to Win an Argument. Watch the AI Pitch Do the Same Thing.

On January 4, 1903, a crowd gathered at Coney Island to watch an elephant die.

Her name was Topsy. Handlers fitted her feet with copper-lined sandals, ran roughly 6,600 volts through her body, fed her cyanide first, and looped a rope around her neck in case the current came up short. A film crew stood by to capture every second. The reel survives. More than a century later you can still watch an animal get executed for a Sunday crowd.

If you want to understand Dario Amodei and Sam Altman in 2026, start with that dead elephant. The men running the two most valuable AI companies on Earth are working the same move that killed her. Manufacture fear. Cast yourself as the only one who can be trusted to manage it. Let the dread sell the product. Amodei and Altman have built their empires on a single emotion, and it is not wonder. It is fear.

The story that got attached to Topsy is that Thomas Edison electrocuted her to prove his rival's electricity was deadly, to win the War of Currents, his direct current against George Westinghouse's alternating current.

That version is mostly legend. By 1903 the war was already over. AC had won years earlier. Edison had been pushed out of the company that carried his own name, and he almost certainly was not there. His film company just showed up to record footage that would sell.

But the legend stuck for a reason. Because a decade before Topsy, Edison ran exactly that campaign, and it was real, and it worked.

In the late 1880s he was losing on the merits. AC carried power across long distances cheaply. DC could not. He knew it. So he stopped arguing engineering and started selling dread. His people staged public electrocutions of dogs, calves, and horses to show crowds what AC did to living flesh. He lobbied to power the first electric chair with AC so the public would tie his rival's current to death. He tried to turn "Westinghoused" into slang for being executed. He printed a pamphlet warning families that AC would kill them in their beds.

None of it was about safety. The better product was winning, so the man behind the worse one manufactured fear and sold it wholesale.

That is the playbook Amodei and Altman are running today, at a scale Edison could never have imagined. The pitch is that they are building a technology so powerful it could end the world, and that you should be grateful they are the responsible adults holding the leash. Trust us, they say. We are the only ones who can be trusted with this.

It is a brilliant sales strategy. It is also, over and over, wrong. And the clearest place to watch them be wrong is open source.

The receipts

Start with Altman. In February 2019, OpenAI announced it had built a language model called GPT-2 and decided it was too dangerous to release. A 1.5 billion parameter autocomplete. They warned it would flood the internet with fake news, spam, and impersonation at scale. Headlines wrote themselves. One outlet ran with the idea that OpenAI had built something so powerful it had to be locked up for the good of humanity.

The open-source community reverse-engineered it in weeks. Comparable open models shipped within months. The world-ending propaganda weapon turned out to be a glorified sentence-finisher that rambled about unicorns and, in its own output, described fires happening underwater.

That episode set the template. Claim unprecedented power. Warn of a unique danger only you can manage. Collect the headlines. Then release the thing anyway once competitors catch up. It has run on a loop ever since.

Now Amodei. He has stood in front of Congress and warned that turning frontier models loose in the open creates risks nobody can contain. Once the weights are public, he argues, you cannot revoke access, patch the safeguards, or trace the misuse. He is careful to say small and medium open models are fine. It is the frontier ones that will take us somewhere very dangerous. He has pushed hard to keep the strongest AI out of China's hands on national security grounds.

Here is what actually happened. Open-weight models keep catching up in months, not decades. By 2026, Western companies are quietly ripping out expensive closed APIs and switching to cheaper Chinese open-weight models that are closing the capability gap. The thing that was supposedly too dangerous to release is now running on laptops, for free, worldwide.

And here is the part the roadshow never says out loud. Open models are the single biggest threat to Anthropic's and OpenAI's paid business. When the man selling closed, metered API access tells you the free open alternative is dangerous, that is not a neutral safety assessment. That is a merchant warning you about his competitor. Even AI industry writers have named it plainly: Amodei's open-source warnings line up almost exactly with Anthropic's commercial interest.

His concern might be sincere. The incentive is still there. And when the warning and the business model point in the same direction, you weigh the warning accordingly.

Now watch the last 30 days

Here is the timeline. Not a metaphor. What happened.

• June 1: Anthropic files a confidential S-1 to go public. Same day, it closes a $65 billion round at a $965 billion valuation.
• June 8: OpenAI files its own confidential S-1, one week later.
• June 9: Anthropic ships its most powerful public model, Fable 5, alongside the restricted Mythos 5.
• June 12: The Commerce Department slaps export controls on both, citing national security. Anthropic cannot verify user nationality in real time, so it pulls the models for everyone on Earth. The trigger was a report that Fable 5 could be talked into writing exploit code.
• June 26: The White House asks OpenAI to limit its new GPT-5.6 models to a small group of government-approved partners. OpenAI complies.
• June 30: The controls come off.
• July 1: Fable 5 is back, worldwide.

Read that sequence again. Two companies file for what could be the largest tech IPOs in history. Days later, their flagship models are declared so dangerous the federal government has to lock them in a vault. Then, right on cue, the vault opens. The danger is not a bug in the pitch. The danger is the pitch.

The tell

Here is the part that should end the argument.

Anthropic itself said the scary capability was not special. In its own notice, the company admitted that other existing models, including OpenAI's ChatGPT 5.5, its own Claude Opus 4.8, and a Chinese open model called Kimi K2.7, could find the same vulnerabilities Fable did. Half a dozen models could produce the same proof-of-concept code. Anthropic called the whole thing a borderline case involving routine defensive cybersecurity work.

So the unprecedented power that took the models offline was, by the vendor's own admission, not unprecedented. It was already sitting inside a free Chinese open-weight model. The exact thing they keep warning is too dangerous to open-source was already open-sourced, and the sky did not fall.

And the export controls did the opposite of what they promised. Demand ran straight to those cheaper open models. One Chinese lab's shares jumped 30% on a new release. DeepSeek closed a funding round near $7.4 billion. The containment made nobody safer. It just moved the buyers and proved, again, that you cannot stuff this capability back in a bottle. Which is exactly what open-source advocates have been saying the whole time.

When you cannot out-build it, ban it

Then last week the point made itself.

On July 16, a Chinese lab called Moonshot dropped Kimi K3. An open model. 2.8 trillion parameters, the largest open-weight model ever shipped, with the full weights going public on July 27 for anyone on Earth to download and run for free.

It did not just keep pace. In blind developer testing it took the number one spot on the Frontend Code Arena, ahead of Fable 5. It topped the Next.js evals. It leads on a stack of coding and agentic benchmarks, matching or beating the flagship that got locked in a government vault six weeks ago for being too dangerous to export. It still trails Fable 5 on some reasoning tests, so it is not a clean sweep. But an open model you can run on your own hardware is now trading punches with the most restricted model in America.

And it costs a rounding error. Fable 5 runs about $50 per million output tokens. Kimi K3 runs $15. DeepSeek's comparable model runs under $4. You are being told the expensive, closed, gated model is the safe one, while the cheap open one that beats it on real work is the threat.

So watch what Amodei and Anthropic do now that they are losing on price and closing fast on capability. They do not out-build it. They lobby to ban it.

Anthropic has been running a campaign since February to get Chinese open models restricted. Blog posts. Letters to members of Congress. A report accusing DeepSeek, Moonshot, and MiniMax of running tens of thousands of fake accounts to distill Claude. A letter to officials accusing Alibaba of doing the same. Every accusation wrapped in the language of national security: foreign labs, military use, surveillance, offensive cyber. The ask underneath all of it is simple. Make the competition illegal.

One of the most respected independent voices in AI research called it what it is. Regulatory capture. Anthropic would gain enormous economic security if the Chinese models it keeps naming got pulled off the US market. TIME said the quiet part out loud too. If Chinese labs release these capabilities open-source, they erode the exact business model Anthropic and OpenAI are about to sell to public shareholders.

This is the Edison move, live. He could not beat AC in the lab, so he tried to get it legislated into the electric chair and scared the public about the wiring in their own homes. Amodei cannot out-ship a free 2.8 trillion parameter model, so he writes to Congress about national security and asks the government to take it off the board.

Same fear. Same target. Your mind. Change what you are afraid of, and you change what you buy, and what your representatives are willing to ban.

The pump you are supposed to buy

Let me be blunt about what this looks like from the cheap seats.

Anthropic is running a roughly $47 billion revenue run rate and chasing a valuation near a trillion dollars, targeting a listing as soon as this fall. Goldman Sachs and Morgan Stanley are steering both IPOs. Anthropic projected its first profitable quarter, then quietly noted the profit might not last because it leans on a discounted compute deal. That is the picture underneath the world-ending rhetoric. A company that needs the number to stay enormous just long enough to sell shares.

We already watched the dress rehearsal. SpaceX priced at $135 on June 11, ran to $225 within days, then gave back about a third of the gains. Pump, then fade. That is the pattern the entire AI IPO class is studying right now.

I am not accusing anyone of a crime. I am telling you what it reeks of. When the story that inflates the valuation is the exact same story the founders have broadcast for free for years, follow the money before you follow the fear. Fear is the demand generator. The scarier the model sounds, the bigger the number gets. That is not a safety framework. It is a sales funnel wearing a doomsday costume.

Y2K wants a word

If you are old enough, you already lived through one of these.

Y2K was a real technical problem. Two-digit year fields, a narrow and specific bug. It was also wrapped in an apocalypse. Planes falling out of the sky. Grids collapsing. Bank accounts vaporizing at midnight. An entire consulting industry got rich selling the panic, and global remediation spend ran into the hundreds of billions.

Then the clock struck midnight and almost nothing happened.

Some of that calm was earned by real fixes, and I will give the engineers their due. But the gap between the apocalypse we were sold and the quiet non-event we got was enormous, and a lot of people made fortunes inside that gap. The lesson was never that all warnings are fake. The lesson was to verify the actual technical scope yourself, and to be very suspicious of anyone selling you both the fear and the cure.

That is the same lesson Topsy teaches. It is the same lesson this month teaches.

What the models actually are

I build with these tools every day. Not on a keynote stage. At a real desk, shipping real systems, auditing the cost of every single inference call because the vendors would love for me to stop looking.

So here is what frontier models actually are, from the terminal instead of the press release.

They are powerful. They are also unreliable in ways the marketing never mentions. They hallucinate. They cut corners. They compound small errors into large ones the moment you let them run unsupervised at scale. The teams getting real results are not the ones chasing the newest model drop. They are the ones building scaffolding, setting explicit rules, keeping a human in the loop, and containing the thing so its mistakes cannot spread.

That is the truth every working engineer knows and no IPO roadshow will say out loud. These are not the world-ending superintelligences in the pitch deck. They are extraordinary tools that need discipline, structure, and adult supervision. Genuinely useful. Nowhere near the god-machines the founders describe when the cameras are on and the shares are about to price.

The distance between the marketing and the command line is the whole story.

Follow the incentive

Edison could not win on the engineering, so he electrocuted animals in front of crowds and called it a public safety demonstration. The current he was fighting powered the next century anyway.

A hundred and twenty years later, the tools are different and the move is identical. Warn that open models are too dangerous, right as open models threaten your business. Declare your own product too powerful to release, right as you file to sell shares in it. Lobby the government to ban the competition when you cannot out-build it. Cast yourself as the only responsible steward of a terrible power. Let the fear do the selling. Then ring the bell on the exchange.

Do not buy the doom. Follow the incentive. When the loudest prophet of catastrophe also has a ticker symbol coming, the catastrophe is usually the product.

Watch the reel. Then go verify it at your own desk.