The thieves are upset about theft
Inspiration for this (thoughtful) diatribe.
Did you know Edison was kinda psycho?
In 1908, Thomas Edison had a problem. People were making movies without paying him.
Edison held key patents on motion picture cameras. So he did what any reasonable monopolist would do: he formed the Motion Picture Patents Company (MPPC), a trust that controlled every major film studio and Eastman Kodak. If you wanted to make a movie in America, you paid Edison.
If you didn’t pay, Edison sent men to break your camera.
The MPPC hired thugs who physically smashed equipment and disrupted film shoots. They limited films to two reels. They forbade actors from being credited by name. They monopolized distribution through a single company that acquired every film exchange in the country. Edison sued one independent producer, Carl Laemmle, 289 times.
When the Supreme Court finally ruled against Edison’s trust in 1917, Justice Clarke called it “the perfect instrument of favoritism and oppression”:
The primary purpose of our patent laws is not the creation of private fortunes for the owners of patents.
Independent filmmakers responded by…running.
They traveled 2,500 miles west to a dusty suburb of in California called Los Angeles. Chosen for three reasons: it was far from Edison’s enforcers in New Jersey, the Ninth Circuit Court of Appeals wouldn’t enforce his patent claims, and if things got really bad, Mexico was a hundred miles south. Edison’s patents didn’t apply in Mexico.
That dusty suburb was Hollywood. Arguably, the entire American film industry as we know it exists because filmmakers had to physically outrun a patent monopolist.
Flagstaff no good for our purpose. Have proceeded to California. Want authority to rent barn in place called Hollywood for $75 a month.
Crazy thing is Edison’s camera was built on Eadweard Muybridge’s sequential photography and George Eastman’s flexible film strips. He packaged other people’s inventions, patented the package, then hired muscle to stop anyone else from doing what he did.
In 2015, a court would have called this “standing on the shoulders of giants.” In 1908, Edison called it “intellectual property.” When others stood on his shoulders, he called it theft.
Sixteen million API calls
This week (February 23 2026), Anthropic published a blog post titled “Detecting and Preventing Distillation Attacks” along with the Post (is an X post a Tweet?) embedded at the top.
The language reads like a Clancy-esque counterintelligence briefing. They identified three Chinese AI labs: DeepSeek, Moonshot, and MiniMax. The claim: they were running “industrial-scale campaigns” to extract Claude’s capabilities through 16 million API exchanges across 24,000 fake accounts.
The details are technically impressive. Anthropic claimed MiniMax alone generated 13 million exchanges targeting agentic coding. DeepSeek used 150,000 exchanges to extract reasoning capabilities. When Claude’s model updated, MiniMax shifted nearly half its traffic within 24 hours. Hydra cluster architectures. Proxy networks running 20,000 simultaneous fake accounts.
Earlier this month (Feb 12, 2026), OpenAI sent a memo to the House Select Committee on China warning that DeepSeek was trying to “free-ride” on American AI research. They claimed DeepSeek employees built code to systematically harvest ChatGPT’s outputs for training data. OpenAI asked Congress to act.
The framing from both companies is national security. The word “attack” appears repeatedly. The proposed solutions involve export controls, intelligence sharing with government authorities, and stronger verification systems.
Nobody in these announcements mentions what the AI labs themselves trained on. Or that Anthropic was founded in 2021 by Dario and Daniela Amodei, who left OpenAI — taking what they’d learned building GPT-3 and using it to start a competitor. The entire company exists because its founders took knowledge developed at one lab and used it to build another. That’s doesn’t look like a scandal to me, it’s how our industry works. But it does make the “distillation attack” framing land a little differently.
But they forgot to tell us something
Anthropic downloaded over 7 million books from pirate sites to train Claude. At least 5 million from Library Genesis. At least 2 million from Pirate Library Mirror. They settled with authors in August 2025 for $1.5 billion — the largest copyright settlement in U.S. history.
OpenAI trained GPT-3 on Common Crawl, a ~9.5-petabyte scrape of the open web. Common Crawl made up over 80% of GPT-3’s training tokens. They also used datasets called “Books1” and “Books2” whose contents they have never disclosed. When the New York Times sued, Bing Chat reproduced 394 of 396 words from a single article. OpenAI told the UK House of Lords it would be “impossible to train today’s leading AI models without using copyrighted materials.”
Meta torrented 81.7 terabytes of pirated books — roughly 17 million ebooks — from LibGen, Z-Library, and Anna’s Archive. Internal employee messages included “Torrenting from a corporate laptop doesn’t feel right” and “I feel that using pirated material should be beyond our ethical threshold.” They did it anyway.
Getty Images sued Stability AI because Stable Diffusion sometimes generates images with a distorted Getty watermark still visible. One output read “imaiges” instead of “images.” Twelve million photographs. The lawsuit seeks to decrease the national debt at a sum of $1.8 trillion.
GitHub Copilot has been said to reproduce licensed open-source code verbatim — with the license attribution stripped. The amended complaint alleges GitHub “optimized the temperature setting of Copilot to produce small cosmetic variations of the Licensed Materials as often as possible.”
Just so we’re on the same page. The claims we’re discussing are active federal lawsuits with discovery underway.
PC Gamer’s headline on the OpenAI-DeepSeek complaint: “The brass balls on these guys.”
The company named “Open”
OpenAI was founded in December 2015 as a nonprofit. The founding blog post shared:
Since our research is free from financial obligations, we can better focus on a positive human impact.
Researchers will be strongly encouraged to publish their work, whether as papers, blog posts, or code, and our patents (if any) will be shared with the world.
OpenAI has changed its mission statement six times in nine years. A November 2025 IRS filing revealed they removed the word “safely” from their mission. In October 2025, they restructured into a for-profit corporation. They’re now valued at $850 billion. Microsoft owns 27%.
After DeepSeek’s open-source model gained traction, Sam Altman said: “I personally think we have been on the wrong side of history here and need to figure out a different open source strategy.”
The company named “Open” AI, who promised to publish everything and share all patents; spends seven times more on lobbying than it did two years ago and wants Congress to stop others from learning from its model outputs. The outputs of models trained on everyone’s copyrighted work.
It’s like they’re trying to anthropomorphize r/LeopardsAteMyFace.
History repeats itself
In 1979, Steve Jobs visited Xerox PARC. The deal: Xerox got 100,000 shares of pre-IPO Apple stock for about a million dollars. Apple got a demo of Xerox’s Alto computer — mouse-driven GUI, overlapping windows, WYSIWYG editing. Jobs was so overwhelmed by the GUI that he could barely focus on anything else shown that day.
Apple built the Macintosh on what they saw. In 1988, Apple sued Microsoft and HP, claiming Windows copied the Mac’s “look and feel.”
The court found 179 of Apple’s 189 claimed GUI elements had already been licensed to Microsoft. The remaining 10 weren’t copyrightable. The ruling said Apple “cannot get patent-like protection for the idea of a graphical user interface” under copyright law. Xerox sued Apple separately, but their case was dismissed because they waited too long.
Apple took the idea from Xerox for $1 million. Then spent five years in court trying to stop Microsoft from doing the same thing.
Quick and dirty OS
Microsoft’s origin story is the same joke told from a different angle.
In 1980, Tim Paterson at Seattle Computer Products wrote QDOS — Quick and Dirty Operating System. It was a clone of Digital Research’s CP/M, the dominant operating system of the time. Microsoft bought full rights for $50,000 two weeks before IBM shipped the first PC. Seattle Computer Products later sued, claiming Microsoft hid its IBM relationship to get a cheaper price. They settled for $925,000.
A decade later, Microsoft licensed the Spyglass Mosaic browser source code for $2 million. Spyglass later sued because the royalty deal was based on revenue per copy, but Microsoft bundled Internet Explorer into Windows for free so there wouldn’t be any revenue or subsequent royalties. Microsoft eventually settled for $8 million.
Microsoft then used its browser. They built on code derived from Marc Andreessen’s work at NCSA to destroy Andreessen’s company, Netscape. Bundling IE with Windows was free while Netscape charged $49. By 1999, a federal judge found Microsoft had committed monopolization under the Sherman Act.
$50,000 for a cloned OS. $2 million for a borrowed browser. Then antitrust for crushing the people they borrowed from.
The ten dollar patent lapse
AT&T’s telephone monopoly has the most absurd version of this.
Alexander Graham Bell’s patent was filed on February 14, 1876. This was the same day as Elisha Gray’s nearly identical patent caveat. Bell was the 5th filing that day. Gray was the 39th. Bell’s patent included a liquid transmitter concept that appeared in Gray’s drawings but not Bell’s own work. The patent examiner, Zenas Wilber, was later accused of leaking Gray’s filing to Bell’s attorneys.
Let’s not be distracted by these two. Remember that Antonio Meucci built a voice-communication device 27 years before Bell in 1849. He filed a patent caveat in 1871. He couldn’t afford the $10 annual renewal. It lapsed in 1874. In 2002, the U.S. House passed a resolution acknowledging that if Meucci had ten dollars, Bell wouldn’t have gotten the patent.
Bell Telephone (later AT&T) litigated 587 patent challenges between 1876 and 1893. They never lost. When the patents expired, 6,000 independent phone companies appeared. AT&T refused to let any of them connect to Bell’s network. Some cities had two separate phone systems. You needed two phones.
It took until 1984 to break up the Bell System. A monopoly born from a $10 patent lapse and a possibly corrupt examiner lasted over a century.
An actress and the chipmaker
Our story takes us back to Hollywood to Hedy Lamarr, the actress and inventor. She co-invented frequency-hopping spread spectrum in 1942 to support the Allies. U.S. Patent 2,292,387. She and composer George Antheil designed it for torpedo guidance. Incredibly, neither received a cent.
Later, Qualcomm built CDMA on decades of spread spectrum research, including Lamarr’s work. They promised fair, reasonable, non-discriminatory patent licensing when CDMA was adopted into cellular standards. Then they deployed a “No License, No Chips” policy that forced phone makers to pay 5% of the entire phone’s price as royalty. Apple was paying $12-$20 per iPhone. Apple reportedly said Qualcomm had “a gun to our head.”
Less than a cent per record
The record industry’s version is the cruelest because it harmed so many individuals who adopted “streaming” too soon.
Little Richard sold the publishing rights to “Tutti Frutti” for $50. He also received less than a penny per record sold. Standard artist royalty rates were 8-15%. Labels kept most of the rest.
Fast forward to when Napster appeared, those same labels sued in the name of “artists’ rights.” The RIAA filed 261 lawsuits against individual users in 2003. Jammie Thomas-Rasset, a single mother, faced three separate jury trials for sharing 24 songs on Kazaa. Verdicts hit $1.92 million — $80,000 per song. Joel Tenenbaum, a college student, owed $675,000 for downloading 30 songs. The Supreme Court declined to hear either appeal.
Then streaming arrived and the labels took equity stakes in Spotify. Warner sold its stake for $504 million. Sony sold half for $768 million. Artists get $0.003 to $0.005 per stream. It takes roughly 200 streams to match a single iTunes download.
They paid Little Richard nothing. They sued a college student for $675,000. They called both “protecting artists.”
Pulling up the drawbridge
None of this is surprising if you look at the staggering shoes the AI companies need to fill.
As I write & revise this in late Feb 2026, OpenAI is valued at $850 billion and Anthropic at $380 billion. These companies collectively spent $1.1 billion on political spending to oppose state AI regulation. Meta deployed 86 lobbyists in the first half of 2025 — one for every six members of Congress. OpenAI increased its lobbying spend seven-fold. Big Tech pushed a provision in a spending bill to ban state AI regulation for 10 years.
The lobbying spend, the security framing, the appeals to Congress are all pointing when a company tries to pull up the drawbridge. It’s definitely not about safety or national security. It’s the $1.2 trillion in combined valuation that depends on nobody else catching up cheaply. Note, that $1.2T doesn’t consider the valuation impact on the tech sector as a whole.
Regulatory capture is the move. Patent examiners grant considerably more patents to firms that later hire them. In agribiotech, USDA regulators who moved to industry jobs approved products 4.2 days faster in the two years before they left, worth $8 million per approval to the hiring firm. Tech is running the same play with AI policy.
Here’s what I believe.
Distillation is not an “attack.” It’s what happens when you publish an API. Every API is a teaching interface. If you serve billions of requests, you’re broadcasting your model’s behavior to anyone with a credit card and a logging script. This is not a flaw in the system — it is the system.
The AI labs know this but they’re scared. DeepSeek built R1 for a fraction of the cost of o1. Alibaba released Qwen3.5 under Apache 2.0; open weights, competitive with Opus and GPT-5 on reasoning, free to download. The drawbridge isn’t being pulled up against one lab. It’s being pulled up against a tide. If distillation works, frontier models are more vulnerable than the fundraising decks suggest. That’s the real threat — not to American AI leadership, but to the specific companies burning $10 billion a year on training runs.
Innovation has always worked this way. You build on what came before. Edison built on Muybridge. Apple built on Xerox. OpenAI/Anthropic built on Common Crawl, Books3, and the entire copyrighted internet. And now the world is building on them.
Isaac Newton wrote to Robert Hooke in 1676:
If I have seen further it is by standing on the shoulders of Giants.
Coincidentally, the phrase wasn’t Newton’s. It traces to Bernard of Chartres in the twelfth century.
Years later, Hooke accused Newton of stealing the inverse square law from their correspondence. Newton was so furious he deleted every reference to Hooke from the Principia. The man who is credited the most famous line about building on others’ work then erased his colleague’s contributions from history.
I wish I was making this stuff up. Even the popular metaphor for collaboration has a backstory about theft and credit.
The question isn’t whether distillation is “fair.” That argument died the moment these companies scraped the entire internet without asking. The question is whether we’re going to let the first generation of AI monopolists lock the door behind them using the same government they’re lobbying to regulate their competitors.
Edison tried it with cameras. Microsoft tried it with browsers. The record labels tried it with file sharing. Disney tried it with copyright — lobbied for extensions twice, despite building a $200 billion empire on public domain fairy tales.
The AI labs will not be the exception. Someone will always build on what came before. That’s not a bug in innovation. That’s the whole thing.