🥛 America’s AI is being copied 👀

GM. This is Milk Road AI, where we track who's inventing the future and who's copying it through the back door.

Here’s what we’ve got for you today:

✍️ The AI distillation wars just went nuclear.
🎙️ The Milk Road AI Show: Why Your First Humanoid Robot Won’t Be at Home... It’ll Be at Work.
🍪 Meta goes all-in on AMD chips.

Okara’s SEO & AI Visibility Agent helps your business rank higher on Google and get cited in AI answers by running daily audits for SEO/GEO visibility issues and recommending high-impact fixes that boost traffic and conversions. Try out Okara’s SEO & AI Visibility Agent today.

Prices as of 10:00 a.m. ET.

THE BIGGEST BRAIN HEIST IN HISTORY

In 1848, the British Empire had a problem.

China had a monopoly on the world’s most valuable commodity: tea.

The Chinese government had made it a capital offense to export tea plants or share cultivation techniques with foreigners.

So what did the British do?

They sent a Scottish botanist named Robert Fortune into China disguised in local dress and posing as a Chinese official.

He snuck into restricted plantations, stole thousands of tea seedlings and seeds, persuaded local experts to share processing techniques, and shipped everything to British‑controlled India.

Source: Tenor

Within a few decades, India’s tea production overtook China’s as Britain’s main supplier.

Britain didn’t just steal a product. They reverse‑engineered an entire industry.

The Chinese called it theft while the British called it “botanical research”.

Sound familiar? Fast forward 178 years, and the tables have turned so hard they’ve broken through the floor.

This time, it’s China accused of stealing America’s most valuable export.

And it’s not tea, steel, or semiconductors but intelligence itself.

Three of America’s biggest AI companies, Anthropic, OpenAI, and Google, publicly accused Chinese AI labs of systematically stealing their models’ capabilities.

And the method? It’s not some James Bond hack; it’s far more elegant and far more terrifying.

The technique is called distillation, and it’s basically the AI equivalent of sitting next to the smartest kid in class and copying every answer until you’ve absorbed their entire thought process.

Instead of stealing model weights (the actual “brain files”), Chinese labs allegedly took the front door.

They signed up for API access like a normal customer, blasted millions of carefully crafted questions at the models, and used the responses to train their own AI.

It’s like recording every lecture a Harvard professor ever gave and using it to train a knockoff professor who teaches the same thing for free.

And the economics are staggering: distillation can cut inference costs by 75%–90% while preserving over 95% of the original model’s performance.

In other words, you don’t just copy the brain, you make it dramatically cheaper to run.

Which raises an obvious question: why would U.S. frontier labs spend billions on compute only for rivals to replicate the results at a fraction of the cost?

They wouldn’t, which is why this fight is escalating so quickly.

Anthropic says the damage may already be done.

In a bombshell report, Anthropic described what may be the largest industrial-scale model extraction operation, naming DeepSeek, Moonshot AI, and MiniMax as the actors behind it:

24,000 fraudulent accounts created to access Claude at scale.
Over 16M exchanges generated across the three campaigns.
MiniMax was the most prolific: 13M+ exchanges focused on agentic coding.
Moonshot AI ran 3.4M exchanges focused on agentic reasoning and vision.
DeepSeek generated 150K exchanges for reasoning and “censorship-safe” rewrites.

They weren’t just extracting intelligence; they were using Claude to help build censorship tools.

It’s like breaking into someone’s house, stealing their toolbox, and using their own tools to build against them.

Source: @TimesOfAI_

And Anthropic wasn’t the only one watching it happen.

The alarms go off across Silicon Valley

Earlier this month, OpenAI sent a memo to the House Select Committee on China alleging DeepSeek used “adversarial distillation” to freeload on U.S. frontier labs.

Google’s Threat Intelligence Group confirmed distillation attacks targeting Gemini; one campaign fired 100,000+ prompts to steal its reasoning traces.

Three companies. Three confirmations. One very uncomfortable conclusion.

And the accused? Radio silence.

Neither DeepSeek, Moonshot AI, nor MiniMax has issued public rebuttals.

Reporters from CNN, Reuters, and Fox News all got the same response: nothing.

Hard to craft a defense when the core activity, sending bulk API queries, is verifiable.

And it gets worse.

While distillation dominates headlines, former Google engineer Linwei Ding was convicted of seven espionage counts and seven trade secret theft counts for stealing confidential AI technology for China.

Source: @Kredo0

It’s the first AI espionage conviction in U.S. history.

Each count carries up to 15 years in prison.

This man is facing more jail time than most actual spies.

But the line between victim and perpetrator isn’t nearly as clean as it sounds; it’s more like a Spider-Man pointing meme.

HOW TO GET YOUR WEBSITE MENTIONED BY CHATGPT & CLAUDE!

With vibe coding, shipping products has never been easier.

But growth and distribution still work the same way they always have.

That’s exactly what Okara’s SEO & AI Visibility Agent is designed to fix.

Here’s how simple it is:

Go to Okara’s SEO & GEO agent
Enter your website

That’s it.

You’ll daily receive:

A detailed SEO audit
5 high potential recommendations to improve your SEO and GEO rankings
Best ways to get mentioned by top AI platforms like ChatGPT, Claude, Grok etc.

Try out Okara’s SEO & AI Visibility Agent today.

THE BIGGEST BRAIN HEIST IN HISTORY (P2)

This is the part where it gets awkward.

U.S. labs accuse Chinese competitors of “stealing” by training on their outputs.

Those same U.S. labs trained their own models on massive quantities of copyrighted material scraped from the internet without permission.

Books, Reddit posts, academic papers, your aunt’s food blog. All of it.

Anthropic, leading the loudest accusations, settled the Bartz v. Anthropic lawsuit for $1.5B last year.

Source: @ns123abc

The largest copyright settlement in U.S. history.

The reason? They downloaded roughly 500,000 books from pirated databases to train Claude.

A company that trained its AI on half a million pirated books is now upset that someone else trained their AI on its AI.

And the legal kicker: no court has actually ruled that distillation is illegal.

The legal landscape is a complete mess:

Can AI-generated outputs even be copyrighted? Unclear.
Does distillation violate copyright law? Untested, no court has ruled on it.
Does fair use apply? Unknown.

OpenAI is literally arguing fair use in its own New York Times lawsuit while claiming DeepSeek’s similar practices are theft.

You can’t make this stuff up.

So, how are U.S. labs handling the hypocrisy problem? Simple, they’re reframing it.

Instead of calling it a copyright issue (where they’re vulnerable), they’re calling it a national security issue (where they hold the moral high ground).

Anthropic argues that distilled models strip out safety guardrails, meaning “dangerous capabilities can proliferate with many protections stripped out entirely”.

Foreign labs could feed unprotected capabilities into military, intelligence, and surveillance systems.

Convenient framing? Sure, but also not entirely wrong.

The inconvenient truth nobody wants to hear

Here’s the thing that makes the “theft” narrative even more complicated:

Chinese AI labs are also genuinely catching up on their own.

Chinese open-source models accounted for nearly 30% of global usage by late 2025, up from just 1.2% a year earlier.

Read that again. From 1.2% to 30% in one year:

Qwen surpassed Llama in Hugging Face downloads.
DeepSeek’s V3 and R1 matched frontier models at far lower cost.
Kimi K2 uses a 1-trillion-parameter mixture-of-experts design.

And since 2023, every frontier model has come from the United States, with Chinese counterparts typically about seven months behind.

Source: @EpochAIResearch

That lead is now shrinking, accelerated by distillation, engineering advances, and massive state-backed investment.

Even under U.S. export controls limiting access to advanced chips, Chinese labs found workarounds through architectural efficiency, synthetic data, and massive state-backed investment.

In Japan, six of the top 10 domestically developed AI models are built on DeepSeek or Qwen.

DeepSeek’s usage in Africa is two to four times the global average.

Distillation may have sped things up, but the convergence was already underway.

What U.S. labs are doing about it

Obviously, Anthropic, OpenAI, and Google aren’t just sitting there taking it.

All three have deployed defensive measures:

Behavioral fingerprinting: Detects distillation patterns in real time.
Identity verification: Stricter KYC for API access.
Output perturbation: Subtle changes that degrade training use.
Reasoning concealment: Hides the full chain-of-thought.

On the policy side, the Advanced AI Security Readiness Act is moving through Congress, directing the NSA to develop an “AI Security Playbook”.

OpenAI wants a full “ecosystem security” approach, restricting adversary access to U.S. compute, cloud, and payment infrastructure.

Translation: They want to turn the entire American tech stack into a walled garden.

But here’s the uncomfortable truth about all these defenses: it’s a cat-and-mouse game with no finish line.

Researchers have already shown that watermarks can be removed through fine-tuning.

Output perturbation can be filtered, and the attackers adapt faster than the defenders can patch.

It’s whack-a-mole with 24,000 moles.

My verdict

So who’s actually right?

Are Chinese labs systematically extracting capabilities? Almost certainly. 24,000 fake accounts and 16M exchanges aren’t a coincidence.

Is it legally “theft”? Courts haven’t decided.

Do U.S. labs have moral authority to cry foul? The entire foundation model industry was built on copying at scale.

It’s a glasshouse of their own construction.

But here’s what actually matters:

The competitive value of frontier AI has shifted from raw model weights to the reinforcement learning, safety alignment, and agentic capabilities layered on top.

The better you get, the more people want to copy you, and the harder it is to protect when your product is accessible via API.

This creates a structural problem that no amount of fingerprinting can fully solve.

The AI race just entered its Cold War phase.

Except instead of nuclear secrets, we’re fighting over reasoning traces.

And instead of spies in trench coats, we’ve got proxy networks with 20,000 accounts.

Robert Fortune would be proud.

That’s it for today.

WHY YOUR FIRST HUMANOID ROBOT WILL BE AT WORK 🤖

In last Monday’s episode, we sat down with Jan Liphart, founder and CEO of Open Mind, to talk about why humanoid robots will roll out vertical-by-vertical, and why early wins look more like front-desk and triage than household chores.

Here's what you will hear:

Why the first real deployments are human-facing roles like intake, triage, reception, and event assistance that need language, memory, and workflow routing.
How Open Mind aims to make different Chinese-built robot bodies behave similarly via a common software layer, with a focus on privacy, data routing, and compliance.
Where humanoids show up first, hospitals, hotels and casinos, conferences, and car dealerships, plus why folding shirts and mopping will likely lag behind.
What could shape the market, China’s manufacturing concentration, vehicle-scale supply chains, and machine-to-machine payments like USDC for charging and identity.

Hit play and see for yourself 👇️