Anthropic Exposes ITSELF as Duplicitous and Desperate
Anthropic’s Claude, Opus, and Sonnet models face allegations of Chinese companies using 24,000 flagged accounts for training, with claims dismissed as fearmongering tied to DeepSeek’s upcoming V4 (83 SWE benchmark)—rumored to rival Opus 4.6’s coding prowess at a fraction of the cost. The speaker, citing Anthropic’s own monitoring and privacy contradictions, argues U.S. firms’ reliance on public data scraping undermines their accusations while China’s open-source AI and STEM dominance (e.g., EVs, robotics) outpaces them. DeepSeek’s potential release could disrupt Anthropic’s revenue, forcing a shift to cheaper alternatives like DeepSeek V4, despite risks like hallucination delays, while U.S. firms prioritize secrecy over innovation. The episode ends with a call for better models—not baseless claims—and plugs the speaker’s AI platforms before pivoting to beet juice ads. [Automatically generated summary]
All right, so Anthropic, which is currently the leader in AI coding, is accusing Chinese AI companies, including DeepSeek, of wholesale theft of their model knowledge through a process called distillation,
which means that what they're saying is that there were tens of thousands of user accounts that were actually proxies for Chinese companies that asked questions of Anthropic's models, you know, Claude or Opus or Sonnet, whatever, and then took the answers and then used those answers to help train their own models.
And this is being called a form of wholesale theft or what, corporate espionage even.
That's an exaggeration.
Well, first of all, Dario Amodi is known to be an alarmist about everything.
He's the one always screaming about how dangerous AI is.
He's the one screaming about cyber threats.
He's the one always screaming.
And now, I mean, pointing the finger at DeepSeek, probably because DeepSeek is about to release version 4, which is rumored to be just as good as Claude Opus at coding.
And if that's the case, then it undercuts completely most of the revenue model of Anthropic.
We could see Anthropic's stock price just absolutely plummet.
In fact, once DeepSeek version 4 is released, we could see the stock prices of many tech companies plummet.
Similar to what happened roughly a year ago when DeepSeek released R1, their first reasoning model.
This version 4 is rumored to be, well, it could score 83 on the SWE benchmark, which is the coding benchmark.
And that would be the highest score ever of any open source model.
And it would be neck and neck with Opus 4.6, which is the latest model from Anthropic.
And notably, Opus 4.6 is the model that I currently use for most of my coding tasks.
And it's a very capable model, no question about it.
That's why I use it.
But if DeepSeek version 4 is just as good and 1 50th the cost, well, of course, I'm going to switch over to DeepSeek version 4.
I'm going to use the best coding model that I can use, whether it's free or expensive.
Either way, it's still worth it to me, you know, because of the AI platforms that I'm building and maintaining and rolling out new features, etc.
And any model that's good at coding will save me time because I spend a lot of time vibe coding.
And often I'm hammering the same issue, trying to get the code to find the bug or fix the bug or refactor whatever.
Opus 4.6 does a very good job, but if there's something better, I'm going to switch.
No question.
But the interesting pattern in all of this is that US tech companies and some US analysts and people tied to the CIA and the State Department, et cetera, they always like to point the finger at China and say, well, all their technology is stolen.
Or they smuggled illegal graphics cards or GPUs into their country.
It's illegal microchips.
Okay.
They do everything and anything to avoid admitting that China has amazing breakthrough technology and that the Chinese people are very good at math and engineering and coding and innovation in technical spaces.
And so, you know, the overall strategy from the U.S. is to try to say that China can't innovate, that China has to steal it from the U.S., that we, the United States, we're the only people in the world who can invent things and create things and we're the best.
That's the narrative.
And it, frankly, it sounds pathetic.
It's like, really?
All you can do, because you know how good DeepSeek 4 is going to be, all you can do is just push a bunch of uncertainty, say, oh, they stole our tech.
Really?
That's it?
And that's really rich coming from Anthropic also, and frankly, any other U.S. company, because they trained their models on all kinds of content, including just scraping the entire internet, you know, scraping books and periodicals and newspapers and everything else to train their models, right?
So for them to say, well, China, you know, China signed up and paid the API fees, but we don't like how they used it.
And we're going to call it theft.
Well, it's not theft.
They didn't hack your system.
They paid for the API.
They paid for it.
And you know what else is a little shocking in all of this is one of the reasons there's so much backlash against Anthropic right now is because Anthropic's announcement on all this admitted that they spy on all their accounts.
They're spying on everybody because they identified, I think, 24,000 or some number of accounts in that range that they say were involved in these distillation attacks.
Well, wait a second.
So you're watching what everybody does?
I mean, you're spying on everybody and yet your message is that everyone should trust you, trust you with their privacy when clearly you're monitoring and spying and collecting metadata, etc.
So, you know, Anthropic is really undermining its own case here.
Instead of really working on creating a better product and staying in the lead, Anthropic is whining and crying about China and admitting that Anthropic spies on you.
So no wonder there's such a backlash against Anthropic right now.
People are waking up.
By the way, with DeepSeek version 4, you know, you can download it once it's available.
You can download it and you can install it locally.
You can run it locally.
You need some pretty serious hardware to do that.
But once you have that done, then there's no way DeepSeek can spy on you because you're running it locally.
Whereas Anthropic doesn't have a local version.
Anthropic hasn't put out anything that's open source.
Anthropic's not contributing to the open source community by releasing open weights models.
Frankly, neither is any U.S. company or almost any company at this point.
They've stopped doing it.
The U.S. AI tech giants have stopped releasing models to the public and they've stopped publishing science papers in the public science journals.
That's why all the best science now is coming out of China because they're publishing.
They're publishing the vast majority of all the papers.
And they're innovating.
So DeepSeek, for example, is run by a group of really intelligent people who author numerous science papers.
For example, papers on sparse attention technology and manifold constrained, what is it, hyper-connections?
It talks about the, well, the mathematical structure of the different layers of model of the model during training and how the different layers talk to each other, pass information in order to maintain information integrity by adhering to a similar mathematical structure.
That's a breakthrough.
That's a breakthrough in the signal-to-noise ratio of these models.
And that breakthrough was used to train DeepSeek version 4.
And all the companies from the U.S. are looking at this and they've read the science papers and they know that DeepSeek is leading the pack now.
That we're no longer in a game where the Chinese models are almost as good, but six months later.
No.
DeepSeek version 4, the way it looks, although we still have to confirm it upon release, but it looks like this is going to be the leading frontier model for the world.
That this will become the standard model for most AI work, unless for some reason somebody, maybe, you know, like a government contractor or somebody has to use a U.S. model, is forced to use OpenAI or Microsoft or Gemini or something.
But for most corporate work and nonprofit work and, you know, whatever, people are going to use DeepSeek version 4 because it will be right there at the top of the list of the best capabilities and it's free, which means even if you're using it through, let's say, a cloud hosting provider for API inference, it's going to be dirt cheap.
Dirt cheap.
By some estimates, it's something like 1 50th the cost of maybe Chat GPT or Google Gemini.
So that's a big deal, especially as so many projects now are burning tokens and you have recursive loops, recursive reasoning, and all these open claw lobsters running around burning tokens all day.
They need cheap tokens and DeepSeek's going to provide that.
China Out-Engineered Us00:03:07
So no wonder, no wonder everybody in the West that's tied to the AI industry is pointing their fingers at China and saying they stole our stuff.
They stole our tech.
And also this claim, oh, they use illegal microchips.
You mean they bought the microchips that you didn't want them to buy?
Oh, they smuggled them.
Oh, you mean they bought them and you didn't want them to buy them, but they found clever ways to buy them.
They didn't actually steal microchips.
They paid for them.
And they just went around your, you know, your blacklist.
So no, China, they didn't steal microchips.
They didn't steal anthropic code.
No, there's been no hacks.
China paid for the APIs, paid for the GPUs, and set up their systems and went to work.
And they just created better products because they have more smart engineers.
I mean, see, I don't know why it's impossible for so many American companies to admit this, that we've been out-engineered by China.
We've just been out-engineered.
Why?
Because China graduates 500% more STEM graduates every year.
And Chinese people, and remember, I speak Chinese.
Know a lot about Chinese people compared to most Americans.
That is, I mean, I lived in Taiwan, and Chinese people are freaking smart, okay?
That's why in the California university system, they have a penalty if you're Chinese.
They penalize you off your scores because otherwise, 80% of the students would be Chinese or Asian, let's say.
I mean, Korean, Japanese, Chinese, Taiwanese, Vietnamese, Thai, etc.
Asian people are naturally very smart people.
And even more than that, the Chinese education system is absolutely-I mean, it's brutal, it's brutal on the kids, but man, it turns out high-level engineers, mathematicians, data scientists, machine learning scientists, and the sheer numbers just dwarf the United States.
But no one in the U.S. can admit this.
Oh, they're just, they're out-engineering us.
Yeah, well, they've done it in EVs, they've done it in robots, they've done it in drones, they've done it in naval dockyard construction, automation, car factories.
Yeah, they've out-engineered the U.S. in like 90% of all technologies.
So, why is it so hard to believe that they out-engineered the U.S. when it comes to AI?
They're just freaking good engineers.
Okay, I mean, we need to get used to that idea.
They're just good.
You know, anyway, it's pathetic for Anthropic and these U.S. companies always sit around pointing the fingers like they stole from us, making a bunch of excuses.
Why They Out-Engineered Us00:02:24
Yeah, you know what?
Excuses aren't going to fly.
What's going to matter is who's got the better models, which models can code better, which models create better images, better text-to-speech.
And, you know, right now, Quinn, Quinn, what was it, version 3?
Is it 3.5 text-to-speech engine?
It's amazing.
I'm using it.
We're going to use that for our audiobooks.
It's amazing.
China is just out-engineering the whole world right now.
And I don't see any reason why that won't remain the case.
You know, they're just freaking good.
All right.
So, anyway, look, I'll keep you posted.
Get ready for DeepSeek version 4.
I don't know when it's going to come out.
It's already kind of late, actually.
So maybe they're having a problem with it.
Maybe it started hallucinating like crazy.
Who knows?
You know, AI is tricky.
Or maybe they're just waiting for some final thing.
Maybe they're ramping up their data centers because they realize everybody in the world is going to want to use this model.
So who knows?
But whatever it comes out, watch for tech stocks to collapse in the U.S. and watch for Anthropic to probably get hammered.
And then the very next step is we have to watch the benchmarks.
We want to see how DeepSeek version 4 actually performs in independent benchmarks, especially the SWE bench, which is all about coding.
So I will keep you posted.
And obviously, this is of high interest to me.
I'll keep you posted.
And I'm rooting for DeepSeek on this.
We need a free, open-source, decentralized model that beats Anthropic.
Because we're all tired of Dario whining all the time, frankly.
Stop whining and start building.
That's my advice.
But thank you for listening, everybody.
I'm Mike Adams.
And if you want to use my AI platforms, just check out brightanswers.ai, brightlearn.ai, and you can catch my videos at brightvideos.com.
Take care.
Beet juice is a refreshing beverage that can help you stay energized.
You can easily incorporate beet juice into your daily routine with Groovy B organic freeze-dried beet juice powder.