Here is the cleaned YouTube transcript, processed line by line and formatted into paragraphs with headings:
AI in Warfare: The Pentagon, Anthropic, and OpenAI
The Wall Street Journal reports that the Pentagon used Anthropic's AI model Claude as part of its operation to capture Venezuelan President Nicholas Maduro. For the first time ever, the Pentagon is using artificial intelligence to scope targets. Our allies are going to be studying how we're using it here. Data centers were just hit in the UAE.
Hi, welcome to another episode of Cold Fusion. By now, there's no hiding it: we live in a world where AI's use in warfare is now a reality. AI was used in at least two high-profile instances by the US military: one to capture the Venezuelan President Maduro, and then in a war between Israel, Iran, and the United States. The AI company that supplied the underlying technology to the US military was Anthropic, but tensions began to grow between the two parties. The government wanted them to sign a contract that allowed for autonomous AI weapon control and mass surveillance of US citizens through data collection. While Anthropic did say no to this, they were completely kicked out of any government contracts.
All government agencies were told to cut off ties with the AI company Anthropic after it refused to agree to the Pentagon's demands on how its tech is used. That news was shocking enough, but afterwards, something pretty astonishing happened. In one of the worst cases of corporate optics in history, just a couple of hours after the fallout, Sam Altman said that his company, OpenAI, the makers of ChatGPT, would step in to do the deal with the US military.
So, the fall from grace is complete for OpenAI. They went from a nonprofit company that was meant for the good of humanity with goals like curing cancer, to a for-profit company that needed investment by any means necessary, to removing the word "safety" from their mission statement and partnering with the military. The result was an understandable, very public, and very loud boycott of their star product, ChatGPT, with hundreds of thousands and possibly millions unsubscribing from their service.
So today, in this massive story, we uncover how AI found its place in warfare only three years after going mainstream. How does AI work in warfare? And most importantly, what does this all mean for us? You are watching Cold Fusion TV.
The Military vs. Consumer AI
So here's a question: If today we can't trust any of the information ChatGPT generates and it doesn't understand how car washes work, how on earth could it be reliable for identifying military targets? Well, as it turns out, the AI that the public gets isn't the same one that the military uses. You can't say it's surprising, though. We know this through an interview with Anthropic CEO Dario Amodei.
"We were the first company to make custom models for national security purposes." So, the main differences were both the model itself and the hardware it ran on. Anthropic created a custom model for the military and ran it directly in a data center full of classified information. Anthropic has also recently figured out how to look inside the black box of neural networks to understand and control ethics and behavior internally and mathematically, which is where scientists in this paper did their magic. Get this: they found the specific geometric direction in the model's brain that represents the assistant persona. It just puts a speed limit on the change of personality.
Furthermore, the model is running on dedicated hardware. In this case, 100% of the available compute is there for a single customer: the military. In many cases, it could be a whole data center just for one request. In contrast, as civilians, we're used to using AI systems where capacity is split across hundreds of millions of users. In short, the military setup could offer five or six orders of magnitude more performance per request versus what consumer AI has.
So, to be clear, a concentrated competing cluster and specialized data makes military AI way better than what the public has. But even then, it's still not perfect. As Paul Scharre, executive vice president at the Center for New American Security, states, "AI gets it wrong. We still need humans to check the output of generative AI when the stakes are life and death." Even Dario himself says the take is still not there yet. "I think we are a good judge of what our models can do reliably and what they cannot do reliably. Fully autonomous weapons there. I actually am concerned that we may need to keep up. It's not, you know, the technology is not ready." If OpenAI can manage to match the performance of the custom Claude model still remains to be seen.
AI in Action: The Maven System
If any of these AI systems get a target wrong, it's a huge tragedy. So there's a real risk here. But as I mentioned earlier, despite the risks, the AI systems have already been used in warfare. So, we have to ask ourselves, how exactly does AI work in war?
Striking 1,000 targets in 24 hours in Iran was only possible with a dedicated AI system. That system is called the Maven Smart system. It was built by none other than Palantir and is powered by a custom version of Anthropic's Claude AI. The system gathers vast amounts of classified data from satellites, surveillance, and other intelligence. For example, the Israeli government hacked nearly all traffic cameras in Tehran so they could monitor who was coming and going from where and when. The system pulled in information from 179 sources of data in total. This sea of data is then processed through Claude, which outputs insights after interpreting that data. In the opening days of the war, the system organized and prioritized Iranian targets for military operations in real time. It "issued precise location coordinates and prioritized those targets according to importance." And that was according to people familiar with the campaign. According to a study by Georgetown University, the system allowed one artillery unit to do the work of 2,000 staff with just a team of 20 people.
The pairing of the Maven system and Claude has created a tool that's speeding up the pace of war. According to some sources, it's turning weekslong battle planning into real-time operations. So, the fact of the matter is, consumer AI and military AI are light years apart. The Maven-Claude hybrid system was also integral in Maduro's capture. Allegedly, Anthropic wasn't aware that their software was being used in the raid.
The Anthropic-Pentagon Fallout
That concern turned into a flashpoint that started the beef between Anthropic and the US government. Emil Michaels speaks:
"And the trigger point was after the Maduro raid. One of their execs called Palantir, who we buy ourselves through, and asked them, 'Was our software used in that raid?' which is, by the way, classified information. Anyway, so we're trying to get classified information and implying that if it was used in that raid, that that might violate their terms of service. It raised enough alarm with Palantir, who has a trusted relationship with the department, to tell me and I'm like, 'Holy cow, what if this software went down, some guardrail kicked up, some refusal happened for the next fight like this one, and we left our people at risk?' And I had to go to Secretary Hagerty. Said this would happen. And that was like a 'whoa' moment for the whole leadership at the Pentagon that we're potentially so dependent on a software provider without another alternative that has the right or ability to do to not only shut it off, maybe it's a rogue developer who could poison the model to make it not do what you want or hallucinate purposefully or do or not follow instructions."
The concept of using AI in war doesn't stop with the United States, though. NATO signed a contract with Palantir last year, and they have their own version of Maven. So, we're going to get to OpenAI and how they play into all of this, but first, since we're on the topic, let's understand the fallout between Anthropic and the US government a bit more.
The Ultimatum and the Boycott
The wild story starts back in July of 2025. At this time, Anthropic signed a $200 million contract with the Pentagon through a partnership with Palantir. As the partnership progressed, all was not well behind the scenes. The Pentagon started demanding that Anthropic allow Claude to be used for "all lawful purposes" with no private sector restrictions. Keep in mind, the Patriot Act is legal. Spying on Americans is lawful.
There are possibilities with domestic mass surveillance: government buying of bulk data that has been produced on Americans' locations, personal information, political affiliation to build profiles, and it is now possible to analyze that with AI. The fact that that's legal seems like the judicial interpretation of the Fourth Amendment has not caught up. No one wants to be spied on by the US government.
In response, Anthropic wanted two safeguards to be written into the contract: one, no mass surveillance of Americans, and two, no fully autonomous weapons making lethal decisions without a human in the loop. On February 24th, the US government gave Anthropic an ultimatum: comply by 5:01 p.m. Friday, or face consequences. Dario Amodei said no, and on February 26th, he stated, "We cannot in good conscience acquiesce."
At one point, they sent us language that appeared on the surface to meet our terms, but it had all kinds of language like "if the Pentagon deems it appropriate" or "to do anything in line with laws." So it didn't actually concede in any meaningful way. The US government was furious. Hours before the bombing began in Iran, the US government banned their agencies from further using Anthropic's tools, giving them six months to phase them out. Then they called Anthropic a supply chain risk to the United States, a label that has never been used before for an American company. It was extremely controversial, and a bipartisan group of senators called it out. And with that, the relationship was frayed.
The publication Axios reported on what the surveillance angle of the AI deal would have been. They wanted Claude to analyze data on Americans. That data included geolocation, web browsing data, and personal finance information purchased from data brokers.
OpenAI Steps In and Public Backlash
And now we come to Sam Altman, ever the opportunist. As mentioned earlier, in a baffling move, a couple of hours after the fallout between Anthropic and the US government, Sam Altman said that OpenAI would take the deal that Anthropic rejected. He himself called the deal "sloppy and rushed." Over 450 verified Google and OpenAI employees signed an open letter calling their own leadership to stand with Anthropic. So, remember in that recent episode when I said OpenAI was in financial trouble? This deal with the US military changes that. They don't have to worry about their enterprise market share falling from 50% to 25% in two years. Now OpenAI has infinite money from the government.
Sam Altman announced on X that OpenAI's deal included the same safeguards Anthropic had fought for: "Two of our most important safety principles are the prohibitions on domestic mass surveillance and human responsibility for the use of force, including for autonomous weapon systems. The Department of War agrees with these principles, reflects them in law and policy, and we put them into our agreement." As the New York Times reported, OpenAI and the government had only been in discussion for about two days. With such a rushed deal, it becomes hard to believe what Sam is saying. As CNN put it, "It's not clear what is different about OpenAI's deal with the Pentagon versus what Anthropic wanted."
The public backlash has been severe. It mainly came in the form of the QuitGPT movement. Some estimates state that millions of users have stopped using OpenAI services. A recent QuitGPT Instagram post has more than 36 million views, and its organizers say that more than 2.5 million people have taken action, either by canceling subscriptions, sharing boycott messages on social media, or signing up via quitg.org. The discontent spilled out into the real world, where a protest took place in San Francisco. GPT has got to go.
The Future of AI and Surveillance
So, this story isn't over. At the time of writing, Anthropic is back in talks with the US military. This is probably an effort for the company to save themselves from being designated as a supply chain risk. Many are holding their breath, hoping that Dario doesn't go back on his word. In the end, there may be no good guy in this story. We just have to watch this space closely.
It's not controversial to say that using AI in war and surveillance is probably the worst possible abuse of the technology. With all the data about us online, a future government AI running purely on massive government hardware could have no problem analyzing a population in real time. Using AI as a surveillance tool reads exactly like a fiction novel with totalitarian technofeudalism themes. All that's missing is for the government to know exactly who everyone was online, like a digital ID or something like that. Good thing that's never happening.
But it's not all bad news. There are some practical things that you can do to protect yourself from this future. This is not a sponsor, but the number one thing I'd say to do is honestly get yourself off data broker sites. There are a bunch of ways to do it, so just Google it. Data collected by private firms, having it bought by the government, and analyzing it in mass via AI, that actually isn't illegal. It was just never useful before the era of AI.
Remember, the US military surveillance contract wanted to analyze geolocation, web browsing, and personal financial information purchased from data brokers. If you're not on the data broker networks in the first place, your risk is reduced. But next, it's clear that we need laws. Sign petitions, talk to people, get the story out there that we don't want this. I'm not even an American, and it's clear that this isn't a good direction for that country or the rest of the world. And this kind of thing is already spreading.
Even in Australia, many of us don't know that Palantir has installed surveillance systems in coal supermarkets. It's not hard to imagine AI being tacked onto that system and then eventually, in five years, being rolled out wherever public cameras are. If the Meta Ray-Ban glasses scandal is anything to go by, those cameras don't even have to be public for us to be monitored. Anything from going to the toilet to undressing is fair game to train Meta's AI. But even if you don't own the glasses, there are plans to use them to identify you in public. Five years ago, Facebook shut down its facial recognition system for tagging people in photos. Now, according to the New York Times, they're bringing it back inside Ray-Ban glasses. The feature is internally called "Name Tag." A wearer looks at you, and Meta's AI identifies who you are. You don't have to own the glasses; anyone with a public Meta profile could potentially be identified.
And in 2024, two Harvard students already proved the glasses can identify complete strangers using third-party software. Meta sold over 7 million pairs last year. An internal Meta document obtained by the Times says the current political environment is good timing because civil society groups would have their resources focused on other concerns.
Growing up as a millennial, I'm unironically starting to think that the iPod was the last mainstream technology product that purely served the user. I'm half joking, but sometimes we have to look at the direction we're going and ask ourselves, who is this really serving? AI was supposed to make our lives easier and more convenient, but it speedran its way straight into the worst use case: war and mass surveillance. Now, in an individual case, AI can make life more convenient. And if you found a way to use AI sensibly to make yourself more productive, great. But we have to look at the big picture here. On aggregate, in the wrong hands, this technology can be catastrophic.
And in short, that's why I decided to make this video. It's a bit spicy, but I think I just needed to speak up and spread awareness on what's coming. It's fundamental for that to happen first before any widespread action can begin. Anyway, so what are your thoughts about all of this? Were you aware of all of this stuff going on, or was this news to you? That's about it from me. My name is Dogo, and you've been watching Cold Fusion, and I'll catch you again soon for the next episode, something more light-hearted. All right, cheers, guys. Have a good one.