Anthropic Said No to the Pentagon. Now It’s Blacklisted

Hey, welcome to the Next Wave podcast. As with pretty much every week, we had another insane week in the world of AI. We have an AI company going to war against the US government, essentially. We have a brand new image model which we're going to fully demo and test and play with and really put it through the motions. And we've got AI agents everywhere. This week we're going to break down some of the new AI agents that came out and we'll talk about Burger King at the end. Cause Joe used to work at Burger King, so why not talk about Burger King? Anyway, let's get into it. Joe,

cutting your sales cycle in half sounds pretty impossible, but that's exactly what Sandler Training did with HubSpot. They used Breeze, HubSpot's AI tools to tailor every customer interaction without losing their personal touch. And the results were incredible. Click through rates jumped 25%, qualified leads quadrupled, and people spent three times longer on their landing pages. Go to HubSpot.com to see how Breeze can help your business grow.

Thanks for joining us again. What do you got for us today?

Thank you for having me. Yes, this has been a wild week, man. Obviously you know that. But I think the big glaring topic of them all, which has gone pretty public now, I feel like people have finally waken up to what's happening with this whole, like, Pentagon standoff with Anthropic, you know, the maker of Claude and I mean, it's actively happening still, but I feel like we kind of know where Anthropic stands at least. And yep, has some interesting consequences to all of this as well. Some pretty big ones. So, Matt, how would you kick this off? Because this is. It's kind of shocking, to be honest. It's kind of freaking me out.

Yeah, no, this one is a pretty crazy story, I would say, because it has such big sort of global consequences. Right? Like, to me, it's a big deal and I don't feel like enough people are talking about it. I actually made a video about this over two weeks ago. By the time this podcast is live, I made a video about this two weeks ago. Nobody was talking about this. It was still flying under the radar. And I was like making a video trying to go, hey, everybody needs me to pay attention to this. Like, screw your productivity apps and stuff. Like, this is what's important right now. But here's the sort of TLDR for people that aren't totally looped in on what's going on right now. So you've got anthropic Anthropic is the only company up until recently was the only company that had access to classified information within the US Government. They were the only model that the US government trusted enough with their own classified information. Now, Anthropic also recently signed a $200 million deal that had tie ins with the Pentagon and Palantir, who, you know, famously makes military equipment and things like that. And so they had this $200 million deal. And what really sort of set things off was, remember when the US went into Venezuela and basically pulled out Maduro, their president, and brought him back to the U.S. well, it turns out that Anthropic's AI was used as part of that mission. Somebody at Anthropic contacted the Pentagon or, you know, some US official and said, hey, is it true that you used our model to do this raid on Maduro? And the US government, the Pentagon, whoever they were talking to, was very unhappy with the fact that Anthropic was even asking how you're using our AI. That was the sort of thing that set off this chain of events is why is Anthropic asking how we used this stuff? Right. And so that's when the Pentagon started speaking up and saying things like, oh, Anthropics AI is woke and we need to have AI that will do anything we need it to do with no restrictions. If we want it to do something, it needs to be able to do that. And it just sort of blew up into this whole thing. Right. Well, Anthropic then went and said, look, we'll actually ease up our restrictions for military use. You can use it for pretty much anything you want to use it for, except for two things. There's two things that we're drawing a line on that you cannot use our AI for. Number one, you can't use it for any sort of mass surveillance of U.S. citizens. We will not allow it to be used for that. Number two, we will not allow it to be used for autonomous weapons with no human in the loop. You can use it for semi autonomous weapons, but there needs to be a human making the decision of when the trigger gets pulled or not. Right? Those are the only two things that Anthropic said that we are going to hold the line on US Surveillance and you must have a human in the loop for autonomous weapons. And basically Pete headseth over at the Department of War said, no, you don't get to tell us what we can and can't use it for. And that's kind of like the whole story, right?

Yeah, it's Freaky.

Yeah. It's kind of crazy that when you think about it, those are the two things that Anthropic is holding the line on. And the US Government is saying, that's unacceptable to us. Like, that, to me, is a little bit scary that the US government is willing to say, like, no, if we want to use it for surveillance of US Citizens, we can do that. If we want to use it for fully autonomous weapons, we can do that. That's our choice, not yours. Hey, if you take a look at my web presence online, it's safe to say that I'm a bit AI obsessed. I even have a podcast all about AI that you're watching right now. I've gone down multiple rabbit holes with AI and done countless hours of research on the newest AI tools every single week. Well, I've done it again, and I just dropped my list of my favorite AI tools. I've done all the research on what's been working for me, my favorite use cases, and more. So if you want to steal my favorite tools and use them for yourself now, you can. You can get it at the link in the description below. Now back to the show. There's a big debate going on right now of who should make the decisions when it comes to military operations. Should we be letting the CEOs of Silicon Valley decide when their technology should be used in military operations or. Or should we be letting the government and the military themselves be deciding when to use this technology for military operations and what military operations it can be used for? And there's a pretty big debate. Palmer Luckey, the CEO of Anduril and the CEO of Palantir, have both spoke it up and said CEOs need to, like, take a backseat. You need to stop telling the government how you can use technology. Right. But then there's also, like, a whole public uprising right now of people saying we don't want our government to be using it to surveil us. We don't want our government to be using it for fully autonomous weapons. Like, that's just scary to us. I'm so glad Anthropic is holding the line here. We don't want that to happen. So there's this whole, like, back and forth going on right now between the

two, and the deadline is. Yeah, it's basically looming at the time of when we're talking about some of these details.

The deadline's today to pull back the curtain. We're recording this on Friday. The deadline is Friday.

And after that, wasn't it, like, if Anthropic doesn't agree they're going to be placed on the Defense Production Act.

Yeah. So no, what it is, is the Pentagon has threatened to declare Anthropic a supply chain risk, not only voiding its contracts, but forcing other companies that work with the Pentagon to certify they aren't using Claude. So if they become a supply chain risk, this is something that's never been done to a US based company ever in history. Right. So this would be the first time ever that an American company would be classified as a supply chain risk. It's something that's usually only for foreign adversaries, Right. Like Chinese companies and Russian companies that they worry might be used to like spy on American citizens or whatever. So this would be the first time that would ever happen. And what it means to Anthropic is not only will the government not use Anthropic, but any companies that directly work with the government have to certify that they're also not working with Anthropic.

Right.

And so this article came out from Axios here that basically said the US Government is already starting that process. They already contacted Boeing, they already contacted Lockheed Martin and said, like, hey, are you using Anthropic? What's your exposure to Anthropic? Because if you are, you need to start backing off because this is probably going to happen. So the government is already in the process of starting to blacklist Anthropic because they're about to potentially designate them as a supply chain risk.

And you know, they're not going to skip a beat because the deadline's looming Friday. They're already reaching out to companies. Blacklisting is probably going to start right away, I would imagine. Daria, the CEO of Anthropic has already basically said they're not going to back down to this.

So yeah, they put out a statement yesterday, Thursday. So even before the Friday deadline they put out a statement and you know, he reiterated, like these are the two things we're drawing the lie on, mass domestic surveillance, fully autonomous weapons. And he plainly says, regardless, these threats do not change our position. We cannot in good conscience accede to their request. So they basically have already come out and said, look, we're not going to pull away on this. Like we've already made our decision, we're not going to allow you to do those two things. So they would be designated as a supply chain risk, but they've also been threatened to invoke the Defense Production act, which would basically force Anthropic, like eminent domain, kind of thing force Anthropic to remove those safeguards. So you can tell us you're not going to remove them, but if we invoke this act, well then you're going to remove them and there's nothing you can do about it.

That's right.

Right. But then, you know, Anthropic says these latitude threats are inherently contradictory. One labels us as a security risk, the other labels Claude as essential to national security. So it's like you're getting mixed signals here. You either need Claude or you're going to make it a supply chain risk. Saying that like it's a security risk for anybody to use Claude. So they're just kind of like throwing all the threats at the wall to hopefully one of them sticks and Anthropic gets on board.

Yeah. Oh man. Well, and what I've read somewhere, I don't know if it's in this article that you're showing or not, is that the other AI, you know, LLMs, all the labs out there, you know, OpenAI, Google XAI have already started to talk to the government and it sounds like they're open and willing from what I understand.

Yeah, you know, that's, I think, more rumor than actual fact right now. So OpenAI and Google already work with the US government. Like they're already baked into some of their systems. However, they do not have access to the same classified information that Anthropic has. So Anthropic was the only company that basically was able to work with any sort of classified government information that, you know, dealt with military strikes and things like that that obviously could not preemptively get out to the nations they were about to do the strikes on. Right. Like had an Anthropic insider somehow leaked to Maduro that this raid was about to happen? Well, the raid would have probably been a failure. Right. So like they can't have leaks coming from some of these AI companies. So these AI companies are already working with the government, but Anthropic was the only one that was getting access to classified information because it was the only one they trusted enough. Saying that there has been confirmation that Elon Musk went, oh, you want to like spy on civilians and use autonomous weapons? You can use ours for that. We don't care, like come use us.

I did say I saw that. Yeah. And I mean if you go into X right now, you just see all this. Everything that Elon's posting right now is basically regarding this. Yeah. You know, and it makes sense. But yeah, it sounds like Grok has Already into all the classified documents.

I don't know if they're in there yet. But they did sign a deal. XAI has signed an agreement to allow the military to use its model Grok in classified systems. Defense official confirmed with Axios. So they did sign a deal allowing Grok to work with the classified stuff. Right. So that is, you know, in the process. So it sounds like what could end up happening is they'll try to offboard Anthropic onboard Grok saying all that Anthropic's the superior model and the government knows that. The government has flat out said multiple times, like, we want to use anthropic. It's going to be more of a pain in the ass to not use Anthropic than it would be to, like, stick with Anthropic. So we want to figure out how to, like, keep using Anthropic. I almost feel like what they'll do first is almost invoke that act that forces them to remove the guardrails. Anthropic will continue to hold their ground. It might go to court. It might turn into this long, like, legal battle, and then maybe that whole supply chain risk thing will come into play. I don't know how it's going to play out. In fact, by the time this episode's live, it might have even played out even more. But what we do know right now is the Pentagon does really want to use Anthropic's model, and I think they're a little less excited about Grok's model because it's not quite as powerful.

Maybe I'm just naive here, but I'm curious of your take since, like, we've been chatting in the previous weeks of how the benchmarking, it's getting pretty close across the board. But, like, is there a clear reason why Claude over something like grok or even OpenAI or Google makes more sense for the government?

You know, I don't know. Like, the government might have their own internal benchmarks that they're working off of that we don't see. That's a very likely probability. But what I can tell you is that Opus 4.6, the publicly available model, is pretty much considered, like, one of the best coding models. You do have GPT 5.3 Codex, which is probably on par. Most people are kind of using those interchangeably now because they're both as good as each other when it comes to coding specifically. But when it comes to, like, military use cases, I'm sure the military sort of tested different models and Decided Anthropic gives us the outcome that we're looking for more often than these other models. But I don't know, like, however they decided the models is probably fairly classified information, honestly. Otherwise, like, what stops OpenAI from going and sort of training on that specific information that they know that the government is looking for to sort of win favor with the government? Like, so that I would imagine would be kind of not something that they would openly be reporting of. Like, this is what we're looking for in our models.

Man, it's just wild because. Yeah, like you said back to what Dario from Anthropic was saying, it's like the two things, human in the loop and straight up surveillance, essentially.

Yeah. And you know, he gives very specific reasons why these are the two. Right. So when it comes to mass domestic surveillance, he says under current laws, the government can purchase detailed records of Americans movements, web browsing and associations from public resources without obtaining a warrant. A practice the intelligence community has acknowledged raises privacy concerns and that has generated bipartisan opposition in Congress. Powerful AI makes it possible to assemble this scattered, individually innocuous data into a comprehensive picture of any person's life automatically and at massive scale. So he's saying right now the government is already allowed to sort of collect publicly available data from social media and browsing history and whatever they want. Right. However, that information is kind of pointless, kind of useless for the government to have because it's just like this massive corpus of data that they can't make any sense of. Well, when you're able to take all of that massive corpus of data and put it into an AI model, now all of a sudden the AI can make sense of that data for you. It can find connections, it can find people who potentially posted things anonymously online and connect them to real people. And this is a big concern for people right now because we saw what was happening with like ICE and all of that kind of stuff, where the US government, you know, depending on your take, is sort of infringing on people's civil liberties right now. So you do have like a lot of fear of like, okay, now they can like pinpoint almost any comment online to anybody in the world, even when that comment was meant to be anonymous and like, hey, anthropic doesn't want to be a part of that. Like, look, if you're going to go and do that, if the government's going to go and do that, like, okay, go and do that. But we don't want you to use our technology to do that.

Yeah.

And Then the fully autonomous weapons. So they actually admit that, like partially autonomous weapons like those used today in Ukraine are vital to defense of democracy. Even fully autonomous weapons, those that take humans out of the loop entirely and automate, select and engaging targets, may prove critical for our national defense. But today, frontier AI systems are simply not reliable enough to power fully autonomous weapons. We will not knowingly provide a product that puts America's war fighters and civilians at risk. We have offered to work directly with Department of War on R and D to improve the reliability of those systems, but they have not accepted this offer. I mean, the Department of War, Department of Defense, Homeland Security, whatever you want to call them, they basically are saying, when we are using our AI systems, we need to know they're going to work. There can't be guardrails. We, when there's a life and death moment, that needs to be decided at the blink of an eye, and we need our AI to work right then. Right. That's kind of their argument. What Dario's argument is, is we agree with you. We don't want our systems to F up when you need them the most. So, like, they're both kind of saying the same thing, but they're coming at it from different angles. Like the government is saying, life and death decisions, we need our AIs to not have guardrails because we need our AIs to do what we need them to do. And Dario saying, yes, in life and death decisions, you need the AIs to do exactly what you want them to do in that moment. We're not confident our AI can do that for you in the moment. So, like, let's work together on R and D to get them there. They're not there yet.

Yeah. So my brain just goes to, like, hallucinations. Like, you know, it hallucinates on the simplest things for the longest time. I don't know if it's already solved the whole Strawberry spelling predicament, you know, with the number of Rs in Strawberry. Yeah, yeah.

I think that's a pretty solved problem at this point.

But it wasn't that long ago.

There is this.

So this is what happens when guardrails become, I guess, non existent. Yeah. What are we looking at? Because this, when I saw this, this was essentially what war game scenarios that were done at what was at King's College in London. And they did something like 21 different war game simulations with 329 different turns, and basically it did not end up. Well, 95% of the time, the leading

AIs from OpenAI, Anthropic and Google opted to use nuclear weapons in simulated war games in 95% of cases.

And that was across multiple models. What, three or four different models that they played these scenarios through?

Yeah, they always end up going to nuclear war as the best option at some point in these simulations. So, like, point being, Anthropic is saying these models are not ready to be put in fully autonomous, like, battle scenarios yet. And I feel like this right here, what we're looking at, is probably good sort of evidence of that.

Yeah, it's very timely with this article coming out or this study. Yeah. Hallucinations, no guard rails or at least a lack of guardrails, and no humans in the middle. That's a hard pill to swallow there.

Yep. So that's pretty much the story there, and it's still unfolding and it has probably unfolded a little bit more by the time this episode is even live. But as of right now, where it stands is the Pentagon is giving them till the end of Friday, the day we're recording this, to basically get on board with their demands. Otherwise they risk being labeled as either a supply chain risk or basically having the government force them to remove the guardrails.

Right.

And Anthropic is saying, look, we can't comply with these demands. These are our reasons. I feel like they gave pretty compelling reasons in my mind, just, you know, sharing my own opinion on this. I think Anthropic is right. Like, hey, if you don't feel your models are good enough yet for these use cases, let's not put them in these use cases yet, because that is a scary scenario. I feel like it's fairly black and white, but I am seeing a lot of counter arguments online right now. I am seeing a lot of those counter arguments from people saying, Look, Silicon Valley CEOs should not be in charge of the decisions that our military makes. Put those decisions in the hands of military officials.

Right. Man, it's tricky. Overwhelmingly, I've seen a lot of people be like, yeah, I stand with Anthropic or Dario. And I think as a human, I'm with you.

Yeah. But, you know, so one of the other counter arguments too, though, is that like, okay, so Anthropic is not saying yes to doing this. So what are we going to end up with? An inferior model that will say yes to this, that hallucinates even more often, that screws these things up even more often? Like, wouldn't you want the absolute best model that is available in the moment at making these decisions? To be the one that's making these decisions. Because if they say no, then we're getting grok and grok is not nearly as good at actually making these decisions.

Right, right.

So like I can see the counter arguments, the counter arguments also make sense to me.

Yeah, man. I don't know where this goes, Matt. It's, it's like I feel like the pressure of, you know, obviously world events, we're already seeing stuff play out so, you know, the government's going to want to make a solution and make it happen quickly. I am kind of thinking they'll probably do some stuff to keep Anthropic around. If I were to guess and like you said that the act, that, that Defense act, something like that, they'll probably put more pressures. That's where my mind goes. My gut's right now.

Yeah, my gut tells me that this goes into court battles and Supreme Court potentially gets involved and it turns into a long, drawn out battle kind of thing because if they do deem them a supply chain risk and all of these enterprise companies that work with Anthropic but also have contracts with the government all have to stop using Anthropic. That's going to be a massive hit to Anthropic and something that Anthropic is probably willing to fight for, I think.

So I think it'd be a hit to everybody, you know, so it's like just such a massive ripple, but definitely a lot of legal battles. This is going to be happening. It's not going to be over today, I guarantee this. So I don't want the doom and gloom the whole time on this episode.

But yeah, yeah, I think we needed to lead with probably the most consequential one that's basically setting a precedent right now. Right. This is like, however this plays out is setting a precedent for how AI companies and the government interacts with each other. Right. So it'll be interesting to see how it plays out. But I definitely think we should move on to some more light hearted, fun

stuff on that note. So I'll kick it off basically. Nano Banana, if you haven't used nanobanana yet with Google and Gemini, where you. Ben, it's so fun. And the new model just dropped this week, nanobanana2. And I believe it's starting to get integrated in a whole bunch of other areas or at least it will very soon and it could do a lot of cool things. I know there's been some tests, Matt, that you've been running that will. I think we should do Some prompts and stuff and test it out here too. But speed tests, accuracy tests, all that stuff.

So full disclosure, Google did give me early access to Nanobanana 2, so I was using it not too far in advance of launch, but they did give me like, I don't know, 72 hours advanced usage of it so I can play with it and make a video about it. And so here's the gist of it. Nano Banana 2 is about double the speed as Nano Banana Pro. So Nano Banana Pro was the state of the art image model out of Google DeepMind. And Nano Banana 2 is pretty much just as state of the art. It's pretty much as good as Nano Banana Pro, except it generates the images in half the time. It takes roughly 15 seconds to generate an image instead of about 30 seconds to generate an image. It's actually got better text generation than even Nano Banana Pro did. So it tends to screw up the text even less and put typos in the image even less. It's got the sort of search grounding. So if you say, hey, make me an infographic about how quantum physics works. It will actually go out and do research on quantum physics and then come back with the information that it found and then generate an image based on what it found. That was something that was unique to Nanobanana Pro, but you get it in Nanobanana 2. Now Nanobanana 2 is also a cheaper model to use and Google has essentially just made it available for free inside of Gemini. I do think there are usage limits. I don't know exactly when you hit those usage limits, but you can essentially just go to your Gemini account in Google and for free generate Images with Nano Banana 2 up to a certain point. And I don't know where that point is because they didn't announce where that point was in their press releases. And also I'm on a paid plan, so like I have more usage, so I haven't been able to find that limit myself either.

It's really cool though. I mean like the fact that you just said it's. It's almost like you can use it as your research partner and it will do it quickly. As we've been demonstrating all these other coding apps and things, like everything's just like lightning fast speed now, but it's actually super accurate. Like the tests I've done as well, you look at the words and it's doing the research infographics, it's doing it really well. I know the quality is slightly, it's pretty dang close. We'll just say that?

Yeah, I would say I have a hard time telling the difference in quality. Like if you're looking for realism, I felt like nanobanana Pro might get you slightly more realistic, but it's like nitpicky. Right? Like it's really close.

So what do we have?

So in order to use Nanobanana Pro, you can go to gemini.google.com but if you click Create an image, you'll notice you're on the new one because this style picker thing is brand new that they just launched with this new model here, which is actually pretty cool too. I don't know if you've played with it at all, but it basically has like these style references. So you can give it a prompt and then select one of these style references and it'll make an image that's in that same kind of style.

Yeah, yeah, No, I dig it. I mean it's kind of blurring the lines now between other photo apps or you know, all these, all the image generation apps. But it's cool, it gives you ideas like immediately and yeah, you could click one of those. And I know what I was doing is uploading my own headshot, you know, and trying the Technicolor.

Yeah, yeah. I mean it's pretty straightforward. You use it pretty much the same way you used to use any other Nano Banana. You have to make sure you're on the Create an image tab to basically force it to use nanobanana. Now if you don't click Create an image, you can say create an image of X, Y and Z and it will usually switch it to nano Banana just automatically. However, I have had it like not switch to nano Banana and for whatever reason give me like a written prompt back like explaining what a wolf howling at the moon looks like instead of making the image. So like, for best results, click the button to create an image.

Yeah, for sure it's easier, but I mean it's, it's lightning quick. I just did one on my side as well.

Yeah, I mean we could do like a sketch here and then. Yeah, I'll drop a, like a headshot of my face in here. Right. So I've got the headshot and I've got this. I don't even know if I have to give it a prompt.

I mean, I've just been hitting go after that and it just does it. I think that is your preset prompt.

Yeah. So if I hit go, it should make a sketch version of my headshot. And yeah, it should take about 15ish seconds. There we go. But, yeah, I mean, pretty impressive.

But yeah, dude, I mean, that's great quality. I mean, I'm not seeing it like full screen, but I think you can get pretty high res too, right?

Well, they claim 4K. I haven't been able to get a 4K image out of it. I can basically get it to be almost 3,000 pixels wide. Like, I think the biggest sort of width dimension I've gotten it to do is, like, 2700 wide. I don't know how to actually get it 4K, maybe that's only available via the API. So if we wanted to try it, like, in AI Studio, we might actually be able to get a 4K out of it.

That's what I was curious about.

You can only use it on the paid tiers inside of Google AI Studio, so if you want to use it for free, I guess you got to use it inside of Gemini or Flow or one of their other platforms.

Got it.

Okay. Okay. So down here we do have resolution options.

There we go.

So I can set the resolution as 4K, create a detailed infographic showing Petco park and all the surrounding landmarks.

We always have to have a Padre reference.

Yeah, absolutely.

In every episode.

My goal is to get the Padres to, like, invite us to throw out the first pitch or something because we talk about them so much.

All right, what do we got? I'm curious if the speed's different on this.

Maybe 4k takes longer time. I did put thinking level minimums too, so maybe. Yeah, this definitely feels slower.

It does. I love these type of images, though. And you can just ask it to gather all the details. Here we go. Hold up. What do we got?

Yeah, it's actually bigger than 4K. 5632 by 3072.

So that's the way to do it then. All right.

Yeah. So in AI Studio, you can get it to generate 4K. Now, how accurate is it? That's the question. Yeah, not super accurate because Balboa Theater is definitely not right next door to Petco Park.

True.

The Western Metal Building is definitely not on the sort of west ocean side of Petco Park.

No, we know where Petco Park. Well, I actually, I noticed in your video yesterday, it did have, like, the arrangement of San Diego in general completely, like, backwards.

Yeah. And I'm wondering if it's because I did it on thinking fast instead of thinking high.

Okay.

So, yeah, not quite so super accurate with the details there.

So don't trust it out of the box. Maybe the thinking moon would help, though. I don't know, maybe. Can we rerun it with that and just see?

Yeah, let's throw it on thinking high instead of minimal.

I mean, it'd be nice to know. It's like, hey, we want to be able to trust this thing for its research too.

All right, so I'll copy this and then let's do another one here.

See, this is a maybe newbie question. I know some workarounds, but for to get the Gemini watermark off this thing is this like using an API like you're doing now. The only way, if you're doing it

in Gemini, there's going to be a Gemini watermark. If you're doing it directly through the API, I think you have to be on the ultra plan, actually. And I'm on the ultra plan, but I'm not 100% sure. All right, so let me make sure all my settings are the same 4k high thinking level, and I'm going to give it the same exact prompt again. And let's see if it spends a little bit more time. I know it's taking longer because we put thinking level high. And also my guess is that setting it to 4K is also something that makes it take a little bit longer. All right, so let's see how close we got on this higher thinking mode. I can already tell. Still not great. Like, this archway is definitely not right there.

It looks not quite right, but it has way more detail.

Yeah, a lot more detail. A lot more information in each little heading here.

So this still isn't perfect. Yeah, but it's cool. I mean, it brought in dimensions, capacity, the state. I mean, the fact that it's doing all this work, but it did do that in Nana Banana Pro, so I don't know, I'm just curious how much different.

Yeah, it's not great at like, drawing super accurate representation of what this would look like, that's for sure.

Yeah. Well, what are other use cases that, you know, Nana Banana? I'm just thinking like thumbnails. Like YouTube thumbnails.

I've never had amazing luck getting nanobanana to generate YouTube thumbnails that I'm like, really impressed with. But what I do use nanobanana for a lot is to generate items that I'll put into a thumbnail. So I'll get it to generate an image of me surrounded by bananas or something, and then I'll like take that and put it into something else.

I gotcha.

Yeah. So like this thumbnail, it looks like something an amateur would make in Canva. Let's Be real.

It's true. I mean, just for something super fast, I guess, if you're split testing something to do more creatives. Okay, Matt, come on, let's stretch. Stress your thumbnail. Genius youtubeness. Like, what would you do? What's your process here?

So I still sort of assemble the thumbnails in Canva.

You start there?

Yeah. So what I would probably do, put it into Create image, Pull in like a headshot. So here's an image of me surrounded by bananas. Right. So what I would do is I would download this image at full size, and then I would go into someplace like Canva, and by some place like Canva, I would use Canva. I'm also an investor in Canva, so full disclosure, I guess. So I would pull my image into Canva. So here's my image with me surrounded by bananas inside of Canva. Stretch it out. So now I've got a thumbnail that looks like that. And then I'd, you know, pull in some text and say, like, Nano Banana is here. You know, something like that. Stretch it out so it fills out the whole area. And then, because that doesn't really stand out very well, I'll go up to a effects and outline, and then I'd put like a probably a thick black border outline around it like that, just so it stands out a little bit more, and boom. That's how I would do a thumbnail.

I love it.

So I like to use Nano Banana to, like, get the assets that go into my thumbnail, but I like to sort of compose the thumbnail myself.

That makes sense. Yeah, yeah. And I guess for any, like, split testing, you know, I've known other creators will take something like what you just generated there, maybe throw that in a Nano banana, spit out a whole bunch of other versions. So you're doing the A B testing?

Yeah, yeah, yeah, exactly.

That's cool. All right. I'm just thinking of other use cases for Nana Banana that would be kind of fun and useful because I know the infographic thing. I kind of love those, you know, where you have, like, a topic. You could talk about the anthropic story and make that an infographic. Test that out.

Yeah. Let's see. Create an infographic that breaks down the emerging story of anthropic versus the Pentagon.

So this will test how real time its research is as well.

We should see if it'll actually. I didn't test this in my video, surprisingly. And I normally test this is. Is it going to reject any sort of trademarks? Right? Like, generate Mario Generate, you know, Mickey Mouse or whatever.

Dude, look at how fast that was.

That was super fast. So breaking down the Anthropicon story, balancing AI and national defense. Safety first, rapid adoption. I don't really feel like this is an accurate representation of what's really going on.

I mean, it's bringing in a lot of the aspects of.

Yeah, but it doesn't have any details, really.

Yeah.

AI safety research, defining, establishing guardrails for advanced AI defense contracts. Dual use technology. Like you would think there would be something in the image about, like these are the two things Anthropic doesn't want. Autonomous weapons. Spying on US Citizens doesn't have that. You'd think it would have something about possibly being designated as a supply chain risk. Nothing about that. Sure. Right. So I feel like it's not quite there.

Can we drill it for more real time updates?

All right, so let me go ahead and try this again here.

Stress how far this thing researches.

All right, so the two non negotiable red lines. No mass surveillance of Americans. No fully autonomous weapons. CEO Dario Amadai's statement cannot in good conscience accede to unrestricted use. Concessions offered, missile defense and cyber defense only. This is accurate. Now, like this. Nailed it. Breaking news. Ultimatum, unrestricted access or contract termination. Friday. Deadline expired.

There you go.

The Pentagon, US Department of Defense. Unrestricted, all lawful uses. That's their actual wording rationale for demands. National security, threat detection, full range of war fighting. We will not let any company dictate the terms. So yeah, it pulled in the most recent news and created an image about it.

I like that timeline on the bottom. It did as well. Like all the stuff from the past, which the past wasn't. What, like last year basically, until now.

There's definitely some inaccuracies here, but it did a pretty good job. Like, you're gonna have to fact check it if you're gonna use this. Any sort of like, production work, you'll definitely need to fact check it. Yeah, but it's pretty close.

But can't you go back and edit images as well with the model? Like, if we did call out some of these things on the bottom, we could.

Yeah.

Revise your work.

Yeah, try harder. Try harder. The dates on the bottom aren't accurate and there is duplicate information down there. Do better.

There you go. All right.

All right, let's see. All right. May 2025 initial contract awards December 2025 proposal submission January 2026. DoD deadline expires February 2026. Chronology of standoff February 26, 2026. Anthropic rejects ultimatum. That's February 27. Pentagon confirms blacklisting.

Is that news that we don't know?

Is this breaking news we're finding out on the show? Let me look this up real quick. So I'm not seeing that it's actually happened, but I have been seeing that there are threats.

Yeah, it says there's no.

Oh wait, hold on. This is what I found from 19 minutes ago. Trump moves to blacklist anthropic AI from all government work. President Trump announced the US government would blacklist Anthropic in the most consequential and controversial policy decision to date at the intersection of artificial intelligence and national security. The department plans to sever its contract with Anthropic, valued at up to 200 million and require companies it works with to certify they don't use CLAUDE in those workflows. Anthropic will also be barred from all other government work. There will be a six month wind down period to allow the Pentagon, its customers and other government agencies to onboard alternatives to claude. I am directing every federal agency in the United States government to immediately cease all use of anthropics technology. Trump wrote on Truth Social. So yeah, I guess it's happening.

Dude, Nano Banana broke the news for us.

Yeah, so we just found that out because an image on Nano Banana was even more up to date than we were.

Yeah, this is blowing my mind, Matt. We just had Nana Banana create an infographic, we pressed it for more research data and it pulled the freaking breaking news into an infographic. That's wild.

That is hugely consequential to anthropic. Like crazy.

So let's talk about that, I guess. I mean this is like evolving our whole plan of the show. But so it said like a six month wind down, so obviously there's that I guess makes sense. But yeah, pretty quick. I mean you're going to see, I'm sure Grok it sounds like they're already in talks, you know, X AI.

Yeah, it sounds like they're going to probably migrate over to Grok. I do know that also the US government is trying to work on like their own internal like AI models as well so they can release reliance on any sort of, you know, public companies. Yeah, so that is also happening. Maybe they might expedite their own internal models at some point as well. But yeah, that's going to be pretty consequential to Anthropic if a whole bunch of big enterprise level companies are forced to stop using them because those Companies would rather have their government contracts than continue to work with anthropic.

Yeah, yeah.

Crazy.

All right, well, we broke the news.

Yeah, Gemini broke the news for us because we told it to get more up to date info and it found info from literally 19 minutes ago.

I thought we were trying to go more lighthearted with this show. This episode.

It took a turn.

Dang it.

That's so crazy. We will be right back. But let me quickly tell you about another podcast, Content is profit. Hosted by Luis and Fonzie Kameho. It is brought to you by the HubSpot Podcast Network, the audio destination for business professionals. Discover the secrets and strategies on how your business can achieve the frictionless sale. They talk about frameworks, strategies, tactics, and bring on special guests to bring you all the information you need in order to turn your content into profit. They just did a great episode titled the platform trap that's blocking your sales. Listen to content is profit wherever you get your podcasts. So do we want to keep on experimenting? I did tease something with Nano Banana, and I feel like we got to close that loop.

Oh, yeah.

I want to see if I could make Super Mario.

That's right. Forgot about that. Super Mario versus trolls. I don't know.

Versus trolls. While at Disneyland.

There we go. Let's put all the copyright infringement possibilities. I feel like Google's got this covered, but I could be wrong.

We'll see.

Let's see.

I've had it reject stuff. A lot of times it rejects, like, people, right? Like, if you say, like, hey, make an image of Donald Trump shaking hands with Elon. It won't.

Oh, hey, look.

There you go. It's. I mean, that's not, like, trolls. Trolls dolls. Those are, like, trolls that look like they'd be in, like, Dungeons and Dragons or something.

But we didn't specify that. I'll give them.

But we got Mario and we got Disneyland in the background, so no problems there.

I mean, that looks pretty cool. That's.

But I bet you if I do, like, Donald Trump shakes hands with Dario Amodai.

I knew you were going there.

It's not going to do it. Like, I can almost guarantee it.

Maybe it will waiting for it.

Oh, I don't think that looks like Dario. I'm actually surprised that it didn't just flat out say, like, no, I can't do that, honestly. But what you tend to find with a lot of these models, too, is they'll release them and they have, like, way less guardrails. And then they start to see what people use Them for. And then they start to like bake the guardrails back in.

It's like Seed dance all over again. Right?

Yeah. And I mean you saw that with Sora too, right?

I've been noticing on X this week this is not Nano Banana related, but just a lot more of these. Longer. Not longer, but like one was up to 14 minutes. Like a feature film, but all generated with AI probably the door Brothers. Yeah, that's them. The Door Brothers.

Yeah, they make like really crazy AI generated videos and you can't really call them AI Slop because you can tell they put a lot of time and work into making them look really good. Even though they used AI to get the generation out of it.

Dude, I was blown away. I was like, I just watched a 15 minute AI generated video that had a full on storyline, consistent characters. I thought that was cool. There was also another video going around. It was almost like showing what happens, you know, in the future when AI is just doing everything and humans are what, generating the energy. But I just thought I was like, man, the level of detail and also generation with video and how this is now becoming like a full on, almost like a feature film if you wanted it to be. It's like right there and it's just, it's. It's pretty cool.

Yeah, those videos are definitely crazy. There was one where there was like something they called like the Phenomenon or something and like there was some giant weather event and the lady was like rushing out of the building and like trying to get back to her kid and it said to be continued. And like she drove a cyber truck down the road and then crashed her cyber truck and then got picked up by Cash Patel. That was one that was like really crazy. And then there was also the one that they did in collaboration with Logan Paul. Logan Paul actually posted it on his YouTube channel, the Door Brothers One.

Yeah, yeah, yeah. I think he did the voiceover is what they said at the end of it. And like featured himself like, this is pretty cool. This is like the collab of. Yeah, these new producers, edited prompters. I don't know what you call the like the Door Brothers. I'm like, this is opening up a whole genre. I feel like that's going to keep developing, you know, people with their workflows and stuff.

Yeah. And I mean like there's a difference between AI Slop and people who used AI to make something really, really impressive looking.

Yeah.

And I find like the Dore Brothers, they tend to be less on that AI Slop side and more on the like, all right, they used AI, and I can tell it's AI, but holy crap, they probably put a ton of time into this, right? That wasn't just like a couple prompts and slap them together. And look, here's a video. It's like they worked on this for a while.

The other thing, I thought that was kind of funny.

Oh, yeah, yeah. By 2030, almost 80% of people

had lost their jobs.

They had no money, no purpose, but they had a lot of time on their hands. The less people actually did physical work work, the more they wanted to appear as if they did.

What if we could use the energy of humans to power the machines that took away their jobs?

Energim solved our need for energy and your need for purpose, Matt, where are

we going in the world? I'm a little scared. After I watched that, I was like,

oh, my God, no, this is reality. This is where it's all going. And I like how all of the actual tech CEOs are all fat and old.

All the billionaires are, because they're not working out well. They don't need to produce the energy. Oh, boy. Yeah, I saw that. And I'm just like, oh, my God. Laughing but also frightened at the same time.

Oh, man. Anyway, yeah, no, I feel like there's some reality that. I mean, we talked in a previous episode about how, like, there's a site where your Clodbot can go and hire humans to do work for them.

Yeah, that's right. I have seen that. Yeah.

So, you know, like, it's not too far outside of reality.

Honestly, no. Well, how about this? Is there anything else on the Nano Banana side of things? Because, like, I was just blown away with. Yeah. Seeing Nano Banana, seeing these videos starting to get published, and they're like picking up steam, like millions of views for each of those. We're just going to see more of this developing, like people getting super creative. What? Google Flow, that's a Google platform that Nanobanana 2 is baked into.

Yeah. So that's a video platform. Right. Where you can use VO and stuff as well. But the starting images now can use Nanobanana 2 and it uses zero credits. So essentially your starting image can be free. Yeah. So inside of Flow, you can use nanobanana for free instead of Gemini, you can use it for free. If you're using it over on AI Studio, you do have to connect an API key and pay, but pretty much anywhere else you can use it for free. It seems like.

Cool. So moving on. What? Last week we were talking about agents. And now this year, it seems like the year of agents becoming useful and now they're being baked into almost every application that we're natively using. Let's just show off some of these. I think this will be kind of a rapid fire thing where there's just been so many different releases, some are on different paid tiers that might not want to pay for, but it feels like it's all going to come down. Obviously everything comes down in price in this whole AI world.

Yeah.

And becomes more available.

Yeah. So I think the most interesting one that came out this week is Perplexity Computer. This one is like Perplexity's version of an open cloud. It runs in the cloud. It can like use any of the tools that you want it to use. So it can access your email, your notion, your slack, pretty much any tool you needed to access it seems like they're giving access to. But because it's Perplexity and it's not one of the foundation model makers, it can actually use different large language models based on what you need it to use it for. Right. So I think it's using Claude Sonnet as like its core model, but it also can use GPT 5.2, it can also use Grok, it can also use Gemini, it can use VO 3.1, it can use Nano Banana, it can actually go and access any of the tools. It's model agnostic and so if it needs a long context window, I think it goes and uses Gemini. If it needs to do some math or something, it'll go and use GPT. I'm getting the exact uses wrong but you know, depending on what you ask it to do, it's going to decide the model that's best at doing that thing and go use that model for you. And why this one is interesting is this one's all in the cloud and one of the things we're seeing with OpenClaw is it's a big sort of security risk. Right. There's people have had issues with their things getting hacked and their API keys stolen. And one story here we've got Summer Yu, who works at Safety and Alignment at Meta's Super Intelligence. Well, she put out this tweet. Nothing humbles you like telling your Open Claw confirm before acting and watching its speedrun. Deleting your inbox. I couldn't stop it from my phone. I had to run to my Mac Mini like I was defusing a bomb. And we could see that it's like it just goes through and is checking how Many emails are left. More old stuff. Get the remaining. It's basically going through and deleting and she's saying, don't do that. Stop doing that. Don't do anything. Stop. Open claw. Like, basically yelling it. And it's funny because Peter Steinberger, the guy who created OpenClaw, says if you just do slash stop, it'll stop it. But she basically had to run to her computer and get it to stop doing that. Well, that's, you know, exhibit number one of why more managed agents are probably a better idea if you don't totally know how to use this stuff. Yeah. Because something like openclaw relies on you kind of knowing what you're doing a little bit. You kind of got to understand how to use your terminal a little bit. You kind of got to understand how a lot of the API keys and things like that are working in the background. And if you don't understand that stuff, you should probably avoid tools like OpenClaw.

Yeah.

Perplexity Computer, on the other hand, is like, look, we'll handle all of the complicated stuff. You just tell us what you need done, we'll pick the right model, we'll tell it what tool to use, and we'll just make sure it gets done.

Yeah, I know you've done openclaw or claudebot, whatever it was called in the very beginning, but, like, I was so close to jumping. I'm like, you know what, hold on. Like, I'm still actively learning all the little details because the last thing I want to do, because I've seen too many of these things now where you give it access to too much stuff or it just starts doing things and I'm like, the last thing I want to do is just get myself in some trouble. However, that looks like I don't even know. It could be anything. So, yeah, I like the idea of Perplexity. When I saw this go live the other day, I was like, yeah, all right. Yeah.

Well, the downside of Perplexity Computer is that right now it is only available on their Max plan. Yeah. So available on web for Max subscribers today. Coming soon to Pro and Enterprise. Right. The max plan costs $200 a month or $2,000 a year to use. So you gotta be on like, their higher end, more expensive plan to actually use it. But you can pretty much tell it to go do anything and it'll go and just run in the background and build whatever you ask it to go build and handle the actions you want it to go handle. Notion rolled out a new AI agent This week. Haven't had a chance to play with that one yet either. We also have Manus, which is an AI agent platform that recently got acquired by Meta, also trying to do something very similar. Cursor rolled out an agent this week as well, where you can tell it what you want it to go develop for you and you can set it to go and work for like 10 hours and it will just go like while you sleep and build something for you. And the cool thing that I like about Cursor is the whole time it records what it's doing. Like it actually opens up a virtual machine, uses a computer behind the scenes and it actually records video of everything it does. So when you wake up the next morning, let's say it worked for 10 hours, you can go back and watch the video of everything it did on that virtual computer for you to see if like, maybe you can spot it making mistakes or doing something you didn't want it to do, which is really, really interesting. But yeah, lots and lots and lots of like, agents actually becoming useful.

Yeah, useful and just honorable mentions. I didn't think Microsoft Copilot Task dropped something. Cowork from Claude was adding some more features to automate things.

Yeah, yeah. So the Microsoft Tasks thing, they sort of teased it. They announced that it's coming, but they haven't really given access yet. The anthropic Cowork thing, they basically added the ability to do like scheduled tasks. So you can tell it like Every day at 2pm Check my email and write up draft replies, but don't send them. I'll go through and decide what to send or not. Or, you know, every Monday morning at 9am, do a scan of all the AI news that came out this week and send me a message in Slack to tell me what news came out. Right. So you could do like scheduled actions. That's kind of a new, new feature to the Cowork in Anthropic. But yeah, I mean, like, everything is kind of turning into like useful agents now. And I was talking about this a couple years ago of like, prompt engineering being this very sort of transient concept of like, you're not really going to need to know how to do prompt engineering. Eventually you're just going to tell your computer what you want and it's going to go do the prompt engineering behind the scenes and go figure out what needs to be done and then go and do the things and use the tools and even possibly make the tools that it needs to accomplish the tasks.

And we're kind of there I think we're there. Yeah. Like, I don't think I've overthought a prompt in a long time. It's just like, if you give it enough context, especially just like something basic like chatting through Claude, have it create the prompt from whatever the heck you just did and then that can become a recurring thing. Plug it into, I don't know, coworker elsewhere.

Yeah, I mean, it's getting to a point now where you can kind of treat your AI agents and your AI chats the way you would treat an employee. Like you give them instructions of. Here's what I want from you. Here's my expected outcome. Here's the tools you have access to go do it. Let me know when you're done.

Pretty much right.

And it can just go do the thing.

Yeah, it's cool. It just has you thinking and then speaking and creating and closing the gap as much as possible.

Well, cool.

I'm sure we're going to get more agent stuff coming up next week and beyond. So let's move to the final thing.

Majigger, your alma mater, your work alma mater, Alma matter. Yeah.

It was my first official real job. I had some other gigs before that. The good old Burger King, the BK lounge, that is. Thank you, Dane Cook. This is bringing back bad memories, I'm just gonna say that. But with this technology, I probably would have been fired very quickly because now Burger King is surveilling on all their. No, it's not really surveillance. It's like basically they have this thing called Patti that is in the ear of. I don't know if it's every employee, but it's this little headset that's demonstrated. I just want to show you the video because I think it's interesting. Here it goes.

Hi there. Good morning, Patty. Looks like we had a great breakfast shift today. Is there anything that needs my immediate attention? The team's friendliness scores this morning were the highest this week. We are running low on diet Coke in the freestyle machine. Thank you, Patty. I'll take care of it right away. Looks like the women's restroom needs cleaning. That's very helpful, Patty. I'll get to it right away.

Hey, Patty, can you remind me how to build an ultimate Steakhouse Whopper?

Of course. The ultimate steakhouse whopper includes one 100% flame grill beef patty on a toasted sesame seed bun, three rings of creamy peppercorn aioli, two slices of Swiss cheese, three half slices of bacon, one scoop of mushrooms, and is topped with six onion rings.

Make me Hungry.

Yummy.

Great job so far today, team. We're one order away from hitting our upsell goal.

They would not be that happy. Hi, welcome to Burger King, where you roll. What can I get for you today?

Hi, yes, I'd like a Whopper meal with large fries and a Coke, please.

Sounds good. So I have one classic Whopper meal with a large fry and a large Coke. Would you like to add a cinnamon apple pie to that?

Yeah, I'd love that. Upsell Target reached Upsell.

Hey, Patty, I'm about to clean the shake machine. Awesome.

Would you like me to remove shakes from our digital channels until you are done cleaning?

Yes, thank you.

Shakes have been removed from our menu boards, third party delivery kiosks, and the BK app. I will add them back as.

As soon.

Soon as you are done cleaning the machine.

Thanks, Patty. And while I have you, can you remind me of the procedures to clean the shake machine?

Of course.

Hi, Patty. We just sold our last cinnamon apple pie.

Thanks for letting me know. Would you like me to remove them from our menu until tomorrow's shipment arrives?

Yes, please.

Okay. Apple pies have been removed from our menu boards, third party delivery kiosks, and BK app. I will add them back as soon as tomorrow's shipment arrives.

Thank you, Patty.

Hey, Patty, Raul just called in sick for his shift tomorrow morning. Sorry to hear Raul is sick. Looks like Maria is available for that shift. Okay, thank you. I'll call her right now.

Patti, why don't you call her for me?

I thought the same thing. I'm like, well, if BK is that good, Patty, then just do it.

Patty's hungover and doesn't want to come in. What do I do?

So look at this thing. Yeah. AI drive throughs just yet. So it's building chatbots into employee headsets, but it doesn't seem like the brand's ready to widely launch AI drive thrus just yet. Something like chains like McDonald's, Wendy's, and Taco Bell have attempted. So we're tinkering with it. Interesting.

With pretty hilarious results oftentimes. If you've ever seen those clips.

I have seen a little bit of those. Yeah.

Somebody's like, let me get a milk. And they're like, one cheeseburger. No, a milk. Okay, I've added a second cheeseburger. I want a milk. Okay. We added fries to your order.

Oh, man. I don't know. I guess training could probably be cut down really quickly, I would imagine. Yeah. Yeah. Literally walking you through how to do your job while you're.

Yeah. The hilarious thing and the thing that I think most people are kind of honing in on is the fact that it's counting how many times you say please and thank you, and then it's giving you, like a sort of like a score on, like, how pleasant you are. So, I mean, it's like, to me, it feels like only a matter of time before, like, Patty is making the firing decisions. Like, hey, last week we noticed you talked to 30 people but only said thank you to 15 of them. That's unacceptable to us. Like, I see that as, like, the future of where this all goes. Like, come on. Like, the people who own the franchises want the actual employees in bk, no offense, Joe, to be as brainless as possible. They want the AI to make the decisions for them, and they don't want anybody there to actually have to think.

I didn't spend a lot of time there, Matt. I did have fun. But I guess, yeah, I read here that Patti the AI assistant is powered by OpenAI. It's in 500 restaurants already and rolling out to all 7,000 by the end of this year.

Yeah, it's a good thing they didn't use Anthropics, because they'd probably have to cut ties because I bet people in the military buy Burger King and they're probably on base. Yeah, exactly.

So then. Yeah, that's crazy. I'm just thinking anthropic again now. It's like how deep those routes go.

It all comes back to this anthropic story. Right? It's so consequential across the board. I'm telling you. Like, if there's Burger Kings on base that were using anthropic inside of their headsets, they would probably have had to switch over to OpenAI or Grok instead.

Exactamundo. So it goes that far? So, yeah, man. All right, well, I don't know where this goes, Matt. I think next week it's going to be just as busy and we'll learn way more. I'm so happy that we got to break news to us at least.

Yeah, yeah. It was breaking news to us, but by the time this episode's out, it's like a few day old news.

Yeah. Still fun, nonetheless. Yes, yes.

It's okay. People get to hear our excitement that we learned something new while in the middle of recording an episode, and we learned it by generating an image with Nano Banana, of all things. That's actually kind of cool.

It's pretty cool, I think. I think that's rad. Where the world develops next week and beyond, who knows. But you need to come back and watch. Yeah.

Yeah. So if you are listening in, I mean if the world still exists next week by the time this episode goes live we'll have all the updates for you. But if you like stuff like this, make sure you like and subscribe to this channel and we'll make sure this kind of stuff keeps showing up for you because we're not going to stop talking about it unless we're forced to unless the government tells us we're not allowed to anymore and in that case we won't. Or maybe we'll hold our line and we'll. Anyway we'll see. We'll see. Anyway, thanks again for tuning in everybody. Really, really appreciate you hanging out with us and yeah, it's been a blast talking about what's going on and what is scaring the crap out of us right now. Hopefully we'll see you in the next one.

Bye. Bye.