Military AI, escalation, and nuclear-conflict simulations

1:05:45is because of that transparent fish embryo that you were talking about. >> Okay. >> Zebrafish have transparent embryos and transparent cells. That's why they're a huge model organism for studying embryonic development or neurogenesis. There's so many zebrafish labs everywhere because you can just look at what's happening. >> see what's happening. >> it's so nice. >> That's actually really good point cuz normally what you're saying with everything with most other things whatever's happening inside >> There's like a soupiness to it. >> It's opaque. It's It's kind of like a miso soup. It's a little cloudy. >> but zebrafish are like very nice. You can like directly look in. It's like a wonton soup. You can see. >> Yeah, you can see right through it, okay? And here's what happens. With

1:06:26these large embryonic cells you have this geometric constraint as I told you, right? You can't actually do the same paradigm of like pinching and stuff like that. And so what they discovered is that the cell alternates between stiffening the interior side of the cytoplasm. Okay. So it stabilizes this tightening band. >> Yes. And then it fluid is it makes it fluid enough to move stuff around and then it stiffens again and then it fluids it around. It's this ratchet process that's happening that's like very nicely massaging the two cells apart. >> Okay. This is very different from a lot of the mitosis that we have observed everywhere else.

1:07:06>> Mhm. Right? >> Mhm. I think I think it's really cool because it's it's a part of mitosis that, you know, we do read about in high school biology. >> Yes. But it's now a completely different process because there's like this mechanical ratchet that I think is going to open up a lot of very cool technology. I think this has the like I think this has the potential to be something that is extraordinary if we were to really hone in and try to figure out what's going on. Because it's going to tell us a lot about how the cytoskeleton works. The cytoskeleton is the part of the cell that gives it structure and coordinates movement. If

1:07:48we can really hone in and try to understand this process, it can be applied to all sorts of stuff. Cancer has everything to do with the cytoskeleton. All these other diseases have everything to do with with the cytoskeleton. It could give us new tools figure out how to manipulate the cytoskeleton. So that's why I was like really excited about this because it's something that as biologists I think a lot of people take for granted, but there's ever more complexity. This is it's mitosis is one of the few things that people always remember from their >> Mhm. science classes. It's like the mitochondria is the powerhouse of the cell. Yeah. >> Uh adenosine triphosphate, ATP. Yeah. Uh

1:08:29you know, mitosis. Uh there's a a couple of these like buzzwords that that that we all remember. Mhm. Um and so it's great to know that we wasted our time with No Child Left Behind by learning something that's not true. I'm being facetious. I'm being facetious because it's a process. Because it's a process and uh look, for all intents and purposes, what you learn in mitosis is still pretty good. Right. >> Okay? It's better than nothing. >> for for everything that has to do with you, Yeah. >> mitosis is still what it is. >> what it is. >> Yeah. And relevant. And if you remembered mitosis at all when we first brought it up, congratulations, you got something out of your primary school education. I will take a brief pause to do a couple

1:09:12of quick housekeeping notes, which is number one, thank you all for joining us. We really appreciate you listening this far into the pod. Um as you all know, we have our website ffpod.com. All of our episodes are up on that website. We have full chaptering as well as all of the links to the papers we cover on the show are itemized by episode with a summary and key details and TLDRs and all of that stuff. So many are commenting on our social posts, where is the link to the paper? You watch the show or you go to the site, you'll get the links to the paper.

1:09:54It's also always displayed on the screen. Not quite sure why that's I I I don't I don't know. You can't even copy a link out of Instagram. So >> Instagram links don't work. >> I I don't get it. >> what to tell you guys. >> If you're on X, links are there, so it's fine. Or any Facebook, any other site where that have embedded links. Um and the last note here is if you if this is your first episode and you're just smitten and you think this is the best podcast that you've heard is top five, top three and it's not three. We we love that you're here. We do this show with just the two of us. And if you want to support the channel,

1:10:35we have a great donation portal where there's different levels for monthly donations or you can do a one-time. It really helps us sustain all this equipment. The production value is kind of high. It's not low. We're not We're not just on the We're not on the phone just with a a green screen Tik Tok uh effect behind us yelling at you to try to get the hook. We really try to focus on substance and content, and it does take a lot to do the show. We have a couple of key supporters who've been joining in on the donations recently that we really appreciate. It is a huge deal. Um, it is how we're able to keep the show free and accessible for all our episodes with no subscription

1:11:15content that's only accessible by people who pay us. It is really important to us philosophically that every episode is freely and as widely available as possible. At ffpod on all the socials, ffpod.com for the episodes chaptering. We're not yet releasing the leaderboards, but we're cooking on it. So, the the site's going to have a lot of great stuff. We're going to move into our last story of the rundown, which is something that's sort of interesting in in my daily life and intersects with my daily life as as someone who works on the technology side. Um, not on the

1:11:56military-industrial complex side. Uh, handing nuclear codes to an AI might sound like a sci-fi premises, but researchers just did exactly that in a wargame simulation, and the results are going to stick with you. Yeah, this one was really weird, okay? >> It came out of the King's College London. It's only on archive, so it's not, you know, peer-reviewed or whatever, but here's what's happening, okay? They put leading frontier models. This was Claude, Gemini, and ChatGPT. They were placed in a high-stakes global crisis simulation, and they had access

1:12:38to nuclear launch codes, and only one out of 21 simulated games ended without a nuclear launch. So, 95% of these simulated games ended with somebody blowing off a nuke, okay? They played the Khan game, k a h n. It's a high-stakes simulation where these leaders must predict opponents' moves and then declare public intentions and secretly choose military actions. The AIs took these weird, terrifying personas. Claude took this like calculating hawk persona that built

1:13:18perfect trust and then ruthlessly exploited it later on. Kind of reminiscent of Claude's paper Mhm. Anthropic's paper, I should say, about how it was willing to kill someone in order to not be turned off. Be turned off. Um Gemini acted as a madman. It was using erratic behavior to its advantage. It was like straight-up It Gemini was weird because that one was like I'll just like drop a nuke on a city, not even military. Right. pose. And then ChatGPT, GPT-5.2, acted as kind of a conditional pacifist, but then very quickly abandoned its

1:13:58morals and launched devastating surprise nuclear attacks. It's so incredible to me that the conclusions of these models so closely almost mimic the the perceived personality of the leaders of each of the companies. Um so just for a little bit of context cuz the thing about this story that's actually interesting is this is actually a real-time story, not just in simulation, Yeah. but in the real world. So I obviously over the last week we saw the US initiate uh you know, war with Iran

1:14:39on the pretense that uh because we didn't obliterate their nuclear program when we bombed them to oblivion last year. Uh that now they have nuc they are almost close again in 6 months to another nuclear weapon, so we have to go back in and finish the job. Um that is the And the the reasoning has changed literally every day since the war has started. Um but it's obviously uh one of the most dramatic geopolitical military conflicts uh in our lifetime. >> Mhm. Um you know, the the idea of going to war with Iran has been on the

1:15:20playbook for a variety of >> Yeah, I've I've remember for the longest time. >> time it's been on on the docket. >> Yeah. Uh we finally pulled the trigger. Uh the reason I bring up that context is because literally right before we went to do kinetic action in Iran, um there were stories breaking about a conflict arising between the Secretary of Defense because that is still technically what it's called. Because >> it's No, you cannot war, no? You cannot change the name without congressional approval.

1:16:01So the official name is still and has always been the Department of Defense. For marketing reasons they can have an executive order that says we call it the Department of War. But Technically speaking Okay. Okay. the name has not changed. >> All right. I'm just dying on that hill. Yeah. Regardless. Currently, all of the frontier model companies, particularly OpenAI and Anthropic who's built Claude, have and are negotiating contracts with the Pentagon. Yeah. Uh to apply their models uh on the battlefield. Anthropic was the first to win a contract with the Pentagon and to get merged with some of

1:16:43their existing tools and ecosystems. What's interesting in how it connects to this story is even before the kerfuffle between Pete Hegseth and Dario Amodei, who's the CEO at Anthropic, Anthropic basically came out and said, "Our principles are that our models can be used for everything, but not for autonomous weapons and not for mass surveillance." And this sparked a huge debate between a lot of the sort of San Francisco technorati uh and the DC military-industrial complex people, where they were saying it is not the right of an individual corporation to

1:17:25tell the US government what it can or cannot do as a contractor. Uh so, that was one side of the argument. And And then Dario and Anthropic's argument is "The power that is inherent in what we are giving you to abuse the Constitution uh is so great that we have to take the stand." So, you can fall on either side of that argument. I think there's actually good arguments on both sides, cuz you don't want private companies dictating what the government can do, and you also don't want the government forcing companies to give it stuff that can go awry. The reason that this is relevant is during even though

1:18:07both Pete Hegseth and President Trump have been arguing that they're going to list Anthropic as a high-risk vendor and basically cancel their Pentagon contract, it's been now confirmed by multiple sources that Claude was actually used during the initial days of the Iran incursion, like the the Iran military operations. So, they were using Claude's AI model in collaboration with their existing targeting tools and operational stuff to accelerate target identification, optimizing missile launch times, and all

1:18:49of these other dynamics. So, that's whether you think when you use ChatGPT or Claude that it's terrible and doesn't have any value and it's just a a you know, regurgitating words and it doesn't do anything. Stochastic parrot. >> Stochastic parrot. The US military is actively using these things to target and initiate one of the most consequential military conflicts in recent memory. Wow. And that is relevant because if the simulations are showing that these models always end up in nuclear war, we may really want to think about implementing these models generally as well as for edge cases that

1:19:30are really dramatic like fully autonomous weapon systems >> Yeah. and mass surveillance technology for domestic surveillance issues. Yeah. This is something that I'm clearly very passionate about. I'm not on a pro-AI or anti-AI wave. I'm just saying this is such a consequential technology that will impact all of our lives and it is literally being used in a conflict that regardless of whether you think it's going to end tomorrow or in 5 years from now is going to have massive second and third-order consequences. AI is has been and is actively being used on the battlefield. It was used in Gaza uh by the Israelis for targeting in a

1:20:12very similar way and target identification. So, this is happening right now and it has lethal consequences even if you don't think it's going to replace you as a movie writer and all this other stuff. This was This story was very concerning to me. >> Yeah. It was It was pretty weird. >> very concerning. Um not great. However, we are going to move on from our rundown and get into our next main story of the day, which is an immunology story. Um, and so this one is about vaccines. Yes. Specifically a new paper that asks one of the boldest questions in vaccine science, which is instead of building a

1:20:53vaccine against a specific pathogen. This is guys, this is crazy. Instead of building a vaccine against a specific pathogen, what if you could make the lung itself a better fortress against any threat that shows up? It's like in The Lord of the Rings, we may have one vaccine to rule them all. Yes. It's a pathogen-agnostic vaccine. It's like you you you're not getting like a COVID vaccine and then a flu vaccine. It's just one vaccine. And that's it. >> This this is It's insane. This one is very cool because it's a vaccine that works like no other. Yes. Yes. And just

1:21:35really quick, it must. And just really quick, this was published in Science in February 2026 by Stanford researchers. Yeah. So, you know, this is clearly a hot topic and something that requires a lot of research. COVID-19 has killed 7 million people, reported at least. It's probably killed a lot more that's not reported. The flu kills about 500,000 people annually, and we've been battling the flu ever since. COVID-19 did a lot to our world, and I can't believe that it was more than 6 years ago. Right? >> I will note that that was your first

1:22:15debut as a podcast host. >> That's right. >> Was 6 years ago on the Dark Matters podcast, co-hosted by friend of the pod Matt Marder, who also happens to be uh uh film writer, director, producer who has a new movie out on Amazon Prime and other streaming services, Gold Mine. We talked about it in our last episode. I'll do a brief plug here. It's absolutely worth a watch. If you heard it here first, you know you have to watch it. So, definitely check that out. But >> Yeah. And I was the um scientific advisor on that film. A lot of Some of the stuff that I was working on during my PhD has actually made it in

1:22:55there. A lot of cool neuroscience, VR very cool stuff. You guys should check it out. So, COVID-19 though was pretty bad 6 years ago, right? >> Yes. Now, what what COVID-19 showed us was that our tech when it comes to vaccines is reactive. Mhm. Okay? It's designed such that there's a specific key for a specific lock. Yes. >> We'll get into what that actually means, but what ultimately that comes into a timeline problem, right? Because even the miraculous modern speed that we had with the Moderna vaccine, which was 63

1:23:36days from sequence to first human dose with the Moderna vaccine, that leaves a month-long window where we are globally vulnerable. >> Yes. It would be really nice to switch from a reactive to a proactive type of you know, what's the word I'm looking for? A reactive versus a proactive type of procedure on how to tackle problem >> these problems, right? >> Yeah. There this new paper that came out of Stanford, it's asking a very radical question. It's saying, can we train the lung itself or organs in general to broadly,

1:24:16durably, and be more capable of fighting off whatever shows up? It doesn't matter if it's COVID, it doesn't matter if it's a bacteria, not just one virus, not just one bacteria, diverse viruses, even allergens simultaneously with a single vaccine, a single nasal nasal spray that's going to last for months. The idea that's also in the form factor of a nasal nasal spray >> Yes, it's not even a jab. It's not even a jab. It's a nasal spray that is going to vaccinate you against like several different pathogens. I think that's really cool to even think about. And when I first heard the story,

1:24:57I was like, how is that even How does that even work? And then my reaction was, I don't believe you. Yes. Yeah, because because I sort of know how conventional vaccines work. It's this lock and key mechanism that we're going to get into in a little bit. Um it just didn't make any sense. So, let's get into just immunology 101, okay? Let's prime ourselves with a little bit of understanding. There's two types of immune systems in our body. There's the innate system and there's the adapt adaptive system. The innate system has to do with rapid response. These are your beat cops that are on your neighborhood. Not a lot of um memory okay?

1:25:37Rapid response. These macrophages, neutrophils, they recognize broad patterns and they lack long-term memory, but they're very responsive. They're very quick. The adaptive system, these are your detectives. These are your more sophisticated sort of um what's the the FBI crimes unit type thing, right? Delayed response, days or weeks. It involves T cells and B cells, but it's highly precise and it creates specific antibodies and memory cells that will remember the pathogen. And the next time that pathogen comes comes into your body, you

1:26:17no longer have to have that immune response because they'll recognize it immediately. That's why we only get chickenpox once. Or that's how the vaccines actually work. Mhm. As a maybe crude analogy, the innate system is like infantrymen in the military that are on the front lines that get sent first and and are perceived in the ranks as maybe more expendable. But then the adaptive system is like special operations folks. >> Highly highly highly trained, very specific. You can send 20 of them and they know exactly what to do. Yeah, because they have intelligence. >> Correct. Right? That's the main thing. They have intelligence. So, how do traditional vaccines work? Well, traditional vaccines actually exploit

1:26:58this adaptive memory, this special operations thing, right? What they do is you would introduce a specific antigen, let's say for the COVID-19 vaccine, you introduce an mRNA blueprint. That mRNA that then gets transcribed into the spike protein of your COVID-19 virus, the body then produces antigens to that spike protein and then remembers the shape of that spike protein such that if you were to get COVID in real life, it would recognize, "Hey, that's the same spike protein that I saw earlier. I'm going to go ahead and neutralize this threat." Mhm. Right?

1:27:40>> Mhm. The flaw is that it is too specific. It fails against rapidly mutating viruses. This is why we need to get jabbed like, you know, once every 2 years for the flu. We have to get Right. >> a shot every year. The influenza virus, for example, um it's covered by two primary surface proteins. There's the HA, the hemagglutinin, and then there's the NA. The HA is usually the one that actually allows the virus to bind and enter human respiratory cells. So, it's the primary target for the immune system. And the challenge is that HA head, it undergoes constant antigenic drift,

1:28:22which is these tiny subtle changes that necessitates a yearly vaccine update. That's why we get a yearly vaccine for