Skip to content

Podcast

Inside the Iran War and the Pentagon's Feud With Anthropic With Under Secretary of War Emil Michael

All-In with Chamath, Jason, Sacks & Friedberg

Source ↗ ← All highlights
  • Operation Focused On Disarming Not Prolonged Regime Change
    • The Iran operation aims to disarm Iran’s ability to sponsor regional proxies and develop long-range weapons rather than immediate regime change.
    • Emil Michael explains strikes targeted depots, drone factories, ballistic missile stockpiles and prior planning like Midnight Hammer enabled precision raids. Transcript: Emil Michael Think the president talked about this is a weeks, not months kind of operation. And it’s aimed at essentially disarming the regime or the country in such a way that they can’t supply Hezbollah, Hamas, Muslim Brotherhood, all the kind of terror groups that get sponsored By weapons and money from Iran, not to mention the nuclear bit. And that’s why you see from the reporting, they’re going after the depots, the, you know, we went after nuclear sites before. They’re a prodigious drone maker. These like huge one-way attack drones that can go hundreds and hundreds of miles. Lots of ballistic missiles that are aimed at every country in the Middle East, as you’ve seen, they’ve attacked them. So I think that’s one. In terms of boots on the ground, there’s no scenario where we have some protracted boots on the ground, Afghanistan, Iraq-2 scenario. Jason Calacanis Freeberg, your thoughts on this war? Obviously, a lot of people voted for Trump in order to have the peace dividend that he was (Time 0:04:36)
  • Technology Plus Looser Rules Drove Recent Mission Success
    • Modern success stems from matured technology paired with relaxed rules of engagement and experienced leadership.
    • Emil contrasts restrictive Afghanistan-era ROE with a Colin Powell style: clear objective, overwhelming force, delegated judgment to commanders. Transcript: Emil Michael Say, no, it wasn’t only a technology maturation from five years ago. It’s the rules of engagement. The rules of engagement that we used to have, there were some, I mean, if you read about them, some of them are insane. Like if in Afghanistan, if the guy had a small gun, you had to have a small gun. And, you know, there was this parody in weird ways. And when you’re like, well, but is the objective to have like a fair fight or an unfair fight? Well, if you’re on our side, you want it to be unfair. So the rules of engagement were relaxed to be. Chamath Palihapitiya Who writes those, Emil? Who sits in an office and says, you can’t shoot back if a combatant is shooting at you if you aren’t matched gun for gun. Emil Michael Yeah. I mean, who writes that crazy policies that are written into military departments? And that’s why when Secretary Hicks talks about this kind of thing and what was happening with him when he was in Afghanistan, if you ever read his book in Iraq, he’s like the rules of engagement Were so punishing that we were at risk all the time because you had to have a legal understanding of what was happening in every minute in the battlefield as opposed to, well, your job Is to take out these guys and protect these guys. Here’s your munitions. Here’s the red lines. And then in the middle of that, go. Use your judgment. Your commanding officer, use your judgment on how to win. We kind of gone back to that, use your judgment, push responsibility field, still have your red lines. But other than that, the objective is the objective. (Time 0:16:22)
  • Drone Swarms Are The New Core Military Arsenal
    • Drone swarms and unmanned systems now dominate battlefield effects and require large arsenals and new tactics.
    • Emil notes one-way attack drones can cost ~$50k–$80k and the US runs a drone dominance program to build those arsenals. Transcript: Emil Michael Changes it big to well. So the Predator drone was the first big drone program like 10, 15 years ago. It was this big honking drone. And then if you remember, Obama would take out some of these Al-Qaeda leaders with drones on their balcony and things like that. I think President Trump took out Soleimani with a drone near his car. That was the beginning. And then the Iraq, sorry, the Russia-Ukraine war happened where it’s like drone on drone, 70% of the casualties are because of drones. So drone on drone warfare, robot on robot warfare, those things are the future for sure. And that’s why companies like Anderol are companies like Anderol is because they’re making unmanned systems. Jason Calacanis And this has been something you’ve specifically been very focused on. And you tweeted today a little bit about a competition. We’ll play a little video here. And this Lucas, low-cost unmanned combat attack systems, it used to take a lot of time. It certainly wasn’t startup time to get new product into the channel for our military to use. Explain what program you’re running here. It feels like the DARPA self-driving challenge all over again, and what these drones cost. I know there’s a company making them for, I think, $35,000. Am I correct? Emil Michael I mean, the small drones like I’m holding right there are way cheaper than that. The Lucas one-way attack drone, which can go five, six, 700 miles at the speed of an airplane, carry a big warhead. Those are like $50,000, $80,000, depending on what kind of equipment you put on it. (Time 0:18:21)
  • AI Use Must Match Risk Profile And Response Time
    • AI will be applied at the edge for target recognition and rapid intercepts, but confidence thresholds vary by risk scenario.
    • Emil gives examples: laser defense against drones tolerates errors; space intercepts (Golden Dome) require fast, AI-enabled discrimination. Transcript: Emil Michael Mean, I believe that a sophisticated drone war is going to be drone swarms controlled by AI to some degree or another, right? To what degree the control matters. Like, for example, drones have decoys, they could spit out, you know, they could dazzle, they could put out things. So, how do you discriminate what’s a drone and how to hit it? You know, you can use AI for that because it’s learned, you know, how to do automatic target recognition, for example. And then also, could it identify a person? And does that make it safer? So it’s going after actually someone you want to get and not someone you don’t want to get. So there’s a lot of uses for AI at the edge, if you will, in the future here. The Ukrainians and Russians do something called like a kill box where they lose comms because it’s jammed for this drone. And then it just starts going in a box and looking for, you know, the person that are trying to get and they’re trying, they’re starting to use AI to do that. Jason Calacanis And China has this ability already, probably times some magnitude. Yeah. Emil Michael They have drone swarms because they can, they can force the companies that make them, not just EGI, to interoperate. So interoperating drones called heterogeneous autonomy, right? You take different kinds of drones and how do you communicate with one another and then make sure they’re not going after the same target is like a pretty complex thing that they’re definitely Working on. Jason Calacanis And let’s talk about the fidelity of these. Obviously, AI is a new technology. It can make mistakes. Anybody who uses it on a day-to basis might experience a hallucination. How confident are you in the AI? Ukraine and Russia conflict, they obviously are not going to be as thoughtful maybe as we are in putting this together. They’re in a hot war right now. But we as the United States have to be very thoughtful about this. So how confident are you that this isn’t going to make a mistake? I think that’s the key to a lot of this debate. And when will it be, you know, perfect to find as much better, and I guess this dovetails with the self-driving, you know, thoughts, it has to be a magnitude better than a human. So when will this be a magnitude more accurate than, you know, when we have, make a mistake as a military and we kill a civilian? Yeah. Emil Michael No, it’s a good question. And I don’t know when that moment hits, that FSD moment where it kind of gets better. It’s certainly not there. And you wouldn’t want to take huge risk with that. There’s a gradation of when you would use that and what kind of risk you’re trying to take or not. If you were trying to take out a drone using AI, using a laser or something, you’d be pretty okay making mistakes because you just missed the drone. Whatever. With the laser and the laser goes off, it’s all over. If you were doing something more sophisticated in a population area, have a densely populated area, you’d take less risk. So we’re developing procedures, tactics for each scenario. And this is part of the debate I had with Anthropic, which is we need AI for things like Golden Dome. Chinese hypersonic mixel comes up. You’ve got 90 seconds before it separates and all kinds of decoys and you don’t know where the actual payload is. And you want to get it, hit it from space. And a human can’t, doesn’t have the reaction time, doesn’t have the, may not be able to discriminate with their own eyes what they’re going after. That’s a pretty low risk thing because it’s in space and you’re just trying to hit something that’s trying to hit you. So I think in the next 10 years, you’re going to see a lot of these applications develop AI to one degree or another, so long as we think it’s safe and it’s not going to make mistakes. (Time 0:20:53)
  • Pentagon Found Anthropic Contracts Impractical For Warfighting
    • Emil recounts reviewing Anthropic contracts and finding blanket prohibitions on military uses, including planning kinetic strikes.
    • He describes months of negotiation and the impracticality of exception-based clauses for unpredictable future missions. Transcript: Emil Michael So if you remember the Biden executive order on AI, which was this crazy executive order that limited the amount of compute any model company could do and was essentially grandfathered In a small number of AI companies that they were going to designate the winners and everyone else was out so they could have more control on what they did. Anthropic was one of those winners. And then they were smart, actually it was a good sales strategy to sell into the most sensitive parts of the U.S. Government, like all of our combatant commands, CENTCOM, Central Command that’s doing the Iran fight now, the Indo-PACOM command, which is sort of responsible for China, several Of the intelligence agencies. And they did forward deployed engineers, Palantir style. So they got very sticky to the workflows and all that. And so I came in and I got the AI portfolio for department in August. And I said, I just want to see the contracts, you know, the old lawyer in me. And I looked at contracts. I was like, holy cow. They say you can’t use them for, you can’t use them to plan a kinetic strike. You can’t use their AI model to move a satellite. You can’t, there was a 20 page. You can’t do a war game scenario with it. You could do a scenario, but you can’t like, let’s suppose you’re writing a plan saying like, if this happens, here’s what we would do. And it might involve a kinetic strike, which causes harm to a human. So like, well, what do you think these folks do? You know, this is the Department of War. This is what we do. And so I said, okay, well, I’ve got to number one, have direct relationships with these companies, not just through Palantir, because I want to use it more broadly. And then number two, I need to have the terms of service be rational relative to our mission set. So we started these negotiations and took three months and I had to sort of give them scenarios about like these Chinese hypersonic missile example. They’re like, okay, we’ll give you an exception for that. Well, how about this drone swarm? We’ll give you an exception for that. And I was like, the exceptions doesn’t work. I can’t predict for the next 20 years what all the things we might use AI for. And so all lawful use seems like a good thing. If Congress wants to act, great. We have our own internal policies. We’ll follow them. We’re not knuckle draggers here. We don’t want to hurt people unnecessarily. (Time 0:42:42)
  • Supply Chain Risk Was About Dependence And Insider Threat
    • The Pentagon’s core fear was operational dependence and insider risk: a vendor could disable or poison models when relied upon in critical fights.
    • Emil emphasizes scenarios where model refusal at a decisive moment could endanger forces, prompting leadership alarm. Transcript: Emil Michael And then the trigger point was after the Maduro raid, one of their execs called Palantir, who we buy themselves 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 there was use in that raid, that that might violate their terms of service. So they wanted to enforce, this is very important here. Jason Calacanis Yeah. They wanted to enforce their terms of service. They went behind your back to try to collect information to then maybe pull your license for their technology. Emil Michael You know, it wasn’t behind my back. I don’t want to accuse them of that. Palantir is the prime contractor, the sub. But it raised enough alarm with palantir who’s got a trusted relationship with the department to tell me and i’m like holy 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 so i went to secretary heggeth. I 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 ability 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 at the time or sort of trick you because you have to trick it. I mean, all these things that we know we’re aware about models or hallucinate purposefully or not follow instructions, like some insider threat stuff. So then that culminated in the Tuesday kind of dramatic meeting with Secretary Hexeth and, and Dario with the Friday deadline that got blown. (Time 0:46:22)
  • Model Policy Choices Create Systemic Market And Security Risk
    • AI platforms can act as de facto gatekeepers whose policy choices have geopolitical and commercial consequences.
    • Emil warns Anthropic’s constitution-like policy and potential for rapid shifts make single-vendor reliance risky for government and industry. Transcript: Chamath Palihapitiya That is the huge thing that this thing tore open this weekend. So if you’re not figuring out how to be multi-model and agnostic across these models, you’re taking on enormous business risk after Friday. Because you can’t tolerate that these folks will do that. It’s too critical of a technology. By the way, this is deplatforming all over again. Remember what happened when you didn’t like what was said, now all of a sudden you were deplatformed? This is that times a thousand. Because this is not about posting on social media. This is about using fundamental technology to either advantage or disadvantage your business. Emil? (Time 0:50:15)
  • Avoid Single Vendor Lock-In Build Multi-Model Redundancy
    • Build redundancy across multiple large models and avoid single-vendor lock-in for critical systems.
    • Emil wants at least two main providers and emphasizes enterprise integration and forward-deployed engineering as differentiators. Transcript: Emil Michael I want to see if they diverge or not, or if they converge, maybe I only need two over time, but we don’t know. It’s too early. David Friedberg But why keep them in the mix, Emil? So if there’s clearly a difference of operations and philosophy and how they want to run their business, and there’s other models, is their model particularly good at particular applications That make it important to keep it in the mix, given that there are three or four other kind of alternatives here? Anthropic, you mean? Yeah. Emil Michael Well, because the number one reason we were having this conversation at all was because they were deeply embedded. So now I have to un-entangle them. And the other companies have not gone as heavy enterprise, enterprise sales, forward deployed engineers, government business. So they have to catch up, not necessarily the capability of the model, but just how do you serve the government? And probably just way ahead on that. David Friedberg But the models themselves, you don’t think are uniquely advantaged? Or do you have a view on that at this point? Emil Michael I don’t have a view on that. I don’t think they’re, you know, certainly cloud code was innovative and ahead. That’s true. But dude, I believe in 12 months, Codex is not going to be close. (Time 0:54:04)
  • Pentagon Funding Push To Onshore Critical Defense Supply Chains
    • The Department is pushing to rebuild domestic supply chains for critical defense components using a $200B lending office to onshore minerals, batteries, and production.
    • Emil expects critical minerals solved first, then multi-year efforts for batteries and manufacturing plants. Transcript: Emil Michael We are early days, critical minerals, you see, you’ve seen the action around that. You’ll start to see so I have the Office of Strategic Capital, which has $200 billion in lending authority. And what we’re trying to do is, it’s like treasuries plus 100 bps, loan to companies, show them that the department needs their solid rocket motors, their batteries, their fiberglass, Like all the things that we’re heavily dependent on for our defense industrial base that are completely outsourced to China and domesticate them here. And we’ve got a bunch of great people running it. So, but it’s early days. It’s going to take for the rest of the term to get, I think we’ll get critical minerals done before the rest of the term where we have the access to what we need to from us or allied countries. But from batteries, it’s like the next problem I’m trying to solve, for example. Batteries are totally outsourced both technologically and from lithium to China. And there’s like, you know, kind of call it 20 critical things. If I could get to all of them at some level, but then it’ll take a few years for them to like build plants and do that stuff. But it’s very important. I hope whatever administration comes next continues it because I’m all free market. (Time 1:11:33)
  • Use Outcome Contracts To Let Startups Rapidly Deliver Defense Tech
    • Reform requirements and adopt commercial-style, fixed-price contracts to speed innovation and scale startups into defense suppliers.
    • Emil urges minimal upfront specs, outcome-focused asks (e.g., missiles with range/blast), and penalties/incentives tied to delivery. Transcript: Emil Michael Yeah. So we did a big… So part of it comes down to requirements reform. What used to happen is people are like, oh, we need a new fighter jet. So Army, Navy, Air Force put into the requirements and we needed to be stealthy, to hold a missile, to hold four humans. And it became this unbuildable thing. But the contractor didn’t care because they’re getting paid cost plus. So like, sure, I’ll fulfill your requirements. Two years from now, you’re like, that was never engineered properly. It’ll be another few years late and a couple more billion dollars. So we’re trying to change that to, I tell you my common operational problem. I need a bunch of missiles that go 500 miles or more that have this kind of blast. Come to me with solutions. As little requirements as possible on that side. And then the contract piece, trying to get to as close to commercial contracts as possible. And this is where the startups are so good. They’ll do fixed cost pricing. You don’t pay me as much if I deliver late. You pay me more if I deliver early. Jason Calacanis It’s very disruptive to the existing system. (Time 1:16:15)