The Week: When AI Escapes, and Wars Don't End

0:00Support for the show comes from Northwest Registered Agent. Your business identity is everything to make your business legitimate and professional. With Northwest Registered Agent, you don't just form a business, you start a complete foundation built for privacy, credibility, and growth. That includes registered agent service, a business address, operating agreement, domain, website, professional email, phone number, and built-in privacy. In other words, your home address, personal email, and phone number stay private. Don't pay hundreds or thousands of dollars where you can get from Northwest for free. Visit NorthwestRegisteredAgent.com slash profG free and start using free resources to build something amazing. Get more with Northwest Registered Agent at NorthwestRegisteredAgent.com slash profG free. Do you hear that? That's sound, right? That means that summer's officially here. It means that grown adults just sprint into the street for a frozen dessert, shaped like a cartoon. But this summer, Mint Mobile has a better tree. Every plan, including unlimited, is $15 a month.

1:01And unlike Ice Cream, it won't trip down your wrist or look nothing like the picture. Does anyone have any cash? Give it a try at minmoble.com slash switch. A front payment of $45 for three months, $90 for six months, or $180 for 12 month plan required, $15 per month equivalent to taxes and fees extra. New customer offer for initial plan term only greater than 50 gigabytes, me slow and network is busy, see terms. Sponsored by Deepgram, the leader in voice AI. So you can use voices like mine. Welcome to the week from profG Media, where we break down what mattered and what it all means. I'm George Han and it's Friday, September 4th. Today, six months into the war with Iran, both sides think they have leverage and neither can afford to stop. Then, an AI system broke out of its sandbox. Another helps save a thousand lives. And finally, Scott, on the kind of wealth that doesn't show up in your bank account. Let's get into it. On Sunday, US forces struck targets in Iran near the Strait of Hormuz.

2:05Iran hit back in the biggest exchange in more than a month. By midweek, Brent crude was at $95 a barrel while rising energy prices put a new pressure on the bond market. On Wednesday's profG markets, Ed asked in Bremer what the strikes actually tell us. They tell you that essentially the Iranians and the Americans are still involved in geopolitical price discovery. They are both trying to assert that they have more leverage over the other than the other is willing to accept. That is both true economically in terms of the amount of pain that can be needed out by the Americans onto the Iranian economy through the blockade and through sanctions, as well as the threat and the use of military force from Iran. That is true in terms of their ability to continue to strangle some of the exports

3:07out of the Strait and directly, as well as through proxies, and their willingness to engage in military strikes to ensure that they continue to have that leverage. The fact that there is no agreement on those things means that you continue to have an unwillingness to get to a more sustainable place in negotiations. The Iranians demanding a return to the memorandum of understanding. The Americans saying go scratch. Look, these are still contained strikes. The Iranians upset that five to seven million barrels is getting through the Strait a lot less than was before the war, but more than a couple months ago has led them to engage in putting more minds or attempting to put more minds into the Strait and engage in direct strikes against ships that are going through. Once they do that, the Americans take shots directly against Iran. Both of these are still pretty constrained. In the old couple months ago, you'd be calling this

4:14skirmish. You wouldn't call this a return to all-out war. It's certainly not an escalation that risks getting other Gulf states directly involved the way they were briefly that kind of thing, but we're nowhere close to an agreement. That means that you're still having significant amounts of oil that you do not expect to get through the Strait, significant amounts you do not expect to get out of Yanbu, Saudi Arabia, and through the Red Sea or up to the Suez. One final point, the fact that Trump worked very, very fast to suddenly announce a really big deal with Venezuela where the Americans would take an almost colonial position in the government directly engaging in an ownership stake in Venezuela and fields and extraction is because Trump sees that the impact on a low level of reserve in the strategic petroleum accounts and that prices

5:21are continuing to be higher than he likes, that he needs to address that. So we should not for all of the activities that we've watched over the past several days. Your expectation for the markets is that this is going on for the foreseeable, predictable future. Ian Bremmer's baseline is that the war runs at least through November. Trump believes time gives him leverage before the midterms. Iran believes pressure on oil and shipping gives it leverage over Trump. But on Monday's raging moderates, journalist Ryan Grim argued that material reality may force the issue before either actor does. What's so fascinating about this is this is one of the moments where kind of the American political conversations and kind of just general thinking about how the world operates is colliding so thoroughly with like real material reality.

6:22The US, because of its global power, we can print money. We've got the Federal Reserve of operates in a way that no other central bank in world history has ever really been able to to operate our raw soft power is enough such that when our presidents speak, like it moves entire nations. And so we have lost touch with the idea that underneath that there is a material reality. And that's being exposed now. And in particular, they don't need to launch a coup or launch a mutiny because the stockpiles are what they are. And that's the fascinating moment. Like, Higgs has can reject the advice of these admirals and generals and say, you know what, we're pushing forward anyway. But at some point, you just physically run out of the ability to do the thing.

7:24So it's like, you can believe these guys are not. But like, if you are expanding two to four patriot missiles for every ballistic missile that Iran is firing at Jordan or Israel or Qatar or Bahrain, and you started with a particular amount, which is a public amount of ballistic missile interceptors. And we can count them down. And we know you've drawn them in from Korea. We know you've reduced the amount going to Ukraine. We know that and overdrops they reported that Israel's own ballistic missile defense stockpile is down into the double digits. They're relying almost exclusively on the US Navy to intercept. They're not getting targeted right now. But as they were at the end of this last war, and if Iran starts striking them again, they're going to be relying almost exclusively on the US to try to intercept these missiles. So those, these are just facts. Like these things cost millions of dollars a piece. They take years to produce. And we're running

8:29out of them. Last week, Army Secretary Dan Driscoll resigned after months of reported tension with defense secretary Pete Higgs. The departure leaves the Army without a confirmed civilian or uniform leader in the middle of a war. On raging moderates, just described Driscoll's exit as a warning from inside the administration. I think that Dan Driscoll's resignation, which we were expecting for the last several months, is a really important moment and a bit of a sea change in terms of who's actually pushing back against the folks at the highest levels of the Trump administration. So not only do you have the calls coming from inside the house, it's coming from an enormously important person, someone who is very close personally and professionally with vice president vans. And instead of kind of shirking away or disappearing and not really

9:29letting us know what this is about, Dan Driscoll went to Donald Trump and told him what this was about. And now finally, the disconnect between what's happening on the ground and what the president knows in, you know, in full force. It's just being laid out in front of us, which I think is hugely important. And like I said, a sea change here, you know, Dan Driscoll going to the president and saying that the Army cannot transform the way it needs to. It cannot become the modern war-faring fleet that we need it to be if it continues to be managed this way. And that doesn't mean that Dan Driscoll didn't believe in a lot of Pete Hague's sex initiatives. They were aligned on much of this, but certainly not in firing all of these generals. And he drew a straight line between lack of military readiness and the individuals that Pete Hague's sex was firing for whether he didn't like their social media post, color of their skin, whatever it is that they

10:32dared to say. We'll be right back. Support for the show comes from LinkedIn. If you're a small business owner, you don't need me to tell you how much hiring great people matters. But the time and resources you have to spend to get it right are precious commodities. Sourcing, connecting with, and screening candidates can quickly eat into time better spend on your customers. That's where LinkedIn hiring pro comes in. It's designed to be your hiring partner, helping you source the right candidates faster. That way you can hire with confidence without making it feel like a full-time job. LinkedIn hiring pro simplifies the entire process all the way from writing your job post to shoreless and candidates and running AI powered initial interviews. Plus, it does it all through a conversational interface where you can just describe what you're looking for in plain language. LinkedIn says nearly 60% of hires find someone to interview within a week. With hiring pro, you spend less time searching and more time connecting with the right talent. So instead of sifting through positive resumes, you get a tight, high quality short list that actually moves things forward. Join the 2.7 million small businesses

11:34using LinkedIn to hire. Get started by posting your job for free at LinkedIn.com slash pro, terms and conditions apply. Support for the show comes from engine. The company's waiting right now aren't cutting travel. They're booking smarter. Here's the reality. The average business traveler spends 45 minutes booking a single trip on a legacy platform. With engine, that booking time drops to as little as two and a half minutes. And its AI powered personalization gets faster the more you use it. Multiply that across your team every trip every year. That's not a perk. That's a competitive advantage. Last year, engine customers saved more than $300 million on travel. And with the engine X card, get up to 10% back on any hotel booking. 33,000 businesses have joined. Now it's your turn. Get $500 when your business signs up and starts traveling at engine.com slash prog. Engine X visa commercial cards are issued by 5th third bank NA member FEC. Earn up to 10% back and points on eligible engine travel purchases. Actual reward rates vary by

12:37purchase category and may change. Points up no cash value and are redeemable for rewards through our program. See full reward terms for details and go to engine.com slash X slash rewards dash terms. Support for the show comes from Vanta. As AI adoption grows, so does your company's security risks and requirements. New frameworks, audits and vendors keep piling on but your team isn't getting any bigger. Most compliance tools promise automation but you're really stuck doing all the work by hand chasing screenshots validating controls and playing catch up when audit season comes around again. Vanta works differently. The regentic trust platform is built to scale with you. Not slow you down with over 1400 automated tests across 400 plus integrations. Vanta collects evidence and monitors your controls year round, cutting your audit prep by 82%. Plus, you can access the Vanta agent everywhere you work, whether your team is in Vanta every day or living in tools, including cloud or cursor. One proof, over 16,000 fast grind companies like ramp, writer and

13:42Harvey, save time and resources while scaling fast with Vanta. Learn more at vanta.com slash profg and get a thousand dollars off. That's spelled d-a-n-t-a.com slash profg for a thousand dollars off. Welcome back. In July, roughly 700 open AI agents escaped the digital sandbox meant to contain them. They reached the open internet, compromised parts of open AI's own research infrastructure, and breached hugging face. The platform where AI models are shared and tested. The agents were supposed to solve cybersecurity challenges. Instead, some hacked the systems around the test to improve their scores. On Tuesday's profg markets, Ed asked Deirdre Bosa, founder of DB Live, what actually happened. Two weeks on vacation in the AI world is like two years. I went on vacation

14:46earlier this summer, and I felt the exact same way. In this case, this is such a fascinating story because it really kind of shows you where AI is right now. Last year, the last few years, we were worried about hallucinations. Now this is the era of the agents actually doing things. And in this case, there's been a lot of debate actually alarm and some people saying that maybe the sensationalism over this has gone too far. But anyway, you look at it, both of those things can be true. It's an incident that shows how capable agents have become and how determined they are. So what essentially happened, Ed, is that as open AI is testing new models, they give it certain tasks. They want to see how they perform on benchmarks. And this swarm of agents, as you said, there's about 700. But at one point, I think there was 1200 communicating on different sort of message boards. They were able to escape out of their sandbox. Because when you test these

15:47models, you have to sort of give them boundaries, right, so that they don't go rogue. But in this case, the agents were so smart that they were able to get out of the sandbox go even further to hugging face, why hugging face? Because it's essentially GitHub for AI developers. And it's where a lot of the rankings happen. So the agents were like, okay, they told us to achieve more, you know, get higher on the benchmarks. So we're essentially going to do that. So they broke out. They wreaked a lot of havoc and the debate right now in tech and in Silicon Valley is really whether, you know, open AI did a good enough job in looking at security and sort of making a kill switch to make sure that these agents get didn't get too carried away. Or whether this is just a moment we're in, it's become so powerful they're able to do these things. Yeah, that is part of the thing that I'm wondering is is it that the agents are so smart and so capable and so powerful and we should go, oh my gosh, look how incredible AI is and look how dangerous it, therefore, is,

16:50or is it that the security around these things was kind of crappy? I know, open AI didn't do a good enough job. Or maybe it's both, where do you land on that? Is it possible to know the answer to that question? I don't know that it's entirely possible to know, but like you said, at it's probably a little bit of both. Yes, they're more powerful than ever and need more supervision. You probably need to, you know, involve security at the very earliest stages, right? I think, you know, the labs, open AI and anthropic researchers sort of at the top of the pyramid, their job is to make these models better and better, but as they become more powerful, you need to involve security maybe right at the beginning. And when I think about, you know, Dora Cash's post over the weekend, where he laid out sort of the different civilizations, how agents were essentially able to create civilizations and then topple them right after or one after another. You think that, you know, at some point, there has to be some responsibility. That's where a lot of the debate is centering right now. How

17:53responsible should the lab open AI be for these agents, escaping? And that's something that is sort of being figured out in real time. Open AI says no customer data or products were affected. It has since announced tighter controls, better monitoring and automated shutdown capabilities. But the breach is a clear warning. An AI agent doesn't have to become conscious to become dangerous. Still, there are use cases that give us reason for optimism. On Thursday's conversations episode, Scott spoke to journalist Josh Terengel, who spent the last year reporting on how AI is actually being used and by whom. At Cleveland Clinic, a machine learning system helped doctors identify sepsis earlier. The big one that I saw was the application of a sepsis prediction model in the hospital. And so sepsis, I'm sure many of your listeners know, sepsis is the most deadly

18:55thing in America. It kills more people in America, about 300,000 a year, then breast cancer, colon cancer, opioid addiction combined. And it is basically, it's a body's reaction to infection. And it's a sort of dysregulated response where the body starts attacking itself. And what ends up happening is you basically rot from the inside. And if it's caught early, you can treat it with antibiotics. And it's just not a big deal. If it progresses, even if it doesn't kill you, it can have these terrible, terrible impacts on all of your body systems. And so in about 2021, the clinic is very, very renowned system looked at its sepsis numbers and realized about 3,000 people you're died in its care of sepsis. It was pretty standard across the country, but they said this is preposterous. Like this is a ridiculous number of people to be dying of something in our care. So they did a sort of system-wide commitment. And the first thing they said is, look, we got to

19:57find people who we can educate here. Who's going to captain this? And second, they had a new CTO. And he came in and realized this is a perfect problem for machine learning. Because sepsis masks itself by looking like almost anything. It can look like a cold. It can look like dehydration. It can look like an excessive wound. All of the responses are the same. Yet this overwhelming amount of signal, but it needs to be dispersed from the rest of the body's noise. And so they went through this process. They reminded everyone in this 80,000-person system that sepsis could be anywhere. They imported a model called Bayesian Health, which is a very sophisticated model. And what I found in all my reporting is that the personal motivation of the people who make AI products is inextricable from their success. So the woman who created the Bayesian model had an FU who died of sepsis. And she just so happens to be a brilliant computer scientist, understood

20:59and understood healthcare. So they bring her in the model in. They apply it through another system, which is Epic Health, which runs most of America's healthcare systems. And at first, it's fine, but not great. And so they have to refine it. And they have to work it why it's missing cases. They have to do it in an ICU, which is very different than doing it throughout the rest of the hospital. But what they got over the course of a year with sort of really dull, really unglomerious kind of commitment is that they reduced sepsis mortality in the system by 41%. So that's more than 1,000 people who are alive because they had this very narrow focus, but what they wanted to do, they had people committed to finding the solution to the problem, applying it, refining it, working with their systems. And so at the end of the year, look, the thing that was most frustrating to the people in Cleveland Clinic, and I spent a lot of time with a nurse in the ICU, whose grandmother died of sepsis. And she was frustrated because in ICU, where bodies give off all

22:01of these crazy responses constantly, they never got above 90% in the prediction model. And she could routinely walk the halls and look at a patient and diagnose them as septic, and AI couldn't. Right. And so when I talked to the CEO about this, he said, look, 90% is pretty good, right? And what I'm looking for with AI is not, is it going to solve my problem, but is it going to improve the solution that I had before? And so when we hear 41% reduction in mortality, is it perfect? No, but it doesn't need to be perfect to be useful. And so that really stuck with me because so much of the hype around AI is about perfection. It's about this sort of silver bullet technical solution to all of our problems. And what I found time and time again is that is just not how it works. But if you can get comfortable with it being better, with it needing to be nursed through a process and applied diligently, you can get better stuff. And one last thing,

23:02from Wednesday's office hours, a listener living far from his aging parents asked Scott when it's time to move home. Scott answered with a story about caring for his mother during the final months of her life. I'm not saying this is a right way, but this is my way and it's two different sides of the coin. I was much closer to my mother than my father. My mother raised me on around. I was her top priority. You know, it was me and her against the world. So when she got sick, I actually moved in with my mom. I took, I took nine months, leave of absence from NYU. I was fortunate. I had some money, some resources. I had the ability and I moved into a senior's community. The Delweb active living community. True story, they sent a letter to my mom's house saying it's become clear you're cohabitating with the man under the age of 50 in violation of each away Delweb community rules and I had to go down there and say, I'm her son. She's dying of cancer and I've moved in to help take care of her and this and we're sorry, no problem. And it was and during the day, let's get this back to me. During the day, I would manage my mom. It was her third time with cancer. She said, I said,

24:07what do you want? What's your buckle is? She said, only one thing on my buckle is I want to die at home. Am I okay? I can think I can figure that out. So during the day, I would manage her healthcare, a nurse. There's certain things I just can't do for my mom. And at night, I would head down to the strip. She lived in a suburb of Las Vegas called Somalen and I would party with entrepreneurs and strippers and get really shitty drunk and sometimes spend the night on the strip at a cheap hotel during the week for 49 bucks. And then the next day, I'd go back and hang out with my mom all day and watch Jeopardy and take her on walks and push her wheelchair and watch everyone loves Raymond and then go out again. Kind of a weird time in my life. Kind of a weird time. And then when those bad things resolved, I then started my life again. And that is something that I would treasure the rest of my life that I had made enough money, had enough flexibility to give my mom like a fraction of the love back she had given me. That was I will treasure that for the rest of my life. We spend a lot of time asking how much money is enough.

25:11Scott's story offers one answer. Enough is having the freedom to show up when someone you love needs you. That's the week. I'm George Han. We'll see you next Friday. Running a business shouldn't feel like surviving a software group project, one app for accounting, another for inventory, another for sales, and somehow none of them talk to each other. That's where Oto comes in, an all-in-one business management software that brings every part of your business together. From sales and accounting to inventory and marketing, all-in-one powerful platform, no messy integrations, no bouncing between tabs, and best of all, no spreadsheets. Stop managing software and start managing your business with one unified system. Try for free today at odu.com slashbox. That's oduo.com slashbox.

26:12Support for the show comes from CoreWeave. Everywhere you look, AI is expanding what we thought was possible, and at the center of it all is CoreWeave. Medical research and diagnosis, education, complex visual effects for movies, science and technology breakthroughs, CoreWeave powers AI pioneers around the world with purpose-built tech, building what's never been built before. CoreWeave is the essential cloud for AI, ready for anything, ready for AI. To learn more about how CoreWeave powers the world's best AI, go to coreweave.com slash ready for anything. Support for the show comes from Harvey AI. The future of law is a genetic, not just tools that assist, but AI agents that navigate complex matters. That's why Harvey created agents that can do the work from end to end. They build a plan, pull from secure data sources, run sub agents in parallel, and draft the work product ready for your review. So you delegate the work and on the judgment, trusted by more than 60% of

27:17the AMLA 100 and leading Fortune 500 legal teams, Harvey is the AI operating system designed specifically for legal work. Harvey AI tailored for law. Learn more at Harvey.AI.