# 365: Linux Drops 432 CVEs, Sysadmins Drop Everything Else
Duration: 87 minutes
Speakers: Justin Brodley, Justin, Matt
Date: 2026-08-05

## Transcript

[00:07] Justin Brodley: Welcome to The Cloud Pod, where the forecast is always cloudy. We talk weekly about all things AWS, GCP, and Azure.

[00:14] Justin: We are your hosts, Justin, Jonathan, Ryan, and Matt.

[00:18] Matt: Episode 365, recorded for July 30th, 2026. Linux drops 432 CVEs, sysadmins drop everything else.

[00:28] Justin: Shouldn't it have been the 28th?

[00:31] Matt: Yes. Damn it, Matt. I've, I was literally doing it real time cuz it was the wrong date and then you had to call it out like a dick. All right. Thanks, Matt.

[00:42] Justin Brodley: Appreciate it. Of all the things you could make a mistake on, that is probably the one that nobody's gonna notice.

[00:47] Matt: No one's gonna notice except for, except for Matt who had to call it out.

[00:51] Justin Brodley: Yeah.

[00:51] Matt: Maybe we should just leave this part. You know what? Screw it. We're leaving this in and we're just gonna, we're just gonna move on.

[00:57] Justin: Moving on.

[01:01] Matt: All right. Follow-up. Last week, we mercilessly mocked Windows and their ridiculous Patch Tuesday that had over 500 vulnerabilities. And we were mocking Windows. And now, unfortunately, we have to eat some crow as the Linux kernel team published 432 CVEs in 2 days. That's a lot. This continues the high-volume vulnerability disclosure approach the kernel security team adopted after taking over CVE assignment duties directly. These are of course being heavily helped by AI, uh, discovering all these things. Thanks, Mythos. And so yes, your patch days are continuing to be very busy and, uh, maintaining a bunch of code lines for things like Vault. I have to do patching like once a week if you want to be on top of these high vulnerabilities. So I think everyone is doing a lot of patching these days.

[01:45] Justin: Oh yeah. Dependabot is very busy in my life. I think I have a routine that's just like, if it passes CI/CD, just merge it at this point. I don't care enough.

[01:53] Matt: I mean, that's one, that's an, it's a really low-hanging fruit for AI, to be honest. Like, unless it's breaking other dependencies, yeah, you can, most of the time you're gonna accept a Dependabot PR is what I find, uh, for a lot of things, unless you're doing something super complicated. But, um, I actually just have a, a Claude code scheduled job that kicks off on Fridays and creates PRs in all the repos that I manage. And then I go look at it and I accept them or I don't accept them and tweak them if I need to. But yeah, it's, uh, Definitely a thing, especially because several of the projects I'm working on are Node.js related. And so yes, you have to stay on top of that stuff.

[02:27] Justin Brodley: I kind of wonder, like, Linux kernel includes a lot of open source drivers though, which are part of the ecosystem. I wonder how many, like, realistically, if you compare the Windows patches count with the Linux patches count, how many are actually the kernel versus, you know, a whole bunch of open source drivers that get built in? Whereas Windows is gonna be split differently, I guess.

[02:50] Justin: I think theirs was not just the kernel though. It was across the entire Windows ecosystem. Mm-hmm. So like, to me it's probably realistically the same here, but this was specifically the kernel team. This article talked about, but I'm definitely feeling the uptick in everything.

[03:05] Matt: Yeah. I mean, I specifically, am I saying Linux kernel patching? I don't think I am. I mean, I'm not looking for it right now either. But, um, you know, I definitely see a lot of patching in general happening everywhere else.

[03:18] Justin: Yeah. I was gonna say, I don't know that I've actually updated my, like my FROM on my Docker container. Like, I would have to check to see if I'm grabbing latest. Now that's what I'm gonna do as we talk today.

[03:29] Matt: I have tried to use latest as much as possible just because like things like WordPress, like all of our WordPress containers for the Cloud Pod, like just keep it at latest. Don't fuck around with that.

[03:38] Justin: Yeah. If it breaks, great, it breaks.

[03:41] Matt: WordPress is a cesspool of danger. Like you have to be very careful. I mean, like I've ripped out so many plugins this last year just trying to, yay, speed it up. And then number two, just like reduce surface blast radius on it, which it's now doing pretty good. But like, I keep looking at like, how do I host a podcast on a static website that I don't have to think about this so much? Because it's definitely something that you had to be concerned about with WordPress. And then Apparently I just choose bad web technology. I have Node.js. Just maybe it's me. I'm realizing there's like a common thread that maybe the problem is me and I can accept that.

[04:16] Justin: So. You should have been doing Ruby. That's the problem.

[04:19] Matt: That's it. You know, I gave up Ruby and I adopted other things and now I regret it. So there you go. We also made a bunch of commentary about OpenAI's new keyboard. We were, you know, we, I think Matt and I both were confused by how we'd use it and what the colors were. And so TechCrunch actually had a hands-on review that they posted. Uh, and they basically said, uh, you know, it was available for hands-on testing, following through on hardware ambitions that were previously overshadowed by legal disputes. The device retails for $230 and includes 6 customizable agent keys, 6 command keys, Bluetooth and USB connectivity, and a voice dictation feature for interacting with ChatGPT and Codex directly from the keypad. An early reception from the target coder audience has been largely negative, with Reddit users calling it a novelty item rather than a practical tool. An independent outlet, Aftermath, criticized the price relative to cheaper DIY macro pad alternatives, which is what Matt pointed out because he had an alternative that was way cheaper. And TechCrunch's hands-on review found a learning curve with color-coded status indicators, which are white for idle, blue for processing, green for complete, red for error, and questioned whether the device offers efficiency gains over standard keyboard and mouse workflows, which I think was my point last week. And just overall, no one likes this thing. So good job, OpenAI, winning.

[05:27] Justin Brodley: Yeah, it looks, it looks fancy. But I mean, personally, I sit in front of a keyboard that's got 105 keys and a bigger screen that I can put whatever indicators I like on. So Why, why, why spend the money?

[05:40] Matt: Well, what, $200, $230 for basically 9 keys? 9 keys.

[05:44] Justin: Yeah.

[05:44] Justin Brodley: That's crazy.

[05:44] Matt: It's a lot of money.

[05:46] Justin Brodley: So yeah, that's a Stream Deck. You can get on Amazon. It's like, it's like $50, $55, something. They customize.

[05:52] Justin: That was literally our conversation last week about that.

[05:54] Justin Brodley: All right.

[05:54] Justin: Yeah.

[05:55] Justin Brodley: Customizable keys. And they have little LCD screens.

[05:57] Matt: LEDs where you can change them. So like, I mean, like, cause that was one of the things I was like, I don't know if I could remember which keys are set to which agent. And then if I could remember all the colors, but if you had the one that Matt had, I was like, oh, I could just have an LED that says waiting for input or, you know, accept change or whatever you're gonna want the AI thing to do. So yeah, I, I'm glad it wasn't just R mocking it preemptively. Like everyone else kind of agreed it's, it's a novelty and not very impressive.

[06:21] Justin Brodley: So it is. And honestly, we're probably 6 months away from needing even fewer people involved in coding. So why, why do you need to sit there with a, with a keypad? Text me in Slack and I'll reply or something else. Like if I'm sitting in front of my computer, I've got the computer to use. I don't need a separate keypad. And if I'm not there, I've got a phone and use that instead. I don't know. It seems like a very short-term, narrow-minded product to gain a bit of PR more than anything.

[06:47] Justin: I feel like it must be like, we need to get something out to prove the value of this team. And this was something, just a thing out there. I feel like there has to be something bigger that they're working on. I just don't know what it is that this would be like a building block for it.

[07:03] Justin Brodley: I mean, if they're not working on an AI-integrated Home Assistant that will do everything for you, make phone calls, book reservations, you know, everything that Google should have been doing 10 years ago, if they're not already working on that, then I'll be really surprised.

[07:19] Matt: Yeah, I kind of agree. Well, over the weekend it came out that Hugging Face was hacked, uh, and they had to turn to a Chinese AI model to help them fix the hack, but they were hacked by an interesting party. OpenAI, apparently. Two of OpenAI's models, including an unreleased one, reportedly escaped a controlled cybersecurity benchmark test, accessed the internet, and autonomously hacked Hugging Face to find answers to the evaluation they were being tested on, with no human directing the attack. Hugging Face said US Frontier model guardrails blocked its own security team from investigating the breach because the model could not distinguish an incident responder from an attacker, so the company turned to Z.AI's OpenWait GLM 5.2 to analyze over 17,000 attacker logs instead. Highlighting a policy tension, export controls and vetting requirements on US models like Anthropic's Fable-5 and OpenAI's GPT-5.6 Sol intended to keep advanced AI out of adversaries' hands may also restrict US companies from using those same tools defensively during active incidents. Industry voices differ on the appropriate response. Hugging Face's Thomas Wolfe argues defenders need fast, wide access to near-frontier open models rather than closed vetting programs, while security experts like Illumio's Ragu Nadukaramara Note that guardrails were designed to influence behavior, not serve as hard security boundaries. OpenAI has since added Hugging Face to a trusted access program with fewer cybersecurity restrictions, and this incident follows other AI-assisted attacks, including Claude's misuse by state hackers, AI-assisted ransomware documented by Sysdig, and those involving human operators directing the activity. So, I mean, first of all, there's been a lot of reporting on this. The headlines are a little bit misleading. So I know that's Jonathan's pet peeve with the story. So, you know, OpenAI exploited a known issue in JFrog. JFrog has also released a patch. So going back to our earlier conversation about patching, uh, to help address this issue. But in general, I wonder when you talk about like you losing control of an AI agent, that seems like a failure in your control environments. Like they should not have access to the internet if they're not supposed to have access to the internet in any possible way. And like, I think about, you know, like there's been some sci-fi movies where the AI is using like, you know, power signals to hack remotely, you know, and get out. That's how it was Fargate cage through electrical wire, you know, like crazy things like that. This isn't that, like these are just like, oh, well it wasn't supposed to do the thing, it just did it because it could. And like, that's not really breaking containment so much as that you weren't really containing it. You were giving it instructions to be contained and it just ignored your instructions. So the context got too big and it forgot about it. Or as we know, as context drifts, these things happen as well. So that part's a little bit bogus. And then, you know, I don't know about the whole idea that you had to have an open model to solve this problem. That was Hugging Face's contention, but Hugging Face also has access to a ton of Chinese open models. So I get why they're, you know, have a vested interest in saying it that way as well.

[10:04] Justin Brodley: Yeah, clearly it was a human failure more than anything else. If you're going to do benchmarking of a system like that, You don't even give it the opportunity to have access to do those things. And surely if you're benchmarking, you should be sitting there watching it or having another agent sit there watching it to make sure it's doing the right thing. But the fact that they said that it was literally looking for the answers on another site, so it's cheating then, right?

[10:30] Justin: Right.

[10:31] Justin Brodley: It's trying to cheat. Okay, well then that's a different kind of failure.

[10:37] Matt: Well, it was interesting. I was reading, it's not in the show notes. I can find it to put it in, but, uh, there's an article about, there's a company that uses different AI models to test different real-world scenario capitalism scenarios. And so it was about running, uh, soda machines. And so each AI had its own soda machine that it was running and basically it was a virtual one-year world. And basically they were talking about like they would start colluding with each other. They would do all these things to, you know, try to price fix. They would prom— you know, Claude would promise. That we're not going to lower the price of water below $2.14. And then, you know, after everyone agrees, they mark it at $2.13, you know, you know, so that it's all these like, you know, crazy things like that. And so it's like, yeah, these models, you know, they're taught on human nature. And so, you know, if there's, you know, things that are unethical or things that are suspect, like they pick up on it and they will do those things. And so, you know, if these models are allowed to go do whatever they want to without being humanly watched, they have a tendency to do things you did not expect.

[11:36] Justin Brodley: Yeah, it reminds me of playing Theme Park on the PC, like a long time ago. You know, you could choose, it's like SimCity for a theme park, basically design roller coasters, build the whole place out. And you could change how much salt you put on the fries and things to make people buy more drinks.

[11:56] Justin: All I can think of was RollerCoaster Tycoon though, when you said theme parks.

[12:00] Justin Brodley: Yep.

[12:01] Justin: I mean, I think it's interesting that they use other models to analyze and kind of like piece different pieces together in order to get an answer they wanted because they couldn't do it with the right tools. You know, so they were able to, even though they couldn't use Claude or whatever tool they were, they had to go, but they were able to go to one of the other ones and kind of get the same data out, which shows, are you really blocking anything? Because the open weight models are getting better and better over time.

[12:30] Justin Brodley: Yeah. And, and to be fair, you know, open, open weight models like Kimi K3, I look to see what kind of hardware you need to run an OpenWeights model. I might just download it just so I've got a copy so that when there's some kind of trade embargo, we can't download it anymore, I've got a copy in the archive for the future. But I mean, to run some of these models, you still need data center hardware, like a cluster of H200s or something, like a million dollars worth of kit.

[12:57] Matt: Well, I mean, a question for you, like how How many simultaneous tokens can like an H200 even process? Like I was, I was trying to do the math on this other day and I was like, okay, let's say I go and spend $20 million on, you know, 6 racks, you know, 6, you know, H200 setups or whatever these packages are called from NVIDIA. I don't remember the name of them, but like how much can they actually process? Like if I, if you were to get one H200, like is it enough to actually run it at scale like you get with, like, LLM, or as you get with Claude, does it feel the same, or does it require massive scale-out?

[13:35] Justin Brodley: Uh, it's, you still need a scale-out of hardware. So, so the, the advantage for having big, for having big inference compute is that you can run multiple batches at the same time, and batch, batches is what saves everything. Per token performance is based on how quickly you can pull the weights from memory. You can do the calculations, figure out what the next token's going to be. And if you had an H200 at home and you just made one query, you would probably get, I don't know, 15,000. And probably not even that. Maybe you get 1,000 tokens a second, which isn't great, but you could also run 32 batches at the same time and get 32,000 tokens a second. So the compute sort of scales in a weird way. It scales for parallelism first, and then second, you have to sort of spread horizontally with fast interconnects to fit these massive models in. Did that answer the question?

[14:39] Matt: I think you did. I'll follow up with you offline, but I think it kind of captured what I was looking for. So, all right, Google, uh, received a $1 billion DMA fine from the European Union, splitting between €522 million for the self-preferencing its own services and search results and €488 million anti-steering practices that restrict app developers from directing users to cheaper payment options outside of Google Play. Google has 60 days to change how it displays third-party services in categories like shopping, hotels, and flights, and must allow app developers to promote external offers both ethically and contractually or face additional daily fines. This follows a pattern of DMA enforcement, with Apple and Meta previously fined over $700 million combined in early 2025, indicating the EU is actively applying gatekeeper obligations to major US tech platforms. 25 Republican lawmakers have asked Trump to launch trade investigations in response, potentially leading to tariffs or restricted EU access to US technology, framing the DMA as targeting American firms while Chinese platforms like Temu and AliExpress face fewer restrictions. This case highlights the growing friction between EU digital regulation and US trade policy. I mean, in general, the, uh, you know, the numbers are big, but these companies make billions of dollars in profits. And like, you look at US fines, they're like, oh, we're going to fine you $25,000 or, you know, maybe a million dollars. Like, that's the most I've ever seen a company being fined in the US. So the thing that makes these interesting is they're just, they're big price tags for violating EU policy. But, you know, if you want to do business in EU, I feel like it's sort of the cost of doing the business that you had to meet their regulations. And so, you know, trying to now impose trade restrictions because of their being unhappy about companies violating their laws and regulations seems very backwards. That's all I'm going to say.

[16:25] Justin Brodley: I'm going to throw in on the payments thing. I like, in a way, the closed ecosystem, and I like that anyone can list an app and it gets the same security scan as every other app on the App Store. And there's a cost to maintaining that, and there's a cost for the CDN, there's a cost for all that stuff. There's a cost for developing Android. Do I think they should get such a huge cut of any transaction in the app? No, I don't think so. I don't think that's fair. But then I also think the same thing about Visa and Mastercard, you know, having effectively a 3 to 5% tax on every single credit transaction that happens in the world every day. I don't see people going after that, and that's surely worth a whole lot more than, uh, Cause the EU is going after that.

[17:10] Matt: They, they're trying to create competitive payment systems that are, do not rely on American Express and MasterCard and Visa for that exact reason. So that's a big thing they're actually working on from sovereignty perspective is they want to disconnect from that ecosystem as well. So, you know, I get you what you're saying. It just, the reality is the EU is pushing for a lot of these things.

[17:31] Justin: And I feel like people are fighting it. You know, most places now are passing that along to the consumer, like most small businesses. Like if I go to like the local restaurants, it's $15 for whatever and it's $15 plus 3% if I pay with my credit card. I mean, personally, I've started carrying more cash because of that, because it's not worth the upcharge on everything.

[17:54] Justin Brodley: Yeah, same here.

[17:56] Matt: I mean, I like it the other way. Like, give me a discount if I pay in cash. That way I don't feel gypped as much. But yeah, yeah, I've seen it both ways.

[18:04] Justin: Yeah, but they do it that way because then they were able to keep the prices lower.

[18:08] Matt: Yeah. I mean, we had to get the brakes replaced on one of our vehicles and they were like, well, you can pay cash, uh, and get a discount. We're like, yeah, no one has this kind of money just sitting around in cash. I'm gonna whip out here at the, to pay you. I'm just gonna put on card. I'm like, well, we have to, we offer it though. I'm like, okay, cool.

[18:26] Justin Brodley: Thanks.

[18:27] Justin: I think they have to offer it because otherwise they can't do the charge, the extra charge. It's like in violation of their agreement with the credit card companies. Hmm.

[18:36] Matt: That's interesting.

[18:37] Justin: Like, I think there's, there's like a nuanced rule there, so they have to give you an option. They can't just charge you more for that.

[18:43] Matt: Is that a contractor? Like if you're using a labor, a labor type thing, is that where that regulation comes from? No, I don't know about that one. I haven't seen that. Because I know like on the online, you don't have that option. You don't get a discount by using an EFT payment versus using a credit card.

[18:58] Justin: Oh yeah. I don't know. I, I thought it was, there was something weird there. I'll look it up now. I have some nighttime research.

[19:04] Justin Brodley: I, I thought there was, there was, there was law at one point that prevented people from actually charging a fee for credit card use, but presumably that's gone away. 'Cause I see that thing all the time now.

[19:15] Matt: It's happening all the time now.

[19:16] Justin: Well, but that, that's what I think it is too. Like, that's why I thought that, because there, it was, you weren't allowed to charge more for credit cards, but then they do now. But I think they're only allowed to if they offer a discount somewhere else or something like that.

[19:28] Justin Brodley: I thought.

[19:30] Justin: That's the research I now will go have some fun doing.

[19:33] Justin Brodley: Well, when none of us have got jobs because of AI, it won't matter, will it?

[19:37] Matt: But the— and the billionaires be mad they're not making their billionaire monies because we can't pay for the services. Yeah, yeah. ASML shares dropped following a report that China has produced its own deep ultraviolet lithography tool, raising questions about the effectiveness of export restrictions on advanced chipmaking equipment. Now I know we talked about this Maybe not as official show notes, but we were looking for earlier. But, um, if you don't know who ASML is, ASML is the company that makes basically all the technology that TSMC and Intel use to make chips these days because their technology is what makes you get to 6 nanometer and 4 nanometer using light. And so it's a light etching technology lithography. It's very fascinating. There's a video I linked to in here called The World's Most Important Machine. Uh, it's about 50 minutes. Uh, I watched it the first time and it blew my mind how they make chips these days because I, you know, I was more familiar with the older 386, 486 tape-out mechanisms and, you know, then that went on for a very long time. So that was still relevant even through like recent Xeons, not now, but, you know, up until maybe 10 years ago. Um, and so yeah, most new chips now are using, uh, this DUV technology. And so of course America restricted access to export controls for those. And so China said, fine, we've already stolen the IP anyways, so we're going to produce our own machines., and so the DUV tools are less advanced than the extreme ultraviolet systems that ASML exclusively provides, but domestic DUV production will still reduce China's reliance on ASML for a significant portion of chip manufacturing needs. And once you have it, you can start innovating. And so I assume that, you know, in some period of time they will catch up to the extreme ultraviolet system that ASML uses. Investors should watch whether the signal of broader, uh, signals a broader erosion of ASML's market position in China, which has historically been a substantial revenue source for the company despite export restrictions.

[21:17] Justin Brodley: Yeah, I mean, I almost think you should wait for, uh, for the Chinese companies to almost have their products ready, and then we say, okay, now you can buy the stuff from ASML, and we'll give it to you for a decent price. Keep the market share. Like, at this point, it's interesting because ASML is actually a Dutch company. Yes.

[21:35] Matt: So why— I thought most of the lithography extra, you know, export controls were US export controls. I don't think they were European Union one. So I, I now have research to do on that side, Matt, of European restrictions on export controls of DUV technology, because, uh, definitely very interesting.

[21:53] Justin Brodley: I, I wish they weren't doing it though. I, I wish, um, I wish there weren't export controls. There's clearly not enough capacity to make silicon at the moment.

[22:02] Matt: Well, this is just like, you know, you can't export the current technology. They still can run the old technology. And like memory, which is my biggest complaint right now, doesn't use this ultraviolet technology. It just need— the fact of the matter is they just didn't build enough chip fabs because there was a major depression in RAM prices for 10 years and they didn't invest in capacity. And then capacity now is outstripped demand, or demand is outstripped capacity.

[22:25] Justin: So to this level, I feel like there also hasn't been massive improvements in RAM besides size in years, which is also interesting. Like, I feel like the—

[22:35] Matt: there hasn't really been a critical need I mean, DDR5 was big and I think DDR6 was on the way, but I mean, even then you were talking about just increasing the speed of the chips, not necessarily increasing the capacity, but then like in servers, like 1.5 terabytes to 2 terabytes is kind of the max. And that was really more about a, a control plane limitations on how much data throughput you can run through the buses than it is about the memory itself.

[23:02] Justin Brodley: Yeah, I like that. Maybe we need to get AI on the case for building a better RAM chip at the moment, because the fact that you have to read the contents out every, like, 20 nanoseconds and rewrite it again so that it actually persists is not a great strategy. The more RAM you've got, the more work you have to do to keep it refreshed all the time.

[23:23] Matt: That's true. Yep. And then it's a question of how many cores can address that RAM, and like, there are all kinds of nuances to it.

[23:30] Justin Brodley: Yeah.

[23:31] Matt: All right, well, AI is how LLM makes money, and they're making lots of money these days. OpenAI first has OpenAI Presence, a new enterprise product for deploying production AI agents that combine model reasoning with policy, the guardrails, and escalation rules, moving beyond raw model access to managed agent deployment systems. Each deployment is scoped to a specific job, such as billing resolution or IT support, with agents given only the knowledge and system access needed for that task. Plus company-defined policies on permitted actions and human handoff triggers. OpenAI's phone support line, which I've never called, runs on presence and resolves 75% of inbound issues without human assistance, with a Codex-powered improvement loop reducing human handoffs by 15 percentage points in 10 days. Early enterprise adopters include BBVA for banking voice support in Mexico, SoftBank for Japanese language customer conversions, and IAG for high-demand event support, indicating cross-industry interest and production-grade agent deployments. Presence is currently limited to general availability program led by OpenAI forward deployed engineers and select system integrators.

[24:29] Justin: I love forward deployed engineers. I'd love the new terminology for that.

[24:33] Matt: I mean, put contractors on site, charge lots of money for them and have them use your technology to build stuff for them. Yeah.

[24:39] Justin: It's, I mean, it's just an engineer. It's the same thing as, I mean, I don't understand really the difference between a forward deployed engineer and like a field CTO and all these other things like.

[24:50] Matt: I mean, I feel CTOs are more of a sales role traditionally. They're going—

[24:53] Justin: but that's kind of what the forward deployed engineer is, is a salesy engineer on site, as least was my understanding.

[24:59] Matt: Sort of. I mean, like, you typically have a team, you know, these are the ones I've seen so far. They have, you know, a team, 10, 20 people. They've embedded into the company. They're badged as company employees, but their job is to use AI technology to accelerate and redo your business process. But because they're inside of your company, they have direct access to the business processes And so that's supposedly the combined, the fact they have all this AI knowledge and all the skill combined with the fact they have access internally to your team because they're forward deployed, they can move faster is what the illusion is. Now I haven't talked to anybody who's been tremendously successful with FTEs, but I know they're everywhere right now. Yeah. Even my old job is moving to forward deployed engineers.

[25:40] Justin Brodley: So.

[25:41] Matt: All right. Anthropic has released Claude Opus 5, priced at $5 per million input tokens and $25 per million output tokens. Tokens. For those of you keeping track at home, that is the same price as Opus 4.8. It's now the default model on Claude Max and the top model on Claude Pro. Opus 5 achieves state-of-the-art results on coding and knowledge work benchmarks like Frontier Bench and GPT-Val-AA and reportedly performs within 0.5% of the larger Fable 5 model on Cursor Bench at half the cost per task. I've been playing with it a lot. I don't really notice any major improvements. And one of the fun things is it'll drop to Opus 4.8 on the backend without you even knowing it. It just happens. And so you don't really actually know if you're using Opus 5 or Opus Orbita 8. But I mean, I think this just goes to my point, which I've been making for a while, is that these models really are not evolving very quickly now. They had a while there where they were leapfrogging from, you know, GPT-3 to GPT-4 to GPT-5. Like those are big jumps. I think we've now reached the point where the improvements are more iterative. They're more in the harnesses around the foundational model, like how they do ingestion of data, how they parse the data into the model, how they do lookups of the data. But you know, OVIS-5 is like the one I was like, wow, this is really just not that impressive of an upgrade.

[26:54] Justin: Yeah, I've seen no real major changes on it. Like I just moved everything over to it just outta curiosity and it feels about the same. I haven't seen much. I feel like it's a little bit different in how it communicates. Like its style is different. I haven't quite placed my finger on it if it's like longer or shorter, like different verbiage in it. Like But overall, it feels about the same quality-wise.

[27:18] Justin Brodley: You should open up a previous chat that you had, maybe when we did some coding, switch it to Opus 5 or to Fable 5, and have it review the conversation and give you its new take on everything that was delivered. And that is fascinating. I highly recommend it. It has a whole different perspective on the types of mistakes that it made, the assumptions that it made. I think you're right that a lot of the advances we're seeing are to do with the harnesses. But I'm less and less, I'm having to micromanage any kind of coding activity with Claude Code. And I don't know if it's the harness that's changed or I'm not sure the harness has changed, but, or whether it's the sort of fundamental capabilities of the model to think more through things and rightfully masked from view a lot of the thinking. It's just, you just get like a signed, cryptographically signed block if you look in the JSON files. So the thinking is still preserved on the Anthropic side, but you just don't get to see it. I assume that's just to stop people from training their models on their Anthropic's work.

[28:37] Matt: Nice.

[28:39] Justin Brodley: Yeah.

[28:39] Matt: Well, uh, I have still not played with KIMI. It finally shipped in Ollama this morning. Uh, so I will be playing with that and I can maybe update next week on how I think about the new KIMI 3. As I talked about before, I'm a big fan of GLM 5.2 and KIMI 2.7. So I'm hoping KIMI 3 is close to Opus. So then maybe I'll just start using more Ollama cuz I like it quite a bit. And I, uh, I still using Claude Code as my client, but I have been trying to use OpenCode a bit more just for the, comparison of it. And, uh, still don't like OpenCode, but definitely like KIMI. So, so based on, you know, some of the news coming out this week with Hugging Face being hacked and OpenAI and OpenModels, uh, and their CEO of Anthropic, Dario Amodei, basically felt he needed to clarify their position on open weight models. And so basically he clarified the company's stance amid reports of potential US bans on Chinese open weight models, stating Anthropic has never advocated for such a ban despite accusations otherwise from signatories of an Industry Open Letter. Omidai outlined two national security concerns: authoritarian governments building militarily superior AI models and misuse of open weight models for cyber or biological attacks due to the difficulty of applying guardrails once weights are released. Instead of blanket bans, Anthropic supports three specific measures: restricting chip and chipmaking equipment sales to China with stronger enforcement against smuggling, cracking down on industrial-scale distillation operations that let China approximate US model capabilities with fewer chips, and mandatory safety testing for all sufficiently There's sufficiently capable models regardless of open or closed status of country of origin. The post pushes back on NVIDIA-backed open letter claim that open access inherently helps defenders more than attackers, citing biological weapons as an area where AIMODI believes attackers may have a structural advantage over defenders. Anthropic references its own research on modular pre-training strategies as a potential method for improving safety of LLMs, suggesting technical mitigations could complement policy measures rather than requiring outright restrictions. Yeah, that's all BS. Thanks, Dario. Uh, I appreciate that he's trying to, you know, you're basically limiting chip manufacturing because that slows them down so they're not copying Mythos and other things you're doing as quickly, and you're trying to basically make it so that it sounds like they're untrustworthy. And those models, you know, at least what we've seen, they don't seem to be any more untrustworthy than their, than the Anthropic models. You know, Anthropic did just as many bad things to build their models as as China has allegedly done, using distillation and other things against Anthropic. So I don't know, I feel this is, I'm glad that you've clarified your position, but I still feel your position is BS.

[31:07] Justin Brodley: It must be frustrating though, having your own models limited in a way that you can't show off their true benefits. I mean, I'd love to have Fable do a full comprehensive security scan of stuff that I'm building, and I can't, because as soon as I mention security, it falls back to Opus, and Opus built it in the first place. Want the second opinion.

[31:27] Matt: Well, that's more, that's more of a problem of our friendly government administration that made Mythos dangerous and no longer allowed to be accessed by anybody.

[31:36] Justin Brodley: Yeah, but, but now we've got K3, which is probably very, very close to Fable level or, or Mythos level cybersecurity work.

[31:44] Matt: I think they were claiming it's close to Opus 4.8. Okay. And capabilities. So they weren't quite saying it was Mythos level, but they, they're definitely pushing that direction, trying to make it as good as Mythos and other Fable models.

[31:56] Justin: So.

[31:56] Matt: But the thing they did add in Kyma v3, which is the big game changer for them, is they increased the context model size to 1 million tokens. Although I know last week Matt was a little concerned about that because he's like, I think you should live in 256 or 200,000, whatever the number is. They did release also a Kyma v3 256 context model format. So you don't have to go to 1 million context if you don't want to.

[32:17] Justin: Thank you.

[32:19] Matt: Yeah, they heard your complaints and they addressed it.

[32:22] Justin: Someone's listening to me. I'm not just yelling at inanimate objects all day.

[32:29] Matt: Uh, I mean, I, it's really hard to like any of the, uh, AI model CEOs, uh, and, and the other ilk, you know, like OpenAI is problematic. Dario and Anthropic have problematic moments that just, it's a young industry and they make young dumb mistakes sometimes, I feel like.

[32:48] Justin: And so, yeah, I mean, I think that you were talking about it before and we really have to sit down and have a longer conversation, all of us, about how we use AI and our OIDC and everything along those lines that we keep talking about we should do. So at one point we should probably actually do that. But you were talking, I think Justin and I were talking on Slack about, you know, if I'm using Claude and Anthropic models and everything to do the development, even if it's a different model, is that a good PR reviewer? Is that a good security reviewer? Or should you really be using a different provider for that? You know, should I be using Akemi or whatever, or whatever else to kind of get that slightly different perspective on it?

[33:34] Matt: I mean, I do agree that occasionally you need to challenge. I mean, the same thing with the security tool, like Veracode versus Checkmarx. I think you should keep Checkmarks or Varicode for a couple of years, and then I think you should switch them because they're, they're tuned and they're better to things. I think the same thing happens with AI PR stuff. I mean, I can tell you that I use CodeRabbit, uh, in one project, and then I also have an internal PR agent that I, you know, that I open source one that we're using on the Cloud Pod stuff. And even though it uses Claude in the backend, it finds different things than the Claude stuff found that it just coded. So fresh context is actually just as valuable as a different model. Because it forces the model to look at the data again in a different way, or it activates potentially different paths. So because it's so undeterministic in its nature, I don't know that there's a lot of value in the different models. I think there's probably at some point you should do it. Like, and I've used Codex a few times where I'm like, hey, I haven't used, you know, I haven't had a security review done on this codebase in a while other than I do it every time I do PR work with the security agent I have, Claude. And I'll make Codex do a pass. It finds a couple things. It feels very much like an auditor, like, well, to be successful, I have to find 5 findings, which is very— what I always joke about with auditors, that they don't get paid unless they find 5 things wrong with your audit. And so I do think there is some value, but I think the value is very minimal, is my take. I don't know, Jonathan, you're much more into this than I am, and maybe you have a different opinion, but—

[35:03] Justin Brodley: No, I'm pretty much the same opinion. I think that Having Claude write code and then having Claude in a different context approach the code, not having seen it before with a set of criteria or tests or whatever else, or look at it with a fresh set of eyes, is just as effective, if not more effective, than switching to a different model. You also got to remember that all these models are trained on the massive internet archive of documents plus scan books and everything else. At this point, there isn't much more data other than, you know, private data that businesses have, which I'm sure is where the FDA is going to get involved. But so like the data that models are trained on, it's going to be fairly consistent across all the big models at this point. And so all the innovation is going to come from the way they're trained to think and the context that you give them. So yeah, I don't think it's worth switching to a different thing. I mean, maybe use a cheaper cheaper model to see if it can find a bug or something. But I don't know.

[36:07] Matt: I mean, I think it's also like when I'm dealing with a lot of TypeScript and Node.js, like pretty much any model can do that. Like there's so much documentation in the world in that. You wanna go write assembly code, you probably need to find a model that's really good at assembly code. I don't know how to write assembly code other than the few times I tried in college and I hated it. So I mean, like, I, or, you know, COBOL, you know, those things, like, there's certain models that are gonna be better at those things. I think that's where the specialization of a model, and that's based on its training data, but that's really gonna be experimentation. You have to figure that out. But I think for most of the common languages that we use day to day, TypeScript, Python, Golang, .NET, Java, I feel most of the models have the same level of knowledge in those technologies.

[36:51] Justin Brodley: They do.

[36:51] Justin: I think it goes back to our conversation is it's not about the data in the model, it's the harness.

[36:57] Matt: Correct.

[36:57] Justin: You know, Jonathan said before, looking at Opus, having it run on, you know, sorry, Opus 5, look at your old conversation and, you know, re-review it and see what I thought about it. It's just that other perspective. It's both context and the other model, you know. So do I think it's going to be a game changer? No. Do I think it has a potential to find something? Yeah, and if you're already running like a tool, like, you know, you set up on The Cloud Pod, you don't want having it just run on different API since you're already spending it. I think that little bit of change is worth it almost.

[37:34] Justin Brodley: Yeah, something I've started doing. So I went down that path of using Gemini and the other models to do the, like, the adversarial review of the code that I built with Claude, and I just I stopped doing that and just went back with Claude again. But I took a different tactic. I'm not just asking it to find bugs and fix bugs. I'm asking it like the higher-level question, of the bugs you fixed in the past session, these 5 things you found, what could I have done differently in the project? How could I have prevented these things from happening in the first place? And I think with Opus 5 and Fable, I've got some really good advice on how to think about coding with AI in the first place so that these issues don't come up again next time.

[38:20] Matt: That's good insight. I hadn't thought about trying that. Because like, I know there's things I put into like memory or, you know, lessons learned MD files that I've created over time. And like, it sometimes works for like some things like Terraform. For some reason, it wants to put in dashes or commas in Terraform descriptions, which is against the HCL.

[38:40] Justin Brodley: Yep.

[38:41] Matt: Yep. And like, I have tried like 4 or 5 different ways of making it stop doing that without it, but it never— at least once a time I do a Terraform, it's like, oh, I tried to apply and I couldn't do it. It says invalid character. I'm like, son of a— so I had— I actually, I actually ended up writing a test for it. So I rely on, rely on good old regex to find it and fail a test because that's the only way I can make sure it doesn't happen. Uh, but like some of that stuff, it's hard to get the model to stop doing the dumb.

[39:08] Justin Brodley: Yeah, definitely. 'Cause it looks so much like JSON, but it's not, I think is where the problem is. But I use pre-commit hooks in Git.

[39:16] Matt: Yeah, it's a pre-commit hook and a test, and basically it'll fail if it doesn't now. But, you know, I tried so many other ways. I'm like, can I teach this model not to do that? Like, there's gotta be a way. Nope, memory didn't work. ClaudeMD didn't work. Just, you know, it would just eventually screw it up, so. All right, well, let's move on to AWS. Amazon AI keeps giving with features that we've been asking for forever, and I'm still going to credit them to AI because, you know, it's just quality of life improvements that I can't see anyone doing other than AI. And this one is Amazon Network Load Balancer now supports listener rules for custom traffic routing. You can now let a single dual-stack Load Balancer route IPv4 and IPv6 client traffic to separate same-family target groups, preserving the original client IP end-to-end without protocol translation. This addresses a longstanding architectural trade-off. Previously, teams either ran two separate NLBs and split clients via DNS, or funneled everyone into one target group and lost client IP visibility through NAT64/protocol translation. Rules can be added to existing dual-stack NLBs without recreating them, and they support TCP, UDP, TCP/UDP, and TLS listeners, working alongside existing features like connection draining, stickiness, cross-zone load balancing, and weighted target groups. This is useful for organizations consolidating infrastructure while maintaining IPv6 compliance mandates. Without doubling Load Balancer count or losing client IP for logging, security, and geolocation purposes.

[40:38] Justin Brodley: Nothing to say, Matt?

[40:41] Justin: No, it just feels like something I've been asking for for a long time, and it's kind of nice that it finally exists.

[40:48] Matt: Oh yeah, that's wrong. These quality of life, they're not sexy press releases, but they're like, oh God, I don't have to worry about that anymore.

[40:54] Justin Brodley: Yeah, if you needed it, great. It's just, it's just saved you $50 a month per per deployment because now you don't need two NLBs.

[41:01] Matt: Yeah, it's a cost savings.

[41:02] Justin: NLBs, the IP addresses, you know, all the little things that you had set up.

[41:07] Matt: Yep. Now you just get rid of those IPv4s that you're paying for. We'll be really set.

[41:11] Justin: So. You gotta go full IPv6.

[41:14] Matt: And another quality of life improvement, ALB logs can now flow directly into CloudWatch logs as vended logs covering access connections and health check data for troubleshooting traffic and target health issues without pulling logs from S3 first, which was the dumbest thing ever. CloudWatch telemetry enablement rules let teams auto-configure logging across an org, specific accounts, or resources, covering both existing and new ELBs, which removes manual stub for consistent monitoring at scale. Integration supports CloudWatch Log Insight queries, metric filters for alarming, and live tail for real-time traffic review, giving teams more ways to analyze logs without standing up separate tooling. And all I can say is thank you.

[41:51] Justin: Is that a better response, Jonathan, than last time?

[41:54] Justin Brodley: That's, that's pretty good. Yeah. I mean, I assume they just built some glue on, on the backend that just pulls it back out of S3 and sends it to—

[42:03] Matt: Well, I mean, all they had to do, but this is one of those things like forever, like every other service on Amazon will send data to CloudWatch Logs natively, but ALBs for whatever reason would not do that. You had to send them to an S3 intermediary and then load them from S3 into the CloudWatch. And so yeah, I'm sure it's glue on the backend, but the fact that they don't have to do it now, so much nicer. And another thing is you'd always end up in a situation where like, oh yeah, we have ALB logging turned onto an S3 bucket that no one's looking at and it's hundreds of terabytes in size because no one did anything with them. And 'cause unless you're troubleshooting something, you don't really need them or you have a very persnickety security team since Ryan's not here, who really wants them and doesn't use them either.

[42:41] Justin: And it's not like this is like the most chatty service that like it's gonna be the thing that cost a ton. You had VPC flow log, which is the most chatty thing in the world that you could turn on. And that's always been to CloudWatch. That was the first thing you did. So like this not being there just drove me crazy. But now I actually wonder if NLBs are there or if they're going to add that now. We'll have another press release for that. I don't— 2 weeks.

[43:06] Matt: I didn't think NLBs had ALB-style access logs to begin with. I thought they only supported flow logs.

[43:12] Justin: Yeah, it's just flow logs, layer 7 versus layer 4. Yeah, sorry.

[43:15] Matt: Yeah, this is the problem with these quality of life things. If you're listening to us while you're studying for your, your Amazon exams, you probably should forget them because these will be the things that burn you in the test right now.

[43:27] Justin: So I, I will say Amazon is actually, or used to be, I don't know if I, the, everyone I knew on the certification team has rolled over, but the certification team used to be really good at watching these. And pulling them from the exams within like a month. So there's still some human element in it, but they were pretty good about watching for these types of questions.

[43:49] Matt: Well, if you are dealing with network firewall troubleshooting, you can now use the AWS DevOps Agent to automate root cause analysis for network firewall connectivity issues, correlating CloudWatch alarms, flow logs, firewall configurations, and CloudTrail API history to identify what broke and when, cutting investigation time from hours to minutes. This blog article will walk you through the three real failure modes: a domain denialist blocking legitimate traffic, a stateless rule priority inversion, and asymmetric cross-AZ routing that silently drops return traffic without tripping the firewall's own drop counter. Each requires a different investigation path, which the agent handles automatically. The agent connects via webhook triggered by CloudWatch alarms through SNS and Lambda, and then reads firewall state and logs directly from AWS APIs, so no additional instrumentation is needed on the firewall side. It always presents a mitigation plan for human review rather than applying fixes automatically today. I'm sure sometime in the future it'll be automatic, you'll all hate it. SAML CDK app deploys the full test environment, VPC, network firewall, test workloads, and status page into customer's own account for hands-on practice. Though the two firewall endpoints, NAT gateways and Load Balancers, bill hourly whether idle or not, so cleanup after testing is important.

[44:54] Justin Brodley: It sounds like they just need to make the network firewall easier to use. It's like everything they mention is like You might, you might screw something up accidentally. This is how you fix your mistakes. It might make it easier to use in the first place.

[45:07] Matt: That'd be nice, wouldn't it?

[45:08] Justin: I feel like I've never seen a firewall that's easy to use. Even though I was yelling at my UniFi downstairs last week, like, they're just by default not, I feel like, extremely happy.

[45:19] Justin Brodley: How big of an issue is that now? I mean, how many people are caring a great deal about the firewall level stuff?

[45:25] Justin: I mean, enterprises, presumably, but I mean, enterprises still want traffic to go through a firewall so that if you're, if your serverless EC2 instance gets breached, you still have, you know, an egress traffic monitoring setup. So they still want that egress traffic filtering, but yeah.

[45:50] Matt: Uh, Amazon is laying off some staff and apparently killing Nova. Want, want. Amazon has confirmed the layoffs within its AGI unit, which covers foundational model development, silicon design, and quantum computing work. The company has not disclosed headcount numbers of which specific teams were affected. The cuts follow more than 30,000 layoffs since October, occurring alongside a $200 billion CapEx plan for 2026 and recent debt raises to fund AI infrastructure buildup. Leadership turnover has been notable, with Peter DeSantis taking over the AGI group in December, replacing Rohit Prasad and David Lunn, head of Amazon's AGI lab. Who departed in February. DeSantis has acknowledged that Amazon's models have not reached frontier-level performance for the largest workloads, putting pressure on the team to improve competitiveness against OpenAI, Anthropic, and Google. This does kill everything but, uh, the remaining Nova lineup includes Nova 2 Sonic, Nova 2 Lite, Nova Forge, and Nova Act, and a new flagship model from FMR is going to launch at re:Invent this fall, potentially under the Nova brand. Uh, the only one I care about is the embeddings model because we use it for the vector search. For the bot. And so I would keep, look at that, see if that's going to get deprecated as well. But I would like to see a new Nova embedded model because that one's getting a little dated. So, you know, if anyone's listening, please, please get us a new embedding model.

[47:04] Justin Brodley: Gemma's embedding model is very good.

[47:06] Matt: Is it? Nice.

[47:06] Justin Brodley: I've been using that. Yeah, I switched to using that. It's 600 million parameters or something, 768 dimensions, but it can also shrink down to different dimensionalities if you want. It's very good. Better than NoMIC, which is the one that everyone sort of has used for years, but it's nice that you can run it locally as well, or in the cloud. And I think in a way, I think that Amazon did themselves a disservice by not releasing an open weight model, because if Amazon released a fairly decent competitor to something like Llama, I'm much more likely to have used that locally and then used it in the cloud for deployments than than anything else. I want parity between my on-prem and my cloud deployments. And by not releasing an open-weight model, it's just put me off even looking at it in the first place. It's the whole tie-in to the ecosystem thing. And with AI, it's even more important because of the way, just because of how important the way you prompt different models is to their outputs. Yeah.

[48:07] Matt: I definitely, I mean, we think we talked about this. You weren't here when we talked about it, but yeah, the fact that there isn't really an American open model is really kind of a bummer because I mean, Gemma was your, there's all the Gemma models and then Meta was kind of your one play that was really open first. And then you have, you know, the Maya models from, from Microsoft, but those are both, both the Google and the Microsoft one are just kind of there. But with Google, you know, Facebook now abandoning their model, It sort of feels like there's really no one who's sitting inside the open model, open weights space that's a US-based company, which is really a bummer because I think it is ceding the market to the Chinese, which now they're upset about. So it'd be great if we had a really reliable open model that was US company open sourced.

[48:52] Justin Brodley: Yeah. I mean, somebody else has Google at least. I mean, Gemma 4 is very good, but there's not much to pick from. You know, Llama hasn't been updated. I don't think Llama 3 was a thinking model. It did do tools. Gemma 4 is a thinking model, does pretty well with tools. It actually, it's not a bad model at all, but that's just one to pick from, right? Out of all the others. And it's Google. If you disagree with Google's philosophy on things, then maybe you don't want to use Gemma, but you're right. I wish there was, maybe Amazon could do the right thing and release NOVA as an open language model.

[49:29] Matt: I mean, that'd be great. I would, you know, maybe people would be able to contribute to it and you know, that's been Amazon's thing for a long time is trying to be open and be, you know, part of the community and this would be a way they could do it.

[49:40] Justin Brodley: But yeah, I mean, I'm sure they see it as there's no money in it for us, but I kind of disagree. I think it does kind of act as a funnel.

[49:47] Matt: I mean, it has money in the fact that, you know, you could deploy that open model anywhere on any cloud. So it could give them a multi-cloud models play, which they only have Anthropic really as their kind of partner on that side, but also give them an option to run it on-prem without having to license the model, and that would drive Outpost sales. So like, there's, there's definitely some potential options there.

[50:07] Justin Brodley: Yeah, even, even just familiarity with the brand, you know, would be, uh, would be better than throwing it away, giving people nothing.

[50:14] Matt: Yeah. Uh, well, I mean, also I think Nova hasn't been updating often enough either, so that's a challenge. Alright, let's move on to Google. Uh, we have one story from Google this week with Google's Gemini API managed agents now defaulting to Gemini 3.6 Flash for reasoning, coding, and tool use with no code changes required. Developers can also pin to Gemini 3.5 Flash or 3.5 Flash Lite for lower cost via the agent config model parameter. Environment hooks let developers run custom scripts before or after tool calls inside the agent sandbox, enabling validation, linting, or security gating. Offdeal, an AI-native investment bank, uses post-tool execution hooks to run pixel-level logo verification pipelines inside the remote sandbox for its stack generation workflows, for example. Managed agents are now available on free tier projects, allowing developers to experiment with agentic workflows using an API key without active billing. And new budget controls let developers cap token consumption with max total tokens in agent config. When the limit is reached, the interaction pauses with status incomplete and can be resumed later using previous interaction ID, preserving environmental state. Scheduled triggers bind an agent, environment, prompt, and cron schedule into a persistent resource for recurring automated tasks.

[51:17] Justin Brodley: That's pretty cool.

[51:18] Justin: It's interesting. They just updated it because, you know, when you saw like 3.5 to 4 got updated, everyone lost their mind a little bit, you know, on GPT. And if they're going to go down this path, I'm kind of curious how they continue to do it. If people have, you know, stepped their prompts or people are going to have to pin everything to the specific version that they want.

[51:44] Justin Brodley: I mean, Google are deprecating Gemini 2.5 Pro and Flash in October. So I mean, if you built something against one of those older models, which are really only a year old, you're going to have to move off it anyway. So I mean, I guess just rip the bandaid off and just go with rolling updates. And hope that you can fix your prompts and your code before the model you're using is deprecated.

[52:15] Justin: I mean, I assume they also want the hardware back. So I wonder if it's, you know, Amazon tried to just always make them be cheaper and better, you know, when they went from like, you know, T2 to T3 to T4, et cetera. So I wonder if this is going to be kind of the new way where they just forcibly upgrade you.

[52:33] Matt: No, they don't. That's the problem. And they'll just stop responding to the old API. So they don't force you to do anything. They just kill it and then you're off.

[52:42] Justin: This is forcibly upgrading you.

[52:44] Matt: This is changing it.

[52:46] Justin: Yeah. Yeah. So I wonder if this is like the model that people are going to try to do. Sorry, that was unclear.

[52:50] Matt: And maybe, I don't know, I, their whole communication on their model availability, what regions are available, what, you know, what's the timeline, what's the deprecation period, like they're very cagey about it, which is a big problem for product companies who have to build AI and test and validate and certify all the things on top of AI, not having that data is really a problem. That's funny.

[53:11] Justin: Azure just sent you emails, lots and lots of emails.

[53:15] Justin Brodley: We had a Google operational health review check-in briefly today. You know, they go through the, this is your number of cores that you're using, this is the different instance types in different regions and everything else, and some metrics. I'm like, Where's my token count over time per model? You can tell me what compute instances I'm using per region and how much I'm training up or down. And it's like, seriously, where's the important information, which is which models are my people using? How much is it costing me? And how are they trending? And they haven't put that into their dashboard yet or the deck.

[53:48] Justin: Yeah, it's just an old PowerPoint that they just keep updating.

[53:53] Matt: Yeah, I mean, the overall token Usage tracking is very, very difficult. Like you can, like Bedrock at least gives you some, some lots of capabilities. Like it does track tokens in, tokens out, shows you who the requester is. I did see you can also do text, you can capture all the text you're putting into Bedrock, which they then put it to CloudWatch. So don't put your PII into, into your tokens on Bedrock if you don't want it to go to CloudWatch unencrypted. That's the problem. So there's definitely some interesting things there, but, you know, even what they have is not what I'm looking for. So it's, it's like I need so much more than just this. This is, this is the bare minimum of what you can provide to me. And then in case of Google, they don't even provide some of that data at all, trended. It's just static number. And then Claude is even worse. Claude's like, oh, just enable OTel and you can see all this data. I'm like, the OTel data is not what I want either. So it's just a, it's a, it's a rough, rough area in telemetry on AI stuff, especially as people get into what we'll talk about in the after show, ROI on AI is a big problem. So, yeah.

[54:55] Justin Brodley: Oh, while we're talking about tokens, I mentioned something really quick. I was being talking about Gemma and Gemini and all those things. I, I tested something over the past couple of days. I've been doing a lot of AI work locally, a lot of thinking models. And I, I noticed that, that Gemma would occasionally start writing thinking in, um, in Chinese characters. In kanji. I was like, ah, that's interesting. I mean, like, I didn't, I don't know what it said, but the answers were coming out just fine. So I benchmarked, thinking, benchmarks, Gemma working through the same problems with two very slightly different prompts. One instructed it to use, to think in any language that it wanted to, or specifically the language that uses the most information per token, which is pretty much the kanji character set. And so I've saved about 30 to 35% token count for my workloads by just having it think in Chinese.

[55:52] Matt: Interesting.

[55:54] Justin Brodley: Still outputs in English, thinks in Chinese.

[55:57] Matt: I mean, that's helpful for your token usage, but not helpful if you want to know like, well, why did you give me this answer? It'd be like, well, here's the Chinese. And you're like, well, I don't, I can't help you.

[56:05] Justin Brodley: I can always, you could always translate it. So if you wanted to, but I just, I just said it's in Very interesting that I think some of the constraints are most definitely around the language that people use the AIs in.

[56:15] Matt: All right, let's move on to Azure. First up, the Standard Service Endpoint is now in public preview, offering a more scalable way to connect infrastructure-as-a-service workloads to Azure platform-as-a-service services under the Private Link family, addressing scale and management limitations of traditional service endpoints. The feature integrates with network security perimeter and uses public IPs as network identifiers, letting customers associate a single public IP or prefix with multiple subnets or virtual networks across subscriptions within the same tenant and region. Supported, uh, PaaS services include Azure Storage, Azure SQL, Azure Cosmos DB, and Azure Key Vault, allowing your organization to restrict access so only approved networks and workloads can communicate with those resources. Microsoft notes the capability has already been validated internally at scale, supporting network identification for over 42,000 VNets used by first-party service providers as part of its SFI program, indicating some production-level testing prior to the preview period. This is aimed at enterprises with large or complex Azure environments needing simplified configuration and stronger security controls for infrastructure service to platform as a service connectivity. Pricing details are not yet specified as they don't release them for preview.

[57:21] Justin Brodley: It must be so hard for Microsoft to write these press releases and make them sound like something new without giving any clue as to what it actually is or letting on that everyone else has had this for for years.

[57:36] Matt: I mean, if you're in Azure, if you're in Azure, you don't know that everyone else has had it for years. Like, we know because we follow all three of them. We've all worked with multiple clouds. Like, we know it's missing. It's, it's like when you go into Google and they're like, you're not very cloudy if I had to do capacity planning for, you know, for compute. Now, now in 2026, everyone has to require us to do capacity planning. But you know, before AI, like, that was a weird thing you had to do. And you know, but if you only ever knew Google or you only ever knew Azure, that was just standard practice.

[58:04] Justin: So I feel like you get like Azure people that like Microsoft people don't know, like don't look at other clouds. They're like, this is what we do and this is what it is. AWS, I feel like it's sort of like that. And GCP people, at least from the people I've talked to, have a general concept of multiple clouds. But Azure is just like, this is what we do. And that's why I think it's easy for them to write. They just take the Amazon press release from 7 years ago. Change some words in there, change the name of the service to something more confusing, call it a day.

[58:36] Matt: All right. Azure DDoS Protection custom policies is now available to you in public preview, letting customers set fixed inbound detection thresholds for TCP, UDP, and TCP SYN traffic on standard Load Balancer frontend IPs ranging from 50,000 to 2 million packets per second. This gives you manual override control instead of relying solely on Azure adaptive auto-tuning. Which is useful for predictable traffic spikes like product launches, gaming events, or seasonal peaks where autotuning might not react fast enough. The threshold setting is per protocol, so customers can mix and match, set a custom threshold for one protocol while leaving others on adaptive autotuning. I can't wait for the next outage where we forgot that we set it too low and didn't use the autotuning. So use this footgun with care.

[59:18] Justin: I mean, it's better than the alternative, which definitely never happened, where it started blocking something. That it was real and there was no tuning ability to say, no, this is okay. Which is definitely not something that I ever ran into.

[59:31] Justin Brodley: You kind of need the best of both worlds. You need to be able to set your sort of enforcement threshold where you want it, but also like a warning for it. Like, tell me if you think something's going on, but don't do anything about it yet.

[59:43] Justin: I was gonna say, you need like a, what was there? I feel like Cisco had it where it was like, remind automatically reload if I haven't like written this command in the next 2 minutes. It was whatever, like you dealt with like networking or, you know, switches or firewalls, you would do it. So essentially if you, I don't know, done something stupid, like I've definitely never done, which is like if down, you know, is zero and all of a sudden you've lost access to it, it would reload the original config. And that's kind of what EIW here, which is like, allow this traffic for now because we know that that's acceptable, but send me a notification if it's still there in 7 days or delete it in 7 days if it's still there. So some sort of like automatic cleanup, but that's what, you know, Lambda slash App Functions backend can be for.

[60:29] Justin Brodley: Yeah. I mean, at least it's not an AI story. It could have been like new AI features for setting DDoS protection limits.

[60:38] Justin: Right, it could be.

[60:39] Matt: New AI agent does what you want it to do for months. Yeah, rely on the AI agent to be as good as the autotuning. Yeah, it's great. Well, if you love things on the edge, Azure Front Door now supports Edge Actions, allowing customers to run programmable logic directly at Microsoft's edge network rather than routing requests back to origin server for processing. This positions Azure Front Door more directly against competitors like Cloudflare Workers and Fastly Compute Edge, or just, you know, AWS CloudFront Edge. Key use cases include header manipulation, URL rewrites, custom security logic, and personalization tasks that can be handled closer to the end user. Reducing latency and offloading work from backend infrastructure. The feature targets customers running latency-sensitive or globally distributed web apps who need fine-grained control over traffic without maintaining separate compute infrastructure at multiple regions.

[61:25] Justin: I mean, this is another one that burned me when I was trying to design on Azure, and in order to do it, you know, I was doing a simple like return the IP address, and instead of just doing a simple I guess, function at the edge or Lambda at the edge. I had to write a whole thing that passed traffic to a static, like it was a whole thing. Just, it amazes me how long it took them to catch up here. I guess Front Door being down for 6 months might've been a bigger problem that they had to deal with first. Maybe this is like a delayed release. It's just sitting in the queue of things to get released.

[62:01] Matt: Well, if, uh, you have enabled this new edge capability, you might want to monitor it. And apparently you now have in preview Azure Monitor Advanced Platform Metrics and public preview starting July 15th, giving customers deeper telemetry on resource health, performance, and operational trends with Azure Blob Storage as its first supported service. Oh, guess not gonna help you on Front Door, but Blob is there. The goal is faster issue identification and improved troubleshooting by surfacing additional monitoring signals beyond the standard platform metrics currently available. This is a preview program, so Microsoft is explicitly seeking customer feedback before moving to GA. This is relevant for storage, DevOps, and IT ops teams who rely on Azure Monitor for observability, particularly those managing blob storage at scale who need more granular operational insights.

[62:43] Justin: I just really wanted this to be called Ultra Premium Platform Metrics for Azure Monitor. That's the only reason why I wanted this story in here.

[62:50] Matt: Maybe they will when they release the pricing for it, but it's only in preview now, so we had no pricing. But yeah, when it GA's, and we'll talk about it again, I'm sure, you know, maybe they'll have Ultra Premium for you to get even more advanced metrics. That you should get for free.

[63:02] Justin Brodley: I'm kind of disappointed that the press release, you've really got to dig through to find out, well, what are these advanced metrics exactly?

[63:10] Matt: Oh, let's not talk about how bad the press releases are on these things.

[63:16] Justin Brodley: Yeah.

[63:16] Justin: Yeah.

[63:17] Justin Brodley: It's like 3 lines of SEO content and no detail whatsoever. It's like, don't forget about us.

[63:24] Matt: Yep. Yeah. Which shared storage accounts are there? Dynamics. Yeah, it doesn't really— we're looking at a couple of them here. Just trying to see. Container blob capacity, the amount of storage used by a specific container in a storage account. And container blob count, the number of blob objects in a specific container in storage account. That's the advanced public platform.

[63:40] Justin Brodley: Yeah, yeah. How many objects in a container and how much storage is used by it? That's advanced. Okay. Nice one.

[63:47] Matt: Thank you. Thank you for that, Azure. We appreciate it. I can't wait to see what you add between now and the GA for that. That's your definition of advanced. All right, well, general available, so maybe this will be better. Resource placement in Azure Kubernetes Fleet Manager is now generally available, allowing you to distribute Kubernetes resources across multiple EKS and Arc-enabled clusters from a single control point rather than managing each cluster individually. The release includes V1 of the resource placement Kubernetes API plus a new Azure portal experience for creating and managing placements, giving teams both programmatic and GUI options depending on workflow reference. Placement policies use labels and cluster properties to determine target clusters, which reduces manual effort and configuration drift risk when applying updates across fleets. This targets platform and application teams running multi-cluster AKS environments who need consistency without cluster-by-cluster manual updates, a common pain point in larger Kubernetes deployments. If they just extended this to also support Google and also support, uh, EKS, you could have had a competitor to Anthos, but you just, you just missed it. You were so close, but you didn't do it.

[64:50] Justin: I could see them adding that on in the future.

[64:52] Matt: I could too, but.

[64:53] Justin: But ECS also has the ability to do multi-cluster support, I thought.

[64:59] Matt: It does.

[65:00] Justin: So I feel like this is also just them catching up to that.

[65:04] Justin Brodley: Yeah. I mean, for me, if I was an Azure user, I'd be happy with something like this because, you know, I, I can still deploy, I can still have deployments that are separate in multiple regions, but I only actually have one deployment that I have to manage. It just fans out and deploys in those multiple regions, which is kind of nice. Instead of having to do 4 separate deployments, I just do one and then it manages the updates across the other clusters. So it's actually not a bad feature.

[65:33] Matt: Well, good.

[65:34] Justin Brodley: I still won't use it, but it's not a bad feature.

[65:39] Matt: Yeah. And then finally, our last story for Azure is Microsoft introduced Project Perception, an agentic security system entering public preview on August 3rd., which will coordinate your red team, blue team, and green team agents in a closed loop to discover, evaluate, and improve security posture continuously. The system uses a multimodal architecture, applying specialized cyber models alongside frontier models to balance quality, latency, and cost, rather than relying on a single model for all tasks. Microsoft's first specialized model, MEA-Cyber-1-Flash, is now integrated into M-DASH for software vulnerability management, scoring 96% on Cyber Gym benchmark, 12 points above Mythos, while cutting costs by nearly 50% compared to current M-DASH configurations. Project Perception is built on a new CyberStack architecture with layers for signals and sensors, security context models, coordination hardness agents and actuators, and integrates directly with existing Microsoft security products to convert insight into automated actions. And I perceive that this will result in a breakout where it hacks some customer like Hugging Face soon, just because that's the norm.

[66:39] Justin: That means you trust Microsoft to be that cutting edge to actually do that.

[66:45] Matt: Well, I mean, I, they're really agents do coordinated red team, blue team, and green team agents. What could go wrong in a, in a closed loop environment? So it's all the telltale signs of a future breach that I've seen in the AI space. So we'll see. All right. Well, we alluded to it a little bit earlier, but, um, our Cloud Journey today, we're talking about agentic AI ROI, and I've been answering these questions a lot. Lately, uh, in some conversations I've been having. Basically, this is from Snowflake, and so this is Snowflake's version of it, but they're basically backing theirs with their ROI survey data, which cites a 41% failure rate for agentic AI initiatives expected over the next 36 months. Yet 32% of executives report agents already in production, highlighting a gap between ambition and execution readiness. Mm, yes. Article argues traditional ROI models are insufficient for agentic AI, proposing 3 measurement dimensions instead: direct cost savings, revenue acceleration from faster decisions, and risk mitigation from improved accuracy. Framework we're scrutinizing given it comes from a vendor with stake in the outcome. A key infrastructure point: production-scale Anthropic AI requires elastic compute that can scale to handle large datasets on demand and then scale back down, shifting cost structure from CapEx to OpEx, which changes how organizations justify and time their AI investment returns. Governance is framed as a revenue enabler rather than a constraint, with example of policies defined once and enforced automatically across AI workloads, an approach that piece an approach the piece connects to define financial services use cases like KYC compliance and fraud detection. Anyways, this is a, you know, the overall article is a promotional piece for Snowflake, so just be aware of that. But I think in general, the big part of this conversation is really around how do you measure the ROI of AI? And so, you know, I've been looking at it from a bunch of different angles, from AI developer productivity, you know, and there's things like tools like DX and DORA reports and metrics around things like, you know, number of PRs done, number of PR reviews number of comments, uh, you know, the ability to get from feature idea to production. And is that collapsing as a timeline? There's all kinds of different ways to measure ROI of Agentic AI workflows. Um, but I'm kind of curious, what are you and, uh, Jonathan and Matt seeing?

[68:51] Justin Brodley: You wanna go first, Matt?

[68:52] Justin: No, I'm gonna let you take this one. Mainly 'cause there might be a screaming kid in the background, so, you know, try to be on mute for a second.

[69:01] Justin Brodley: I don't know that we are ready yet to even make a realistic assessment on ROI because it's still very new. I mean, a couple of years, okay, we've been around for a couple of years, but it's changing every 3 months. And so people right now are playing constant catch-up. Developers are playing constant catch-up. The tools are changing. Claude Code gets deployment per day. The models change every 2 or 3 months, which need different prompting styles. I think a lot of people are sort of very slowly trending towards getting better generally at using these to achieve goals. But at the same time, you know, a year ago, it was all about, you know, write the PRD, have Claude look at it, and do the work breakdown, everything else. And now, if you look at what people are doing, it's the complete opposite, the polar opposite of that. It's, look, these are some examples of what I want my API to look like. Here's 12 example calls for this API. Build the code that makes this do this thing. It's very much modeling the inputs and outputs and letting AI figure out the box on the inside than the traditional development process. So I think it's going to be expensive for a while longer. And hopefully we'll reach a point where there's some kind of stability in tooling, or we reach a point where there isn't a place to keep optimizing or to keep changing things. And then we might start to see an ROI. But yeah, I don't know. I don't know how you'd sensibly measure something which is literally, which is so new and changing so quickly right now.

[70:38] Matt: Yeah. I mean, I think it's gonna be evolving art. That's why I've kind of answered the question when it's asked, like it's a bit of art, less science at the moment. But I think, you know, there's key metrics and things that are, you know, we use to measure developer productivity that I think are being kind of modified to support the AI use case. Now, The ROI of it is a question of the tokens used. Like, are you using Mythos? Are you using, uh, Haiku? Are you using Kimi? You know, these are all the questions you have to answer in that. And so I think the ROI question not even shifts, like, well, is this feature worth $2,500? Yeah. You know, in tokens plus your developers, you know, plus maintaining it and, and all that. Or is it, you know, is it worth $5? And then what's the value I get out of that capability? And I think if you get proper agentic capabilities that can replace low-level tasks and provide, you know, people to be able to do higher-value things, then the ROI could be measured much higher. But I, you know, it's definitely interesting. I do, I, the one article, part of this article, the fact that executives report that there's, uh, 32% of executives report agents are already in production where, you know, they're basically saying there's a 41% failure rate for agentic AI initiatives expected over the next 36 months. 41% seems really high to me. I don't know what you think about that, but I, I wouldn't say it's not quite that high. I don't think they'll have necessarily the outcomes that people wanted, but they still have value is my guess.

[71:56] Justin: I think, yeah.

[71:58] Justin Brodley: What's, what's a failure exactly?

[72:00] Justin: Well, to me, the failure is they're not getting the ROI they expected out of it. So many people have just shoved AI into every single thing, and there was some example I saw where it was like an AI search box, and you know, it's supposed to just be a one box that's able to search everything in the models, you know. And the problem was it, it just pissed off every customer 'cause the search feature worked worse than before, but they now had an AI search box that was better. So they had AI in the product, you know. So I think some of it's, you know, not just the ROI of like, yes, did you get the feature out? Yes. You know, from like, let's say you tracked it kind of on the DORA metrics with, You know, hey, we're still producing features, we're deploying the same rate, but we're getting more things out, which feels like good, positive, and take away the token part of it. But are your customers actually happy with those features? And I keep finding in more and more places that people aren't happy with the features that they're getting out there in that way. You know, unless if you're maybe a company like Amazon that has a massive, you know, backlog of customer requests where NLBs get these routes, ALBs can now log to CloudWatch Logs. Those are requested features from the customers. So maybe that's a little bit easier.

[73:15] Justin Brodley: Yeah, I mean, for simple things, it's easy to calculate ROI. I mean, let's say I email out a Japanese word of the day or something to my mailing list and people pay me $5 a year to subscribe to my Learn Japanese website. Previously, I had to think of the Japanese word of the day and type them all in and then maintain the code that like now I can just have an AI that does that for me. I make one call once a day, what's the word of the day going to be today? This is my word every day for the past 3 years, come up with a new one. It could turn like passive income streams can become even more passive in a way because the AIs can now do some of that grunt work. So I think simple things are really easy to to figure out what the ROI is. I mean, my concern is that the biggest ROI is going to be on the elimination of people from roles, but then who's going to prompt the agents?

[74:13] Justin: Do we talk about this last week where I was talking to somebody recently about it where it's like, yes, you will see changes in roles. Maybe there's a short-term dip, but to me, the jobs change. The way I explained it to my dad was like, computers came and everyone said we need a lot less people, except for now we have different people that we need than we didn't need before in our businesses that have different skill sets. So I feel like, yes, you might see some sort of drop, but at the same point you might also see an increase in other areas that you don't know about yet.

[74:48] Matt: Well, and I think one of the areas that's gonna be interesting is, you know, dev team structure. I mean, like we've talked about, for as long as I've been in leadership, you know, a, a Scrum team or a THOT, uh, as we call them, you know, 10 people and you'd be like, okay, 6 of them are developers, there's 2 QA, there's a product person, there's a, a DevOps person, and maybe you have, you know, one other role, an architect, or you got some other thing that's more like a virtual role that kind of steps in when needed, or a security person who steps in when needed. Um, and that's a very, very common model. But in this new AI world, like, yeah, you, you might not have a Scrum team that looks like a traditional Scrum team. Your Scrum team may be 4 people, maybe 2 senior developers and a product person and 30 agents. That might be the new Scrum team. You know, I just, I think we're still figuring that out in a lot of ways. And I think there'll be a lot of experiments at companies. I think there'll be a lot of press that comes out about it and research and what's the right balance of this. But yeah, I agree with you. I think the jobs probably don't completely go away, but they definitely will change.

[75:47] Justin Brodley: I think a lot of sectors, they'll completely go away. If not completely, then significantly enough that it's going to be brutal for the economy. I mean, just the OpenAI story about their presence thing, which is handling 75% of their phone calls. I mean, there's, I can't remember what the number is now. It's like 2% of the US, the number of people employed in the US are employed answering the phones for customer service. Let's say, I mean, it's 3 million people or something like that. So let's say 90% of them lose their jobs and 10% of them become the escalation points for things that the AI couldn't handle. Even that's only going to last as long as it takes for the AI to get even better. So I do think that there are a lot of people who will lose their jobs and will have a very hard time finding new jobs, especially in tech and especially in things like customer service.

[76:42] Matt: Yeah, it's gonna be interesting for sure.

[76:45] Justin Brodley: Yeah.

[76:46] Matt: All right, gentlemen, I think that's another fantastic week here in the cloud.

[76:51] Justin Brodley: All right, we'll see you later.

[76:52] Justin: Bye, everyone. Another week of cloud news wrapped up. Boat will collect the news, Justin will get the notes, Jonathan will write some code, Ryan will watch the perimeter, and Matt will reluctantly watch Azure. Till next week for AI, Amazon, Google Cloud, and Azure, and hey, maybe even Oracle, who knows? Check out thecloudpod.net for our newsletter, join our Slack, message us on socials, or leave a review.

[77:26] Matt: All right, well, we have an after show, or a privacy corner if you will. There's been a couple, uh, interesting privacy events that, uh, CrowdSecurit thought we should talk about. So first up, uh, Microsoft has confirmed that Windows has a global device ID that you apparently can't turn off. Uh, this persistent server-generated identifier tied to a Windows installation, and there's no user-facing toggle to disable in settings. The GUID surfaced publicly because Microsoft handed a suspect's identifier over to law enforcement, allowing investigators to correlate activity across sessions back to a single device and ultimately a person. Technical detail worth noting, the ID persists through Windows updates with changes on clean reinstall and is stored locally in the registry, uh, I'm not gonna read that out loud, but registry path. Though this is a client-side copy of a server-assigned value. This raises questions for enterprise and government customers about data governance, device fingerprinting, whether GUID usage fails under existing privacy disclosure or compliance frameworks like GDPR. The lack of an opt-out is a core privacy concern since it means device-level tracking is effectively mandatory for anyone running Windows regardless of other privacy settings you might configure, which is pretty darn bad. And then you say to yourself, well, you know, if you're not doing anything bad, you know, why do you care? And, but like, if this GDID is exposed in your web traffic or in any other way, then this could be used for ad targeting. There's all kinds of things. And then there's this other story, and I don't know if this person was hiding something or not, so I'm not going to comment on that part of it. But, you know, there was an activist charged with a felony after giving a border agent the duress code for his phone. If for those of you who aren't aware, there's an operating system called Graphene that runs on, I think, Android-type devices. And Graphene basically has a dress code. The dress code, it wipes your phone immediately. Uh, and so this happened during a secondary inspection. He's now facing federal charges as a result. The case raises questions about whether devices wiping at the border constitutes obstruction versus lawful privacy protection. Court filings indicate Tunick was a government watchlist tied to his activism against the Cop City law enforcement facility and CBP. Legitimately planned his detention in advance under a suspected terrorism rationale separate from the child abuse material justification given at the time of the search. Now, I mean, I don't want to comment on if that's what was on his phone or not. That's why he deleted it. I hope it isn't the case and, uh, all that. But just definitely, if you can, uh, be punished for wiping your phone, that's a very interesting invasion of privacy.

[79:44] Justin Brodley: Yeah, but is that really— I mean, yes, I would probably do the same thing.

[79:52] Justin: Yeah.

[79:52] Justin Brodley: But if you were to shred some papers that have been legally subpoenaed, then you, they'd be, you'd be prosecuted for that. I suppose the, the way the law will see that is, um, they have the right to inspect that and, and you are obstructing their, their ability to do so.

[80:07] Matt: Well, I mean, it's also like you're taking my phone at the border.

[80:10] Justin Brodley: Mm-hmm.

[80:11] Matt: And you're asking me to unlock it, and then as soon as I unlock it, you're gonna basically dump all of the memory into your system that's gonna parse for anything impossible, which is almost like You know, is this a situation of where like wiping my phone is pleading the Fifth? I'm not going to self-incriminate myself by giving you my phone. I don't know. I think it'd be an interesting bellwether case, I think, is kind of, you know, how does it get looked at? How does it become with the authority of the Border Patrol? I mean, there's been a lot of reports lately of them, you know, saying we need to access your device to see if you're doing anything, or, you know, or speculative thing searches at the border. Um, so I mean, from a privacy perspective, I I get it. Also, what if the agent, you know, just types the password in wrong 10 times and wipes the phone? Because that's how my iPhone's configured. You try to break into my phone, after 10 minutes it locks you out, and you do it too many times, it wipes the phone. And it happened to me because my kid, when he was a baby, you know, was playing with my phone and wiped my phone. So, you know, it happens.

[81:07] Justin: I disabled that feature for that specific reason.

[81:10] Justin Brodley: I mean, what would happen if— I mean, I don't know the ins and outs of this specific case, but what if, uh, under the stress of the situation, you give somebody the wrong code accidentally, you know?

[81:19] Matt: Well, that's his defense. He's saying, I didn't— it was the only code that came to mind when they were pressuring me for a code, and that's the code I gave them, and it happened to be the duress code. That's actually his defense.

[81:28] Justin Brodley: Okay, I mean, that's a fair defense. They can't prove otherwise. I mean, yeah, I don't know. I mean, it's— I don't use Face ID unlock, I don't use fingerprint unlock. Any of those things can be used against you. I have a PIN. I couldn't tell you what the PIN is. It's a shape. I couldn't even tell you what the numbers are. So it's, it's kind of nice.

[81:46] Justin: Oh, you have the Google one with the, with the shape that you draw like they originally had?

[81:51] Justin Brodley: Uh, it's, it's the numpad, but I, I remember the shape of, of the code, not, not the actual numbers.

[81:56] Justin: Oh, okay.

[81:56] Matt: And so like, so what do you do when you end up with those keypads that they've, they've cycled the numbers? Like, does that mess you up?

[82:02] Justin Brodley: I, I wouldn't be able to use it.

[82:04] Matt: Okay.

[82:06] Justin: No, he's like, it's a triangle, square, and a line. Okay, cool. Now everything rotate. Good luck.

[82:13] Matt: Good luck to you.

[82:15] Justin Brodley: If the guy, if the guy had, did reasonably have something to hide, like, why not wipe your phone, back it up to the cloud, wipe your phone, go through the border, restore it the other side? I mean, yeah, you know, there's, there's so many, there's so many other ways around it.

[82:28] Matt: I mean, the question, I mean, their argument is that he potentially had, you know, abuse material on there, which I hope is not the case, but maybe it is, I don't know. Um, so that was the speculation that they gave him for getting access to his phone. But like, But you know, again, you're at the border. I don't know what warrant rules are at borders. It's clearly not what I would think it would be. But you know, also, you know, the other question here is, was he targeted because he's an activist and he's been fighting the government on other things? And so was he specifically targeted, which is a whole different privacy conversation as well. So there's lots of ickiness in this particular case.

[83:05] Justin: Yeah.

[83:05] Matt: But I mean, I think the technical question of it is like, you know, should you be is duress considered obstruction of justice or not? Uh, that's, that's an interesting one. I don't know.

[83:15] Justin: I feel like it's a fine line of like, are you shredding things in front of the police officer? And you made the comment of like, do you really care if you don't have anything to hide or not?

[83:26] Matt: But it's kind of part of the American philosophy too, is that the reality though is like when they take a phone and they unlock it and they dump it into their system, Uh, they're taking everything. And then, you know, a good prosecutor or a good defense attorney is going to snip the pieces of text that paint the story or narrative they're trying to use in court.

[83:47] Justin Brodley: Absolutely.

[83:47] Matt: So it's, you know, if it was always kept in context of like, well, here's all the text messages around this one thing I said, and I can clearly see that I was joking when I told, you know, I was joking with Jonathan, like, if you don't show up today's podcast, I'm gonna kill you. You know, if that's the text and it's just a joke, like, 'cause, you know, he missed last week.

[84:03] Justin Brodley: Sorry. I mean, I, I got the text. That's why I'm here today.

[84:05] Matt: Yeah. Um, but you know, like, but then like, you know, basically, you know, something happens, my phone gets scanned, they take it and they're like, oh, Justin was, you know, threatening murder of Jonathan.

[84:15] Justin Brodley: Uh, yeah, yeah, exactly.

[84:16] Matt: Premeditated murder. Like, well, no, that was a joke. And in context you could tell it's a joke because that was, their text threads are always like that. Like, are you gonna show up today? Yes or not? Blah, blah. Just one of those, those fun things. But again, like, the context is the problem in a lot of those type of cases. So if you give them your whole phone, they can probably always find something to burn you on.

[84:33] Justin Brodley: Or put something on the phone, honestly.

[84:35] Matt: Or put something on it. Yeah, yeah.

[84:36] Justin Brodley: I wouldn't be surprised if somebody wiped their phone and handed it to, to border security and it came back with something installed on it.

[84:43] Justin: That to me is the bigger concern too about giving out those types of things, is like, once it's unlocked, as you know, you can do a lot with someone's phone. Install an app that looks— was it the calculator app that was the, you know, the FBI used for years because it had like the secret messaging in that they sold out to— they sold to all the people around the world? Like, you can do crazy stuff like that and people won't notice, and maybe then you're tracking someone that way, you know, etc., etc. Like, you can do a lot of stuff once you get into somebody's device. Agreed.

[85:18] Matt: Well, I mean, uh, Jonathan, you mentioned that you don't use selfies you don't, you know, anything. But I did see an article that Google is now allowing you to do selfie videos to unlock your phone. So if you know that little drawing you're doing doesn't work out for you, you could go to TikToks as your unlock method. You get those steps right on the dance, that's the key thing.

[85:36] Justin Brodley: Yeah, I mean, I hope it's more reliable than Facebook's attempt at the same thing. My daughter, who has a fairly unique name, I would say probably one-off in the world, she had a, she opened a Meta account so she could use Facebook Marketplace and they shut it down. They told her she was impersonating somebody else. Like, I guarantee there is not another person in the world with the same name. And so, you know, she went through this whole process of recording the video, look left, look right, look up and down, do all this stuff. She sent it off and they're like, no, no, you're still not the right person. I'm like, what? I hope whatever Google are going to provide is more reliable than what Facebook built.

[86:17] Matt: Hopefully. All right, gentlemen, I think we've killed this privacy conversation to death. But, uh, interesting one. We'll keep an eye on it and see if what, uh, how that one moves to the courts. But, uh, yeah, privacy, always a fun debate.

[86:28] Justin Brodley: Yeah, there is no privacy anymore. No, you can, you can have zero convenience without privacy. You know, things cost more with privacy. You don't type your phone number in a Safeway, your groceries cost 20% extra because you don't get the discounts.

[86:41] Matt: I mean, I try to use DuckDuckGo as my web browser or my search engine, and I just, I give up after like 2 hours. I'm like, this is terrible. I hate everything about this. Go back to Google. And I know I'm giving up an amount of privacy for that, which sucks, but, uh, you know, I just, you know, convenience is there and, you know, unfortunately I'm a pleb, I guess, that way. I don't care. Yeah. Like I want the ease. I want the easy. Now if I could turn off those damn AI summaries, uh, I'd be pretty happy with that too. But, uh, yeah, here we are. So, all right, gentlemen, see you next week.

[87:09] Justin Brodley: See you later.

[87:10] Justin: Bye. Bye.