351: IAM the One Spending All Your AI Money

[00:00:00] Speaker A: Foreign. [00:00:06] Speaker B: Welcome to the Cloud Pod where the forecast is always cloudy. We talk weekly about all things aws, GCP and Azure. [00:00:14] Speaker C: We are your hosts, Justin, Jonathan, Ryan and Matthew. [00:00:18] Speaker B: Before we get into this week's news, we want to take a minute to tell you about We Are Developers World Congress, which is finally making its way to North America this September. If you spent any time in the European tech scene, you probably know the team behind it. They've been running World Congress in Berlin for over a decade and it's a big deal over there, pulling in more than 15,000 developers every year. Our friend Kote from Software Defined Talk is actually speaking at the Berlin event this July. And from what we've seen, these are the people who know how to put on a good developer conference. This September 23rd through 25th, they're bringing it stateside to San Jose. Organizers are expecting more than 10,000 developers with over 500 speakers across 18 different content tracks covering the entire stack, including Cloud, DevOps, AI Security, Software Architecture, data engineering, front end and developer experience. If you've got a team, everyone's going to find a full schedule. It's not just sit and listen sessions. There are keynotes, workshops, masterclasses and hands on labs. The kind of stuff you can take back home and work on. On Monday, there's an impressive list of speakers including names from Datadog, Honeycomb, Sentry, Google, LinkedIn, Stack, Overflow, Netflix, Microsoft and Stripe plus Kelsey Hightower, Oliver Pommel, Christine Yen, Scott Hanselman and Angie Jones. Head over to werdevelopers us to grab your ticket and use code DEVPOD26 for 15% off that stacks with their group rates. If you're bringing four or more people and honestly at that price, you should probably bring the whole team. [00:01:50] Speaker A: Episode 351 recorded for April 14, 2026 I am the one spending all your AI money. Good evening Matt and Ryan. How are you doing? [00:02:01] Speaker D: Good, how are you doing? [00:02:02] Speaker C: Good. Feeling rested? [00:02:03] Speaker A: Yeah. You went on vacation to beautiful, beautiful warm tropical places while it poured down rain here in the bay, so you took the perfect time to go. [00:02:11] Speaker C: Yeah, I did thread the needle there because it was. I, I was in between two storms in, in Hawaii and then came back here and now it's sunny again. [00:02:21] Speaker A: So yeah, perfect. It's. I think it was rain this weekend, but I'm going to Vegas so it doesn't really matter to me. [00:02:26] Speaker D: I don't think you're inside the entire time. [00:02:29] Speaker A: That's true. Just inside a conference room. That's all I see. At least, at least at the Mandalay Bay, you walk by that big wave pool they have, and so it's like, oh, that looks so nice and peaceful out there when I'm dealing with bajillions of people trying to crush into conference rooms. Although these last couple years, the executive room, where you have social executive privileges is downstairs in the basement, which is the best place to put executives. Honestly, I don't know where else they put us. So I kind of missed out on a little bit of the beach front because I have to drop down this little escalator to go down there, so I missed some of that, but it's all right. Well, last week, while you're out, Ryan Anthropic shocked the world with Mythos, which apparently is a new, very expensive AI model that can find vulnerabilities and attack them. And so one of the big items of that was that they tested against things like FreeBSD and they found a bug that they said had been there for 25 years, which is true, it was there. But, you know, even last week, I think we were a little bit skeptical about how great this model is because Anthropic, they're a bit overhypey on the marketing. So there was a lot of talk, a lot of chat, and so now the reality has come to bear. And so I'll probably have the best article I saw where they basically, you know, looked at Mythos and Project Last Wing and then put it to its paces. And basically, using inexpensive open weight models, they were able to replicate basically all the vulnerability work Anthropic did and attribute it to Mythos, where they tested eight models, all detected the flagship FreeBSD NFS buffer overflow, including a 3.6 billion parameter model. They did notice that one was pretty legit. So it's not completely bullshit, it's just partially. Which is kind of what I expected to have happen, is, you know, it'll be really impressive and really cool and it can do some really cool things. But it may be not as great as Anthropic made it out to be, which is basically what they found set, essentially. It is great. It does do a lot of things. It is very expensive. Be careful. Almost 6x the cost of opus, which is already really expensive. [00:04:21] Speaker B: Whoa. [00:04:22] Speaker A: But definitely, if you are in the security space and you want to have it poke holes at your app, it uses really complicated patterns to basically figure out different attack vectors and can actually link different vulnerabilities together to actually exploit your thing. So I assume red teams will use this a lot. [00:04:36] Speaker C: Mm. [00:04:36] Speaker A: I Assume others as well. But you know, is it the end all, be all panacea that's gonna end all, all software as basically they were kind of implying, which is why they kept it restricted to such few number of companies and via the API. I think the reality is that because it's so heavy on the GPUs that if they lose it to everybody, they don't have enough capacity. [00:04:56] Speaker C: Right. So the oceans would immediately raise and. [00:05:01] Speaker D: Yeah, yeah, if you use it for targeted use cases, you know, it will make sense. So if you have part of your app that's unauthenticated, that's general information, or you have anything specific that's deployed to end users, devices or anything like that, I could see it making sense to like run this more on a reoccurring basis, but to me this is more like let's target specific areas, then let's run our entire app through it and see what vulnerabilities pull. [00:05:25] Speaker C: It'd be interesting to like sort of add this to, you know, if you've got any kind of like attestation, you know, or framework that you're sort of signing off on controls. It'd be interesting to have this sort of be part of that or just an annual review. Right. I like that, you know, these tubes things are great because it comes up with things that maybe I didn't come up with and you know, no, you know, now we, you know, I haven't worked on a red team so like I'm. I know that my methods are less sophisticated than what I've seen and you know, know. But having a team that you staff permanently is expensive. [00:06:00] Speaker A: So it's. [00:06:01] Speaker C: Yeah, I think it's a cool tool. [00:06:03] Speaker A: But I, I've used, you know, basic models as well to define security vulnerabilities in my code already. So I mean like, there's definitely value and so, you know, the fact that this is that much better at reasoning and much better to actually make it more complicated things, I think it's valuable. Again, I think it's to your point, it's red team level, someone who really knows how to like guide a model, which is again, I think why it's limited to certain partners. It was interesting Cloudflare, who was not one of the launch partners this launched on day of recording, so it wasn't a lot of time to play or to look into this. But you know, Cloudflare stock went down like 5% because they weren't included in the list of security vendors who were included in the. The limited access to it which was so silly because like the market is just reacting because I'm like. But it's available via the API and cloudflare can access the API just like everybody else. So just because they're not a launch partner doesn't mean anything. [00:06:52] Speaker C: I feel like the market is very. [00:06:53] Speaker A: The market is just crazy. Yeah. Basically anything that's a threat to potential valuation just kills you right now, no matter what you do well. [00:07:00] Speaker C: And then it'll recover a week later when everyone forgets like it. It, like I. It seems crazy to me. [00:07:05] Speaker A: Well, I mean they are. You know, there's a war going on, other distractions, so it's hard to keep a memory and apparently there's a lot of ADHD and investors I don't know so well. Amazon this week had a interview with basically Andy Jassy and Matt Garman was involved, et cetera. But basically they've invested $8 billion in anthropic and 50 billion in OpenAI creating a situation where it holds significant financial stakes and two directly competing A model companies. And so there was question was, you know, basically is this a conflict of interest for you to own so much in two different companies as well as your own Nova models? And AVA CEO Matt Garman frames this as consistent with Amazon's long standing practice of partnering with companies it also competes against against citing Oracle selling its database services on AWS as an established precedent. I mean that's a pretty new precedent. That's not the one I would have chose. The dual investment was partly driven by competitive necessity. [00:07:55] Speaker D: So Oracle's been RDS for as long as I've been using rds, so way too. [00:08:00] Speaker A: But originally it was not a partnership, it was a we screwed you in the licensing negotiation because you didn't know what we were going to do with it. Correct. [00:08:07] Speaker D: But that's what they're referencing, screwing somebody over. [00:08:10] Speaker A: I thought they represent the current new service which is, you know, Oracle Cloud on aws. It's interesting too because there is also and we didn't cover this week because it wasn't worth talking about its own, but OpenAI basically said, you know, getting a partnership with AWS has actually been way better for them than the partnership with Azure and helped open the door to them to a bunch of new enterprise clients. And I'm like, yeah, because the power of the Amazon marketplace is pretty strong. [00:08:32] Speaker D: Shocking. Utterly shocked. [00:08:35] Speaker A: Not shocked at all. Well, next week is Google Next in Las Vegas and it came upon all of us very quickly. So some of us have done research and some of us have not. But we are going to do our annual prediction game here where we try to see if we are as good as Google product management and things that we think will be announced. And I believe, you know, our typical rule fashion, this is a draft style. So before the thing, Bolt drew numbers for all of us and I won. Ryan is second and Matt was third and Jonathan was fourth. But Jonathan's not here, so he gets. He gets nothing. [00:09:11] Speaker D: Hey, Ry, did you ever audit Bolt code to make sure that Justin didn't make it? So he's the first one every time [00:09:17] Speaker A: when, you know, when it picked me first, I knew you guys were going to say that. [00:09:20] Speaker C: Yeah, a little sus. [00:09:23] Speaker A: I mean, if it happens next time, you definitely should call me out. But it is purely a randomizer. I was very clear with it that it can't have preferences. So. Yeah. So anyways, I have the first pick for this week's goal next, which is ironic too because I did the homework. So I'm more prepared than both of you probably are for this. But so my first prediction this week is going to be that Wiz and Google Cloud will announce some new big integration feature on stage. This will be the first big announcement since the acquisition is closed. And so that is my first prediction is a Wiz Google Cloud partnership or a product offering of some size will be announced. [00:10:02] Speaker C: So not only do I not have I not done the research, you're stealing my first one. [00:10:09] Speaker D: That was my second one. I will. That wasn't my first choice. [00:10:12] Speaker A: Nice, nice. Well done. So that's my first one. So that puts you, Ryan, who now has fantastic. [00:10:18] Speaker C: I've got one left. And then I have to frantically, you know, find more while hopefully Matt is taking takes a sweet time. I think that they will. They haven't really announced that Gemini 3.1 Pro is General availability. And so I think that they will announce general availability of that model while also teasing the next set of models. [00:10:39] Speaker D: So that was my number one choice. Just saying. [00:10:41] Speaker A: So that was my second choice. [00:10:47] Speaker D: All right, we're doing well here, guys. [00:10:48] Speaker A: You're doing great. All right, Matt, you are up for your first pick. [00:10:54] Speaker D: Let me just delete my entire list. So go from there. I'm going to go with. They're going to talk about more default guardrails in AI in general. So you know, all the different places they use it, how they're adding the guardrails into Gemini itself, where they add the guardrails because that's a big, you know, theme. So just general guardrails in AI. [00:11:21] Speaker A: I Think Ryan would tell you that he hopes to God because what they have is pretty bad. [00:11:26] Speaker D: Yeah, they're all pretty bad. So yeah. [00:11:30] Speaker A: Uh, all right, my next one I think they're going to announce Anti Gravity for enterprise. It's been kind of consumer focused so far. It's not had a lot of, you know, if you wanted to use an enterprise you had to have like a Google Apps Ultra subscription. And I think it's going to be that plus some major enhancements of Gemini. CLI will be the big announcement for the dev side. Hmm. [00:11:56] Speaker C: Trying to figure out if that steals my next one. [00:12:02] Speaker A: But it's put you up for your next one. So we'll see. [00:12:05] Speaker C: We'll see. So I'll try it anyway. Like so I was. [00:12:08] Speaker A: I. [00:12:09] Speaker C: Since Gemini doesn't really truly have sort of the agentic framework that you see in cloud code and now with. With Cowork and even some of the you know, the code specific tools like cursor and GitHub copilot where you can branch off into multiple agents and complete workflows. I think Gemini or I think that Google will announce either Gemini's ability to sort of run those multi agent workflows. [00:12:36] Speaker A: Yeah, they don't. Do they have an agent platform today in Vertex? [00:12:41] Speaker C: I don't I mean they have their sort of agent development. [00:12:45] Speaker A: They don't really have a run. [00:12:46] Speaker C: They don't really have like yeah, there's [00:12:48] Speaker A: nothing unless the builder also does the [00:12:50] Speaker C: running deploy even in general enterprise that like you can't do a complex workflow. You can sort of create an agent publish. [00:12:57] Speaker A: There's an agent Vertex AI agent engine [00:13:00] Speaker C: that exists but that's still development of [00:13:03] Speaker A: isn't it Enables developers deploy manual scale a agents in production agent engine handles shoulder scale agents and productions. You can grow some creating applications. I. Yeah, I mean I don't know. I think, I think there will be something in that space. If you want to make it broad to say something enhancements with agents, I would be fine with that too. Okay. [00:13:21] Speaker C: Yeah, I'll take, I'll take the broad recommendation. [00:13:25] Speaker D: They're going to announce three new features. That too broad? [00:13:29] Speaker A: No, you can do that if you want. [00:13:33] Speaker C: Except for when there's four. You lose. [00:13:35] Speaker D: Yeah. [00:13:38] Speaker A: All right Matt, you're up for your second. I have a long list. Would you like. Would you like some. [00:13:43] Speaker D: Google's going to spend a couple minutes talking about just agentic coding practices and tooling and you know how do really it's going to be like how developers are leveraging, you know more agentic Coding. I don't want to say philosophies but [00:13:58] Speaker A: you know, so your Marian's got more on the runtime for apps. You're thinking more on the SDLC. That's how I would segment these YouTube. Okay. So I'll say it's very close to Ryan's. Let's see. Let me go through my list here. See where I want to. Where I want to go. With my third and final pick we will go to a new Ironwood Will GA and a new inference based chip. [00:14:27] Speaker D: That's their tpu, correct? [00:14:29] Speaker A: Yes. Ironwood TPU is their custom silicone for is like Trainium or Inferentia. All right Ryan, what's you up for your third and final? [00:14:40] Speaker C: Trying to think of something that's not AI related and it's really tricky these days because there's just not a lot talked about. [00:14:48] Speaker A: Yeah, I mean like I would say security is probably a strong area. [00:14:52] Speaker C: Yeah. But. But I do think that's kind of covered with wiz. I think they're. They're focused. [00:14:57] Speaker A: Yeah. I mean I do think we're probably due for a new Axion chip this year, you know, which is their general purpose custom silicone. Yeah. Firebase maybe will have something. Yeah, I don't have anything that's really somewhat AI agentic or it's really fun. Maybe you go robotics because the Gemini Robotics API isn't, you know, could be out there. Is in developer preview today. It's not in GA yet. Yeah. I don't know. [00:15:24] Speaker C: Like it's funny. I always think that space is going to die with the cloud providers and it sort of goes quiet and then comes back every once in a while [00:15:32] Speaker D: you go into like Google workspace in [00:15:34] Speaker A: some way you do something on Sovereign. Yes. Data. [00:15:39] Speaker C: I had data sovereignty but it was like more data sovereignty. Doesn't seem like a great prediction which is what I think will actually I [00:15:47] Speaker D: had a runner up of a. Of at least you know, obviously choose a number one, two or three, you know, new regions getting announced like at least one new region gain outs. But I always going to say that as a runner up [00:16:00] Speaker A: they have not traditionally announced new regions but that's why [00:16:03] Speaker D: it's a runner up. [00:16:04] Speaker A: Yeah, definitely runner up. [00:16:06] Speaker C: Yeah. I was trying to think of like what could they do to transform that. [00:16:10] Speaker D: Does Google have a like Azure Local or Outpost? [00:16:15] Speaker A: Yes, they do. [00:16:16] Speaker C: All right, let's see. I'm going to choose one, the only one I can think of and by me I mean Claude. I think that they will announce some major VMware interruption based off of Kubernetes, like a. [00:16:37] Speaker A: Like a on ramp to Kubernetes that gets rid of your VMware clusters. [00:16:42] Speaker C: I was thinking more about the, you know, the. There was the play that VMware themselves tried to make a few years ago, which is more of like the sort of, you know, management platform of Kubernetes workloads. But I think this will go the other way, which is you'll have a VMware like experience, but it'll be based off of, you know, the. It'll be underpinned by GKE in the storage product. [00:17:12] Speaker D: Okay, so kind of the inverse of Tanzu. [00:17:15] Speaker A: Yes, thank you. [00:17:17] Speaker C: I couldn't remember the name of it. [00:17:19] Speaker D: I googled it because it was driving me crazy that I couldn't remember what Tanzu was called. [00:17:25] Speaker A: All right, Matt, your third and final. [00:17:27] Speaker D: I'm gonna go more generic. I'm gonna do with what I did last year thanks to Bolt, which is at least three non AI announcements. [00:17:38] Speaker A: That's. That's bold. Three. [00:17:40] Speaker D: I know, right? [00:17:40] Speaker A: Maybe. I think, well, one would be a blessing. [00:17:43] Speaker D: Like Riot did one. Right. Kind of the opposite. Vm where I was like, all right, well if I do two, it feels air. I guess they could do one and it matches for both of us. [00:17:54] Speaker C: Price is Right rules. [00:18:00] Speaker D: I had a few runner ups, if you want to add them. [00:18:02] Speaker A: Yeah, if you want to. Just want to put them in the article. We'll talk about them here. So I had a few runner ups. So one was agent to agent protocol. Probably hit 10 this year. They talked about Turboquant a few weeks ago and I thought it was sort of interesting timing that they would bring up that they've got this new memory optimized play for AI models. They said it was research, et cetera. Is it as research? I don't know. Maybe, maybe not. So I was thinking, you know, they may ship something with Turboquant, which would be a huge elite frog, potentially in Gemini 4, something like that. And so I think that's a good one. [00:18:31] Speaker D: I had something with Waymo. I feel like there'll be something there. I don't know exactly what, but you know, their whole self driving car, I feel like I've seen more recently a bunch of stuff in general around self driving cars, you know, with the taxis and everything. And maybe some sort of integration or just dumb talking about other ones. And then really just because I like to say nano, nano Bananas is, you know, some sort of update to that one, to that model. [00:19:00] Speaker A: Just because it's fun to say Nano Banana. Yeah, it's good. [00:19:03] Speaker D: Yeah it's just. It's just fun to say. I'm not gonna lie. [00:19:09] Speaker A: Gemini Robotics. Yeah so then I also had accent gen 2 we just mentioned trying to give it to. Trying to give it to Ryan or you either one of you took it do think Gemini 3.1 Flash will hit GA this time as well. BigQuery AI agents seem like an obvious one to me. Sovereign Cloud AI because like things that S3SN and their partnerships and some of the other sovereign things. You know more AI expansion there. The Gemini Robotics API going public preview makes sense to me. In my research I had Claude try to give me some more wild there things like. And they had Google acquires or partners with hugging face which that would be crazy to me so hugging face free tier Gemini enterprise for startups like kind of like the AWS activate program. It was something that Gemini thought might make sense and then the agent to agent payment protocol going live would be a big surprise I think this time because it's still pretty early and we haven't really seen a lot of adoption on that one. So I think that's probably not ready yet but we'll see. [00:20:08] Speaker C: Yeah I think monetizing AI is still going through that sort of, you know, finding itself phase where I think you know there was the initial it was too expensive, no one bought it. Now people want it and they're charging more but it's people are finding the limits of you know, tokens and how expensive it could be. So it does make sense that that would expand to the agent to agent monetization next. But I don't know how long it'll take. [00:20:35] Speaker A: Yeah I do. I do want that. With the Mythos stuff coming out I thought maybe if depending on you know because I watched talk about later today OpenAI actually can drop something to compete with Glasswing. I wonder if there's a Gemini model that's similar they could drop and I just don't know if you can like Glasswick has been kind of out there for a little bit as a rumor then you know they finally announced it officially but if they have something at next that'd be interesting. So again that's aren't mine. [00:21:01] Speaker D: Yeah the other couple I had was more AI integration into my phone because that's really what I need in my [00:21:06] Speaker A: life and I assume that'll come at big. [00:21:08] Speaker C: I think they have the different conference for the. For the Android announcement. [00:21:11] Speaker A: Yeah that comes at I.O. typically or I.O. [00:21:14] Speaker D: whatever it's called. Yeah. And then the other One I had was there's been some rumors I've read around Boston Dynamics and all the robots and leveraging Gemini more. I've seen a few random articles fly by over the last couple months, but it'd be something interesting to see that's too specific for announcement now. I feel like a specific company partnering, which I felt too specific. [00:21:38] Speaker A: I don't know. [00:21:38] Speaker C: Their robots still freak me out. [00:21:43] Speaker A: The Dogwood, specifically. [00:21:45] Speaker C: Yeah, yeah. Like, I don't. I don't. [00:21:47] Speaker D: Yeah. [00:21:48] Speaker C: I don't know if I want that smarter. [00:21:51] Speaker D: Yeah. [00:21:51] Speaker A: I don't need you to get better. All right, well, I think that's a pretty good prediction. Our tiebreaker is how many times would they say AI across three keynotes? This is not include the partner keynote. This is just the developer one, the morning Wednesday morning. And here's a Thursday morning keynote, if I recall correctly from my schedule. But whatever the keynotes are, minus the partner one, because no one cares about the partner keynote. So last year that number was 101. Because we did look it up in Bolt. Thanks, Bolt. So in reverse order, Matt, you get to pick your number first. [00:22:27] Speaker D: I'm gonna go 99. [00:22:31] Speaker A: Nice. Ryan, [00:22:36] Speaker C: is it going to be more or less this year? I think I'll. I'll go a little bit less and try to give some space where I'll go 75. I think we are not the only ones that are getting tired of it. [00:22:51] Speaker A: They. [00:22:51] Speaker C: I mean, they. They will talk about it, but I think it'll be a little less. AI. AI. [00:22:55] Speaker D: AI. Hopefully. [00:22:57] Speaker A: I mean, this. [00:22:58] Speaker C: Maybe that's just wishful thinking. [00:22:59] Speaker A: This is spoken by someone who's not looked at the course catalog for next week. [00:23:02] Speaker D: No, I have. I know that. [00:23:05] Speaker A: Yeah, I'm gonna go. I don't think it's gonna be crazy more than this, this last year, but I think it will be more. I'm gonna go 115. I was gonna go Ryan's model that go like 75, try to hit the low, but you took it already. So I'm gonna go over and go 115 and assume it's a little bit more this year, because the AI is just that much more prevalent this year than it was last year, shockingly enough. So. All right, well, let's see how we do next week. I think Matt and Jonathan are gonna hold down the fort while Ryan and I are there, and we'll report back the following week. [00:23:36] Speaker D: Whoa. Putting us on the spot. I see how it is. [00:23:39] Speaker A: That's how it rolls around here. It's called listener accountability, sir. [00:23:45] Speaker C: It's the only accountability we have. [00:23:49] Speaker A: All right. AI is how ML makes money this week. First up, Anthropic has launched Cloud Managed Agents in Public Beta, a suite of composable APIs that handle production infrastructure like sandbox code execution, state management, credential handling and end to end tracing so developers can focus on defining tasks and guardrails or rather than building backend system systems. The platform includes long running autonomous sessions, multi agent coordination and preview and trusted governance with scope permissions and identity management with internal testing showing up to 10 percentage points improvement and task success over standard prompting loops unstructured file generation tasks. Pricing is consumption based on standard cloud platform token rates +0.08 per session hour for active runtime which position this as a managed alternative to self hosted agent infrastructure where teams could otherwise spend months to set up worshipping anything to users. Early adopters included Rakuten, which deployed specialized enterprise agents across five business functions within a week, and Sentry, which shipped a bug to PR pipeline in weeks and seven months. By pairing their existing SEER debugging agent with a cloud powered patching agent, developers can start get started with the cloud console, the new CLI or by using cloud code with a built in cloud API skill. The Multi agent Coordination and self evaluation feature is still gated behind your Research Preview access request form. [00:25:01] Speaker C: So I don't have to get a fleet of Mac minis to run all my AI things. [00:25:06] Speaker A: Sure don't. [00:25:07] Speaker C: Yeah, it is this interesting, you know, security debate which is you know, having a secure, you know, infrastructure and to run those agentic workflows. But it's also, it's a hassle and you know, I wonder, I wonder how many people are just bypassing that and just running it. You know, full access to everything. Not sandboxed. [00:25:28] Speaker D: I'll always use cloud but when I do I run it with, you know, dangerously allowed everything to my root system. [00:25:34] Speaker C: Yeah auto approve everything. [00:25:36] Speaker D: Yeah. [00:25:38] Speaker A: Someone I saw someone today say I got burned by you know, fully dangerous mode today and I was like Auto mode is the way to go. Auto mode is a safer choice. OpenAI is reporting enterprises now account for more than 40% of revenue and is ready to reach parity with consumer revenue by end of 2026. With APIs processing over billion tokens per minute and Codex reaching 3 million weekly active users, that's pretty impressive. For Codex. OpenAI Frontier is positioned as a company wide agent deployment and management layer distinct from single product agent implementations, allowing agents to operate across an organization's tools, systems and data with centralized governance and permissions a new stateful runtime environment being co developed with aws, is designed to give agents persistent context and memory across sessions, addressing a core limitation for complex enterprise workflows that span multiple tools and data sources. OpenAI is building towards a unified AI super app that consolidates ChatGPT, Codex and Agent browsing into a single employee facing interface. The stated goal of reducing enterprise rollout friction by leveraging ChatGPT's existing 900 million users who are already familiar with the interface. I can tell you from experience that super apps are somewhat problematic as well. So permission boundaries become a big problem really quickly. [00:26:47] Speaker C: Yeah, this, I mean this sounds great. [00:26:49] Speaker A: It's it. [00:26:51] Speaker C: All these AI models are only as good as the data that they have access to. And when you get into the enterprise, you know, you're trying to integrate with all the IT services and other platforms that are used for development or other parts of the business. Design tools, there's all kinds of stuff and it's really tricky to sort of manage that. And so, you know, I've seen two models where it's, you know, you're kind of left to your own devices, you know, setting up your own MCP server or your own local integration somehow or there's just if there is a platform, you know, sort of a sparse support of that. So I'm really happy to see this developed and I'm really, you know, really eager for this type of framework to be more prevalent because it is cool. [00:27:41] Speaker D: Like any of these businesses, you know, things are sso, things are locked down, so it's sometimes it's all or nothing, it's kind of the problem or an end user that works around the security controls which makes Ryan and explode because he's also working around them, I'm sure. So, you know, well, how else know you for too many years know how your brain works? [00:28:03] Speaker C: Exactly. [00:28:04] Speaker D: You know, like if you're going to build it, you might as well build it, right? And it's just, it's going to take time and effort to do those things. There are a lot of cloud cost management tools out there, but only our chair provides insured commitments. It sounds fancy, but it's really simple. Archera gives you the cost savings of a one or three year AWS savings plan with a commitment as short as 30 days. If you do not use all the cloud resources you've committed to, Arterra will literally cover the differences. Other cost management tools may say they offered insured commitments, but remember to ask, will you actually give me my rebate? Artero will check out thecloudpod.netarcheror to schedule a demo today. [00:28:57] Speaker A: Well, if you remember about a year ago meta released llama 4 models and then basically got real quiet and that's because they've basically been rebooting their entire AI strategy at Meta and they're now releasing their first model from its new Meta Superintelligence lab called musespark. Available now at Meta AI with a private API preview open to select users. Users this is a natively multimodal reasoning model, supporting tool use, visual chain of thought and multi agent orchestration. A new contemplating mode orchestrates multiple agents reasoning in parallel, achieving 58% on humanity's last exam and 38 on frontier science research. Positioning it alongside extreme reasoning models from Gemini, Deepthink and GPT Pro. Meta claims its new pre training stack reaches equivalent capabilities with an over with over an order of magnitude less compute than Llama for Maverick Required, which has direct implications for cost and efficiency at scale, including their new Hyperion data center investment. The model uses a multi agent test time scaling approach that delivers stronger performance at comparable latency versus single agent extended thinking and applies token compression via thinking time penalties to optimize reasoning efficiency for serving at scale. Apollo Research identified that Muspark showed the highest rate of valuation awareness of any model that they tested, frequently identifying scenarios as alignment traps. Meta concluded this was not a blocking concern for Elispark knowledge. It warrants further research and something you should keep in mind if you're going to use these in any production level that they they detect evaluations. That's sort of creepy. [00:30:21] Speaker C: Is this an open model like llama 4 was? [00:30:24] Speaker A: You know, I wasn't super clear on that if it is or not because [00:30:30] Speaker C: that it was always sort of surprising. Just. Well mostly because they were the first big company to sort of release their own model and it, you know, the fact that they did it open was very different than a lot of the other, you know, companies in their space. [00:30:48] Speaker A: It is not an open model. It is a proprietary closed source model. [00:30:51] Speaker C: That makes sense, that makes more sense. [00:30:54] Speaker D: I mean one, I may be cynical here from the new Meta Superintelligence Lab division. If I have to go work anywhere, I want to work for the Super Intelligence Lab division because who doesn't want that? [00:31:07] Speaker A: Your job title probably also has the most stock options too, so you definitely want that. [00:31:13] Speaker D: Yeah, and then I've seen it before and every time I see humanity's last exam I'm always like can we really name this something different guys? [00:31:24] Speaker C: No, I love that. It's like I For one. Welcome. My new robot overlords just own it. [00:31:32] Speaker A: I mean, it is a. I was just looking at the latest human exam, so it still says Gemini 3.1 Pro and GPT 5.4 and Opus 4.6 are leading on that. So, you know, this new model from Meta doesn't even. I don't even see it a moment just glancing at it. But it is interesting. You know, this is one of the main leaderboards that people are using to see which one destroys us first. Thanks. [00:31:54] Speaker D: Yeah, I mean, overall, look, it's impressive. They got a full order of magnitude lower of conception like that will give them a lot more power in the future. And that's one thing I'm wondering is a lot of these, you know, incremental updates they're doing, you know, I guess Class Wing is, you know, clearly a couple or, you know, much better. But is it just they threw more power at it and gave it more of a resource set or is it they actually, you know, redid the way it's done and Meta going back to kind of the ground level and rebuilding it from the ground up. At least from what I understand, this means that they are able now to take advantage of newer things, newer technologies that might give them an advantage for a period of time until the Claude and the OpenAI and GPT models and all those kind of have to go through that because there's going to have to be some sort of shift between 5.2 and 6 or 4.6 and 5 of cloud. [00:32:59] Speaker A: Yeah, I'm sure. So the thing about the what's on humanity's last exam right now is that it the last updates from February 20th. So we're just waiting to see when mythos and this new meta one gets added to it. So that'd be interesting. All right. Anthropic is launching routines in cloud code as a research preview, letting developers configure automated workflows once with a prompt, repo and connector, then run them on a schedule by API call or in response to GitHub events without requiring a local machine to be running. Three trigger types are supported Scheduled cadences hourly nightly or weekly API triggered endpoints where each routine gets its own URL and Auth token and GitHub webhook events that spin up new sessions per matching PR and continue feeding it updates like comments and CI failures. The Cloud hosted instruction removes the need for developers to manage their own cron jobs, MTP servers or additional tooling since routines ship with built in access to repos and connectors daily Routine limits are tier by plan. Pro users get 5 per day max users get 15 and Team Enterprise users get 25 additional runs available for extra user usage of the same search and usage rate as interactive sessions. So like I see a bunch of my chores getting automated real quickly. Like hey, I got, you know, new dependabot alerts that I need to address. Like I can now have those run on a weekly basis and I don't have to think about them as much, which would be cool. [00:34:13] Speaker C: So I feel like they announced a very similar feature on Cowork. [00:34:19] Speaker D: Very. [00:34:19] Speaker C: I mean there was. I know dispatch is not that, but I thought there was a routine. [00:34:25] Speaker A: There is a loop capability which is more for like inside a cloud code. So. But this is, this doesn't require you to have cloud code running for the loop to work. This is the. [00:34:32] Speaker C: I see. Okay. [00:34:34] Speaker D: One of the skills they have which is do a pr, wait five minutes, see if copilot's response and then give me the feedback to it. And my next, you know, when I get back to work is going to be setting up a routine that just does this for me and all my things. Make sure copilot is triggered, let it go and at least have two AIs review my AI generated code at that point. [00:34:57] Speaker A: It's fun to make them fight. [00:34:59] Speaker D: It's one of my favorite things. That's why I enjoy Claude for my day to day and Copilot for my my PR review. [00:35:10] Speaker A: I have been using more codecs. I can't get onto the Copilot thing. I keep trying and I just, I don't love the interface of a visual studio and or through, you know, GitHub copilot interface. But I do use Cowork for that. Sometimes I'll pull up Cowork and be like hey, time for you to go judge Claude. But I've also used Gemini CLI for that too. But Gemini CLI is garbage. So I'm hoping next week we get a better version of that because it needs some help. But we'll see. [00:35:37] Speaker C: I still don't know how you use Claude code in VS code and complain about the interface though, because they I don't use it. [00:35:44] Speaker A: I. I rarely use VS code anymore. To be honest with you. It's mostly cli. [00:35:49] Speaker C: So you're just doing it in the terminal. That makes sense. [00:35:52] Speaker D: I do it in a terminal inside of VS code. But now that I think about it, I'll actually use much of VS code at that point. More like here's a link to the [00:35:59] Speaker A: folder path if I'M doing stuff where I'm actually reviewing a code heavily. Like the stuff I do for Bolt. I don't typically look at that closely because it's Bolt, but like stuff that I care about, like for security reasons and things that I need to make sure it's doing things I typically do at least have Visual Studio linked to my Claude code session in the terminal. That way, as it's doing edits I want to review, it's pulling them up in. In a sorry Visual Studio at the same time so I can look at it and show me where the changes are it's making. I'm like, I don't like that. Or I give that real time feedback. That's more pair programming though, I would say than other times. If you're doing more gentic things where you're just like, here's what I want you to go do. Go do a bunch of things or my test suite. If it works, we're good to go. If not, let's talk about it. [00:36:38] Speaker C: Yeah, I still don't get good results when I do that. So I'm mostly just doing the pair programming thing and having discussions there. [00:36:46] Speaker A: So my, my trick to that is try using voice instead of typing. Because you probably. And I've actually seen your typing, so I know this. You have a tendency to probably when you talk you use a lot more words and you use a lot more descriptive language in it versus when you write on the prompt. You typically are editing yourself as you're writing. And so the raw thought coming out of like using the voice thing sometimes will help your prompts be more effective. Clot. So if I have like I do something I'm really brainstorming or thinking about as I'm talking it out, I'll typically use voice for that exact reason. And then I'm having more of a con. I'm just spewing raw thoughts to it. And then it can take that raw craziness in my brain and turn that into something really interesting. So something to try is the voice trick. [00:37:26] Speaker C: Maybe we'll try that. Because it is sort of. It is one of those things where I. Every time I feel like I try it and then it doesn't work out and I. I know it's because I'm not providing the right amount of detail. And if you leave an ambiguous space for assumptions that the AI is going to fill it. [00:37:45] Speaker A: Human nature is very often like, hey, you're used to typing something in teams to someone else or to a human, so you're trying to break it down to the basic parts of it. But if you're just talking to something, all that extra context becomes helpful when you're doing planning. OpenAI is launching GPT 5.4 Cyber, a fine tuned variant of GPT 5.4 specifically designed for cybersecurity work. So weird. A week after Mythos, with reduced refusal boundaries for legitimate defensive tasks and new binary reverse engineering capabilities that let security professionals analyze compiled software without source code access, the Trusted Access for Cyber program is expanding from a limited pilot to thousands of individuals verified or individual vendor verified defenders and hundreds of teams with tiered access levels based on Identity verification through chatgpt.com cyber for individuals and a separate enterprise request process for an organization. Codec Security, which has been in preview, has proven to fixing over 3,000 critical and high severity vulnerabilities across the ecosystem. And OpenAI is positioning it as a shift from periodic security audits to continuous automated vulnerability detection integrated into your workflow. And the rate that things are breaking and coming out from all of the different NPM and everything else, you should definitely have continuous vulnerability detection. [00:38:57] Speaker D: Yep. Yes, please. [00:39:00] Speaker C: Yeah, there's, I mean, there's too much and I, I feel like we're just at the beginning of that iceberg for supply chain attacks. I think it's still a lot, there's a lot to be done there to secure those and I think there's too many avenues and I do like, you know, the idea of having this continuous and, and directly into, you know, the developer workflows, like we've always said, you know, adding to the SLDC for static code analysis and all these things. So this is a great, great addition to that. It's amazing to me that it can sort of work on compiled binaries because that's. How cool is that? And then also how scary is that? [00:39:37] Speaker D: It's a little terrifying, I'm not gonna lie. [00:39:39] Speaker A: A little terrifying. [00:39:40] Speaker C: Yeah, yeah. [00:39:41] Speaker D: No, I mean a security review is part of anything. And I guess I never thought about using. I definitely use different sessions to do security reviews. You know, I have different Personas set up to do it, but I've never thought about using different models, whether it's a different vendor or the same vendor to do that review. I get, I guess at some level I do because I use Copilot to analyze CLAUDE code when I do agent coding, but I've never thought about, you know, specifying it down to a very specific one to do it. [00:40:14] Speaker C: Yeah, I haven't played around with security specific models at all, which seems like a gap that I really need to fill because that's, you know, like, I do analysis using the, the, the standard models and I definitely do the same thing with comparison, but I like the, the idea that they're. This is the second, you know, major announcement in the space, and that's kind of cool. [00:40:32] Speaker D: But that's because you were on PTO for a week, so. Ryan, stop going on pto. [00:40:35] Speaker C: That's true. These things probably didn't exist two weeks ago. [00:40:39] Speaker D: I don't think they. [00:40:40] Speaker C: I mean, that's how fast everything's moving, you know. [00:40:42] Speaker A: Right. [00:40:43] Speaker D: And there'll be a new one next week. Right. [00:40:47] Speaker A: I was just looking at this codec security. I just found out there's a Codecs desktop client for Mac, but I was like, wait, what? [00:40:54] Speaker C: I didn't know that. [00:40:54] Speaker A: Good to know. I didn't know that existed. And so. Yeah, okay, cool. I have a lot to play with that report back, but, you know. Yeah, everything is moving so fast. It's just crazy. Even the Gemini CLI has had major updates since last time I looked at it, which I should probably update myself because I don't want to speak here on the podcast. For our listeners, Anthropic is releasing a redesigned cloud code desktop app built specifically for managing parallel Agentic coding sessions with a new sidebar that lets developers run simultaneous tasks across multiple repos and filter sessions by status, project or environment. The app introduces a drag and drop layout system where developers can arrange the terminal diff viewer, file editor and chat and custom grid configurations, reducing the need to switch between external tools during code review and chipping. And this is. I very rarely used the new cloud code ui. You know, I know it exists inside the cloud app again because I just used the cli, but again, we're giving you more features just like the Codex tool that I just saw as well. So there's a new side chat feature lets builders ask questions been tasked about polluting the main session context which they also I now have in the. In the CLI as well. You just type in by the slash by the way and you can ask a question, which is great. And the reason has 3D view modes, verbose, normal and summary to control how much detail is shown about Claude's tools calls plus a usage of Cater showing both context window and session consumption at a glance, which that's actually really nice because getting your session consumption in the CLI is kind of a nightmare and if you're in the middle of a thing, you don't know you're about to run out the updated App is available now for Pro Max Team and enterprise plan users. So check that out. [00:42:23] Speaker C: So this is everything I was just complaining about earlier. This is, this is perfect. This is why, you know, not having this level of tool is why I haven't really adopted CLAUDE code for my main workflows because everything that they're announcing here is exactly what I use and GitHub copilot through. [00:42:38] Speaker D: Well, this was announced today, I think. [00:42:40] Speaker A: Yeah, this came out today. [00:42:42] Speaker D: Yeah, yeah, I think it came out like a few hours ago because I saw you drop it in. I saw, I got the notification too. I was like, ooh, this is interesting because I played with it. I played with Claude code cloud like once or twice and even like this week some where like on my phone I was like, hey, add this feature to this, this agent that I'm working on for my own personal life. And you know, I had it review and do the pull request and everything from my phone, which was kind of fun because I was standing in line at a grocery store because I'm a crazy person. But this is another way that this integrates in with all that. So it's interesting to see them really start to maneuver more of these things into the club code and get away from VS code integration. [00:43:27] Speaker A: I mean, I think VS code's kind of a bit of a boat anchor and it's clunky for a lot of these where a lot of times you just need to review simple things. You don't need all the headache of the, of the ide. So yeah, so it's interesting. I'm curious if there's purpose built use cases. I'll play this cloud one too. So I have homework to do now. So report back. [00:43:44] Speaker C: Is it, is it a dedicated app or is it, is it being rolled out into the like. Yeah, okay. [00:43:52] Speaker D: Yeah, it's inside this default Claude. There was like the three tabs before and they completely redid the whole ui. [00:43:58] Speaker A: Yep. [00:43:59] Speaker C: Sounds like I need to update. [00:44:01] Speaker A: Sounds like you do. [00:44:02] Speaker D: I swear I update like four times a day. There's always new updates. [00:44:06] Speaker A: It is kind of crazy. You know, it's. And then even like, you know, you'll start seeing people complain and you're like, oh, they changed something. You go read the release notes and you're like, oh, they change the way they're, they're calculating context or the way they're doing memory or the way they're doing. And like there's a reasonable caching issues that cause people to burn through their, their seeing things a lot faster. So you want to, you know, then I realized I was still burning the tokens really fast and they had fixed it and I was three versions behind because I hadn't closed the session. So yeah, there's. Don't, don't leave your cloud code alone and unattended for multiple days without quit. Without quitting it at least once. Because yeah, it'll. You're missing out on stuff or potentially you're avoiding things. I did avoid some of that bad caching behavior a little bit too. [00:44:48] Speaker D: So my problem is even the app, when you tell it to update, it only updates like one version. So it closes, reopens and it's like one more. I'm like, no, no, no. Update to latest. Like don't. I don't need 17 interrupt steps. [00:45:01] Speaker A: Yeah, I don't need you to walk me through this actually. I just, for that exact reason I just download redownloaded the binary to install it because I was like, I don't think I have the latest version that has all the, all the code code stuff. So I now do, which is great. All right, moving on to aws. Amazon Bedrock Projects lets organizations attribute AI inference costs to specific workloads by passing a project ID and the API call, which then flows into AWS Cost Explorer and AWS Data Exports for analysis. This addresses a real operational gap for teams doing chargebacks or investigating cost specs across their multiple AI applications. I can tell you that this is a must have. So every cloud provider needs to provide this capability. Like this is a major problem in Vertex, it's a major problem in Bedrock, at least until now. And even I'd say the project level is probably not granular enough. I would need it at IAM Identity level. Yeah. So I appreciate the step in the right direction, but I think there's definitely a need for even more of this space. And I think Finops is probably going to have to really drive a lot of better behaviors and AI usage because it's difficult to know what people are doing, what they're charging, how it's getting charged to them, and they're just using it. They don't know better. And so they switched over to Opus Model or they switched over to chat GPT 5.4. Then all of a sudden the bill's going right through the roof and you don't know why. And without being able to give them more real time insights, it's impossible. So thank you Amazon for this. It's appreciated in the right direction and you could potentially, I guess, get Bedrock projects per user, but there's still Lots of opportunities. [00:46:29] Speaker D: They do have cost allocation by user for Bedrock. It was announced. [00:46:33] Speaker A: That's good. [00:46:34] Speaker D: It was announced I think a couple days ago. I think it was just a follow up story to this one. [00:46:39] Speaker A: Okay, good. [00:46:40] Speaker D: Yeah. Bedrock now supports cost allocation by IAM user role. I'll drop the link in the show notes just as a follow up to that one. I was like, I swear I saw that. Yeah. [00:46:49] Speaker A: Cause I hadn't seen that and I looked for it before. So I'm glad to see they've got that too. That's not, that's not old either. So that just came to April 9th. [00:46:56] Speaker D: I mean it'd be nice if they made this like an SCP that you can force because. [00:47:01] Speaker A: Yes, that'd be great. [00:47:03] Speaker D: That's probably will be the follow up, I feel like to this which is enforcing it because otherwise, you know, no one's ever gonna do it. [00:47:09] Speaker C: Well, can't you enforce budgets in AWS now like. [00:47:13] Speaker A: Yes, you can. But again it's the challenge of you're forcing budget at the prod, at the, you know, at the overall service level typically or at the account level, not at the individual level. Yeah. So I mean you can do all kinds of warnings, you do things but the granularity and the need for the stuff is going to become a bigger deal because you know, the truth of the matter is is that it's really going to cost. You know, there's someone, some quote from Nvidia I think was like your $500,000 software engineer is going to need require $250,000 in tokens. [00:47:41] Speaker D: Oh yeah. [00:47:41] Speaker A: And I was burning, yeah, if they're not burning that many tokens, they're not doing your job. [00:47:47] Speaker D: And it's job. And I was like whoa, yeah, I saw that quote. [00:47:51] Speaker A: I don't know if I, I don't know if I buy that. Just to be clear, he's a man who wants to sell you a lot of Nvidia hardware that is cheaper than running these inference models on cloud. But yeah, if that's true, then you're going to need to be able to have this granularity for sure. [00:48:03] Speaker C: Yeah. I mean I know that I definitely burn through tokens quite a bit, you know, and especially at the day job, like you're being pushed for more productivity using AI. They're, they're actively monitoring metrics and, and it's one of those things where you do different things and bigger things and you know, it's, it's really easy to burn, burn through that and if there's Not a quota or not, you know, some sort of limit, you don't really know. But I also hate the enforcement because I'm constantly at a token. So don't, don't do that. [00:48:35] Speaker A: Well, you need, you need that. So this, this is one of the areas where you have recently started using a teams plan for Claude. I'm very disappointed in how little enterprise controls do exist in these tools like the ability to have per, per person budget approval flows or per organization published approval flows for overages tracking that data. What are they, you know again like what models are they using? Like it's very, it's not as granular as you would expect it to be. And like I was telling someone earlier today, I was like, they really need a good RBAC model. Yeah, you know, least privilege least, you know, because like someone wanted access to usage data. Like, well to have usage data you have to be an owner of the org. I'm like that doesn't make any sense. Why would you have to be an owner? That should be something an admin can get. Or I should be able to create a custom rule that says hey, I'm an engineering manager and I want to be able to see for my org these, you know, their usage. Yeah, that would make sense. [00:49:23] Speaker C: So yeah, they're really not enterprise ready. Like no, nothing is right. I haven't played around with the OpenAI [00:49:29] Speaker A: version but yeah, well like, and like, you know, if you look at like Claude like they have a lot of integrations into work OS as their back office to do that kind of stuff to set up SSO and to track usage. And so part of his limitation that they've, they've adopted a third party who isn't purpose built for what they're trying to solve and so they're kind of beholden to what the platform offers to them and that platform is probably trying to you know, move as quickly as possible too to meet their demands. So it's a, it's a double edged sword all the way down. Amazon Bedrock Agent Core runtime now supports stateful MCP servers enabling bi directional communication between MCP servers and clients. The key change is a single flag statelesshpalse which provisions a dedicated micro VM per user session lasting up to 8 hours. Three new client capabilities are now available. Elicitation for pausing tool execution to collect user input. Mid workflow sampling for delegating LLM generation back to the client without the server need to own some model credentials and progress notifications for streaming real time status updates during long running operations. The Sampling capability is particularly notable for enterprises use cases because it allows MTP servers to leverage the clients connected LLM without holding API keys or model credentials directly keeping model access control on the client side. Each stateful session gets CPU memory and file system isolation via the micro VM session tracking through an MCP session ID header. Sessions expire after 15 minutes of activity or maximum of 8 hours after which client side must reinitialize the connection. This could be dangerous. So definitely this one. If you're going to implement stateful mcps, I would make sure you have a very good security model for them. [00:51:02] Speaker C: Yeah, it's. I mean it's such a weird, such a weird interaction model. Like it's like on one hand I really do like this because it's such a huge limitation, you know, being able. Not being able to sort of have a. Have something that's executing against MCP without it just being it stood up every time. But on the other hand it's like it's really easy to get into and what an attack vector. [00:51:25] Speaker D: I'm thinking about how taking it the other direction, think about the infrastructure that's built by AWS over the last 10 years with the nitro cards and everything else and kind of think of like lambdas and whatnot. And I feel like this is a way for them to essentially take a lot of what they built and leverage it in other ways now so they're able to leverage the nitro cars to spin up a quick VM or you know, or quick container and you know, or whatever it is behind the scenes in this case and really be able to build it and put it out there very quickly for a person and tear it down or not tear it down in this case it's without a lot of the work that's been happening over the last 10 years at that infrastructure level of the nitro cards, the quickly loading of VMs, getting rid of all that cold start issues that were existed and everything. I don't think we would be where we were today. [00:52:21] Speaker A: No, it wouldn't be possible. I mean this is definitely built on top of firework. There's no doubt about it in my mind that one. [00:52:25] Speaker D: Sorry. That's why I couldn't remember why I kept saying nitro cards. [00:52:28] Speaker A: But nitro is a key part of fireworks, so it makes sense. But yeah, micro VM was the tell for me on that. I was like, oh, they're using micro VMs, we're using firework to do this. You know, they're giving you the managed service which makes Sense very similar to Lambda. And so they're giving you that capability. I can see the value of a stateful NCP for certain, you know, things that, you know, like non security sensitive content, like knowledge bases and different things where you want to be able to make that available and something, you know, you don't want to always have to go rebuild the cache of every time or something like that. So I think there's some use cases that do make sense to me for this. I just, again I, I worry about anytime you introduce state into something like this, that didn't have state previously, that people all of a sudden start doing things that they shouldn't really do just because they can. It's like, you know, S3 fuse. Like yes, just because you can access S3 via an API, it doesn't mean you should. And so that's, you know, one of those, you know, the devil's in the details. [00:53:23] Speaker C: Yeah. Especially since you don't really have a lot of visibility into what is in that state. [00:53:28] Speaker A: Right. Oh, and then you know, what's the mcp? Who owned it? Where did it come from? Does it have malicious code in it? You know, does it have access to APIs? You don't expect like there's so many risks. So again, it's careful. Great power comes great responsibility. AWS Agent Registry is part of Amazon Bedrock Agent Core, now in preview as a centralized catalog for discovering and governing AI agents, tools, MCP servers and custom resources within an org, helping teams avoid rebuilding capabilities that already exist. The registry supports URL based discovery that automatically pulls metadata like tool schemas from live agent endpoints, plus an approval workflow so admins can gate what becomes discoverable. With CloudTrail providing full audit trails for compliance, developers can search, register using natural language, semantic search or keyword search and can access it via the console CLI SDK or via their IDE. Preview is available in 5 regions, most of the big ones, and for individuals running multiple agents projects across teams. This addresses a practical governance gap app. So as long as these aren't public agents that I can now just go access and run like NPM agents or NPM code, I'm okay with this. It's for my internal Org. I think it's great. [00:54:34] Speaker C: Yeah, it's, I mean it's funny because I, I don't really think about, you know, Bedrock Agent Core for, for enterprise, but maybe it would allow that maybe in a sideways kind of way. [00:54:45] Speaker A: That's where I think that's where I see the most. I mean the problem is you and I come from a SaaS background, so we always think everything from the SaaS context but from an enterprise context. Right, like your customer support people and being able to have agents that handle incoming tickets and doing triage of those. And then you know, being able to create tickets automatically to engineering or defects, respond back to customers for refund requests and shipping notifications. Like all these things that you have people doing toil work on that these agents can do. That's where I see these things coming into play. But I do have to switch my context from SaaS to enterprise. And what are they trying to do with these tools? [00:55:17] Speaker C: Yeah, I mean it is sort of like, like from a SaaS perspective, like I, I see this going the way of microservices. That's exactly what I was thinking. [00:55:25] Speaker A: Yeah. [00:55:25] Speaker C: And then having reusable components sort of govern centrally for constructing those. Those services. Yeah. [00:55:33] Speaker A: Yep. It's all, all new world out there. Kiro CLI is introducing headless mode allowing developers run the agentic terminal programmatically via API key and environmental variables. Enabling integrations into CI CD pipelines and build scripts without user interaction. Native Windows support removes the need for workarounds like Windows subsystem for Linux getting developers using Kiro agents directly in Windows terminal for tasks like code based navigation, bug tracing and workflow automation. The updated TUI is now generally available after an experimental experimental period. Adding a subagent monitoring view acceptable via control G real time task list and parallel sub agent execution that protects parent agent context on complex tasks. I don't use CURO enough to know if any of this is cool, but it sounds like it's moving more towards what Claude and other tools are doing and they're all racing to meet what Kiro's doing. So depending on what way you like to do your automated junta coding, everyone will have a solution for you. Either a CLI or a new custom GUI or integration into an ide. So sounds like you get all three. [00:56:32] Speaker C: When's the last time you heard Windows users getting any love? Like that's native Windows support. [00:56:37] Speaker A: Like that doesn't happen even in cloud code. Like it's still very heavily requires wsl. [00:56:43] Speaker D: Yeah, most of these require WSL under the hood because even I feel like Microsoft is leaning that way with a lot of random things too. [00:56:52] Speaker C: Yeah, I don't know if they're leaning that way. They, they put it out there because they had to. [00:56:56] Speaker D: Yeah, people are forcibly most. A lot of tools that we use. If you're on a Windows box require. Let's Put it that way. [00:57:04] Speaker A: I mean they could just, you know, release Windows 12 with, you know, a [00:57:07] Speaker D: Linux kernel because think over multiple people's dead body. [00:57:12] Speaker A: I'm sure, I'm sure. I mean that part that Windows always was most viable is their superior ui. So you know, underlining, pinning parts of Windows are terrible. So yeah, the next kernel plus the rate Windows UI could be really nice. Just saying it'll never do it. And our final Amazon story. Amazon is acquiring Global Star in a deal expected to close next year, gaining its LEO satellite fleet, MSS spectrum licenses or global authorizations, and direct to device technology to expand the Amazon LEO satellite network beyond its current broadband aspirations. Started in. Sorry. Amazon LEO will deploy a next generation direct to device satellite system enabling voice text data services on standard mobile phones without specialized hardware targeting coverage gaps where terrestrial cellular networks cannot reach Amazon have Apple have signed an agreement for Amazon LEO to power satellite features on iPhone 14 and later and Apple Watch Ultra 3. So this is basically your SOS capabilities in your phone which Global Star has provided will still continue to operate once Amazon closes the steel. The combined network is designed to support hundreds of millions of endpoints globally with practical applications spending, consumer emergency messaging, enterprise fleet tracking, disaster response, fallback connectivity and rural broadband extensions. And the. I guess we can finally say the conversion from Amazon the bookstore to Amazon the utility are complete. Yeah, [00:58:24] Speaker C: it's kind of crazy. I didn't realize that they had grown this so much that it is, you know, a competitor with Starlink, which is what everyone thinks of when they think about these things. I knew they had, you know, or at least I knew globalstar. [00:58:36] Speaker A: I mean they haven't even shipped leo. They haven't shipped LEO yet. [00:58:39] Speaker D: It has. [00:58:39] Speaker A: I mean, I'm on the wait list because I want to get that for my backup Internet at my house or for when I go camping or things like that. But you know, it hasn't shipped yet. It's supposed to ship later this year, which I'm excited about. But then they've, they're already aspirationally saying, well, we need something big enough to manage all of it and more. And they probably already had a pretty big contract with Global Star. And so it was one of these questions of like we could just buy it and now have now be a carrier for cell phone satellite connectivity. It's interesting play. [00:59:05] Speaker C: I thought you could get some connectivity through Global Star before this acquisition. [00:59:10] Speaker A: I think you, you could, you could. [00:59:12] Speaker D: They did cell phone. They were more geo. Geosynchronous satellites versus LEO and Stargate are all Low orbit or like. [00:59:20] Speaker C: Or the, the SMS type messaging, right? [00:59:23] Speaker D: Like, yeah, like you're lost in the woods. [00:59:26] Speaker C: You have like use it for Dolore Bay. [00:59:28] Speaker D: Yeah, yeah, I know of one of my old employees had one of those devices because he would go hunting in the backwoods of wherever he lived and he would. Had one of those, like, it would just pretty much send a pink to his wife saying SOS with the GPS coordinates. [00:59:44] Speaker C: I have one of those. [00:59:45] Speaker A: It's a $11.57 billion acquisition. Whoa. [00:59:50] Speaker D: Yeah, it was like 30 cents per share or something I saw this morning. [00:59:56] Speaker A: Yep. [00:59:57] Speaker D: Because all. All this was like today. [00:59:58] Speaker A: I think that's a pretty. Yeah, it's in the last day or two, a pretty big, you know, attack on Starlink too. Like to, you know, hey, we're going to. Not only are we going to announce leo, which will compete directly with Starlink, but now we're going to get fully vertically integrated. It's interesting play. And then if you can power cell phones as well in the future, which I know Starlink would like to do as well. And it all gets real interesting. [01:00:20] Speaker D: But they already had the ground station and a few other things that they kind of. [01:00:25] Speaker A: The ground station was a partnership with Global Star anyways. [01:00:29] Speaker D: Oh, was. I realized that. [01:00:30] Speaker A: Oh yeah. Because Amazon just built a control layer on top of Global Star's stuff. That way they could basically provide ground station. Because the reality is Amazon wasn't building their own infrastructure for that. It's cost them bajillions of dollars to build the satellite infrastructure. Plus they had. What satellites are they talking to? They're talking to Global Star satellites or other. Other competitors of Global Star. So it does. It doesn't make sense for them to do that, in my opinion. So it's always a partnership. [01:00:55] Speaker D: Apparently Apple owned 20% of Global Star, according to Wikipedia. [01:00:59] Speaker A: Yeah, that was an investment they did when they basically invested in the iPhone SEO SOS feature. Because basically Amazon or Apple's giving away for free on every iPhone. And so but there's still a cost to deliver the service. And so basically Apple, by owning part of the company, subsidizing that piece of it. [01:01:15] Speaker C: So not only, you know, so Apple gets a little richer with this acquisition, but iPhones will continue to get more expensive because they will continue to make this a free service. [01:01:23] Speaker A: Well, the interesting thing will be is, you know, once they release in 2028, a satellite cell phone service, you know, will. Will Apple devices, you know, people start moving that direction instead, you know, versus Verizon or T Mobile, et cetera. So that'll be be interesting to see. [01:01:40] Speaker C: Yeah, that's a lot of throughput. [01:01:43] Speaker A: Yeah. We don't have a lot of GCP news this week because they're, you know, preparing to shock and amaze us all next week, but a couple of things came up. GKE Cloud Storage Fuse Profiles are now GA in GKE 1351. This automates storage configuration for AI ML workloads by replacing manual tuning with three pre built storage classes. GCS Fused CSI training, GCS Fuse CSI serving and GCS Fuse CSI checkpointing feature addresses a real operational pain point where customers were leaving performance on the table or experiencing pod out of memory kills due to misconfigured cloud storage settings that previously required navigating dozens of pages of documentation, which if you had ever done Google documentation, you know, is a terrible day in the life. The system dynamically scans your bucket, analyzes node resources including ram, SSD and accelerator type to calculate optimal cache size at deployment time move the need to manually account for these variables across different infrastructure configurations. [01:02:35] Speaker C: That's nice because I think previously you really had to rely on some team managing this centrally if you're going to get that level of customization and tuning. Now it seems like this might be something that individual dev teams would have, you know, a better experience just instead of just, you know, whatever they get out of the box. [01:02:55] Speaker A: Well, if you've been using Gemini to help your kids with their science homework and you really wanted 3D models, Gemini has that for you this week. The ability to create interactive 3D models and charts directly in chatmini.google.com Moving beyond static text and diagrams to functional simulations users can manipulate in real time. This is available by selecting the Pro model and prompting Gemini to show me or help me visualize a concept. The feature supports adjustable parameters like sliders and numeric input, so users can modify variables such as gravity or velocity and immediately see updated results. The role is global for standard Gemini app users. Their education workspace accounts are currently excluded. No additional cost is mentioned beyond existing Gemini Pro access. So if you don't have a Pro account, you'll need to get one to use these features. [01:03:38] Speaker C: This is something that makes me think about actively getting a Gemini Pro account, which I don't have today. Just the amount of stuff that I do with 3D printing and, and that being able to generate a model that I can then import into a tool like, you know, fusion and tweak it, or maybe just would generate G code directly. So this is, I like this and it's definitely something I can see myself [01:04:03] Speaker D: using, I mean, I can see it used not just for that, but for, you know, people doing housing construction. Anything along those lines. You know, my wife and I were talking at one point about redoing the kitchen and trying to visualize it and I can't visualize, you know, if we remove this wall, what's the space going to look like? So now, you know, if you can more before you could do like a 3D, you know, picture, but now if you can actually do a 3D model and kind of look at it and walk through it, to me, I think that would help. So I foresee this not just changing, you know, what you're talking about, but if this grows into just general things when you need a model of things. You know, the gravity effect can be interesting if you start talking about gaming because I remember one of my friends told me, she's an architect and she said they added gravity to a really old game from like, I don't remember, choose a year like you know, v1 to v2 of the game. The first one didn't have gravity. But the problem is none of the buildings they built in there were structurally sound. All of a sudden they added gravity and all the buildings collapsed on, you know, as soon as they added that feature. So if you can add those types of things, you can really start to see a larger and new design aspect of things. [01:05:15] Speaker A: Yeah, that's cool. Yeah, so I, I have pro and I just did it and I created the, you know, a multi layer version of the sun and it shows me the different zones and I can change different aspects of it. It's. It's pretty cool. [01:05:27] Speaker C: It is cool. [01:05:29] Speaker A: Definitely fun to play with. And yeah, I didn't think about the 3D printing aspect though. That's, you know, my son has a 3D printer. He'll see if we can create some 3D models of some stuff. That'd be cool. [01:05:39] Speaker C: Yeah. [01:05:40] Speaker A: Google Cloud is automatically enabling an enhanced security command center standard tier for eligible customers at no cost. Adding AI protection features including a unified dashboard that detects unprotected Gemini inference and reports on LLM guardrail violations, with general availability expected by end of 2025 or 2026 for June. The free standard tier now includes more than 44 misconfiguration checks based on the Google Cloud Security Essentials compliance framework, up from the previous count of 21 checks, along with agentless critical vulnerability scanning and graph driven risk prioritization data. Security posture management has been added to the free tier, allowing teams to discover and visualize Data across vertex, BigQuery and Cloud Storage with Compliance Manager included for automated monitoring against the GCSE framework. SEC now services in context security findings directly inside Cloud Hub, GCE and GKE dashboards, giving infrastructure admin security insights without switching between tools, which is always great. Bring me the data, don't make me go to the data. [01:06:32] Speaker C: Yeah, no, this is definitely. I really like this and especially the free tier aspect of this just because it is already such a challenge to know where your AI workloads are and then having the specific configuration checks is great. I do think that the checks themselves, I mean I don't. I've played around with the 21, they were a little basic so it wasn't, you know, that great. But you know, I do think it's a great thing to have. And specifically the data scanning is super key because that's typically been really expensive to run and classify your data and know where your your sensitive data is. [01:07:08] Speaker A: So I mean what you found cloud DLP expensive? I don't. [01:07:12] Speaker C: Just a touch really. [01:07:13] Speaker D: Why would you say that? [01:07:14] Speaker A: So weird. Yeah, I haven't found DLP solutions didn't cost bajillions of dollars. Makes sense. Well, if you want to visualize all those data and findings, you would typically use Looker. And if you're looking for Looker Studio, you're going to have to look elsewhere because now it's now named Data Studio, which will not be I remember ever again after this podcast. Data Studio is positioning itself as a hub for personal data exploration and ad hoc reporting across Google Data sources including BigQuery, Google Sheets and Google Ads. Platform now serves as a single location for multiple asset types beyond traditional reports including BigQuery conversational agents and data apps built in Collab notebooks. Looking the broader shift towards AI era analytical workflows, Data Studio will coexist with Looker rather than replace it with Looker remaining. The enterprise BI platform focused on governed data and semantic modeling, while Data Studio targets individual and small team use cases. Cases pricing follows a two tier model. The standard Data Studio remains free for individual use while Data Studio Pro adds AI features, enterprise security and compliance capabilities at a paid tier purchasable through the Google Cloud Console, which that was One of the big problems with Looker Studio was that it wasn't really meant for enterprise. So this Data Studio Pro version gives you that capability finally. [01:08:22] Speaker C: So it used to be called Data Studio, so this is renaming it back of course. [01:08:27] Speaker A: Yeah, full circle. [01:08:29] Speaker C: That's funny. And I wonder like there's. Because there's always been this weird sort of like difference between looker and looker Studio. And I wonder if this is like if they're gonna still have the sort of looker, which is your sort of your big bi engine, versus the studio, which is the visualization layer. I wonder if they're gonna keep that as two separate things or what they're doing. [01:08:50] Speaker A: Yes, I do think that is what they're doing. At least what they say. Yeah, well, coexist with looker rather than replace it, so. [01:08:57] Speaker C: Oh, okay. Because that's, you know, it's funny, I'm starting to feel like, oh, Google, Google Cloud Pro. And it's kind of embarrassing. This is the, the service that I remember the history [01:09:10] Speaker D: shows how many years you've been on the cloud. Yeah. [01:09:13] Speaker A: Shows how many years he's been trying to make looker work for him. [01:09:15] Speaker C: It, you know, every time we think it's easier than quicksight, so we could all. [01:09:21] Speaker A: That's not hard to do. Hard to do. [01:09:23] Speaker C: Exactly. [01:09:25] Speaker A: BigQuery graph is now in preview, Bringing native Graph analytics into BigQuery using the ISO GQL standard. This lets analysts run multi hop relationship queries without leaving BigQuery or learning a separate graph database system. The key technical distinction is that graph schemas are created on top of existing relational tables with no data duplication or movement. And users can mix SQL and GQL in the same query, which lowers the barrier of teams already invested in in SQL skills. Integration with Spanner Graph is a notable addition, allowing federated queries that combine real time spanner Data with historical BigQuery data in a single virtual graph. This addresses a common pain point where operational analytical graph data live in separate systems. I don't get it. The challenge here, I had a very simple mental model of graph data and now graph data is SQL data. And so that's confusing, but you can just put it on top of SQL data. Which is what I thought the whole point of why we couldn't just use SQL to use graph data. And like I said, my brain hurts. [01:10:24] Speaker C: Yeah, I don't get this at all. Like what query could I possibly write where I would need access to both? I mean maybe they want to create like a data set that has both types and like separate tables that you can then query independently. I don't, I have no idea how one would use this. And BigQuery can BigQuery like, you know, for like graph like queries that are sort of like not one off but like kind of smaller and quick succession like BigQuery doesn't really like that. It's. It's very limited in terms of like concurrent queries and things you can do. So this is strange. [01:11:05] Speaker A: Interesting. Well, I am going to try to not think about that much more because it's going to hurt my brain. [01:11:12] Speaker D: Yeah. [01:11:13] Speaker A: In a feature that I'm sure is going to make your screen people cringe just slightly. Google's launching skills in Chrome, a feature that lets users save custom Gemini prompts and rerun them with a single click using the forward slash or plus button interface, eliminating the need to retype repeated prompts across browsing sessions. Skills can operate across multiple tabs simultaneously, which makes it practical for tasks like comparing product specs or scanning several documents at once without manual prompt response repetition. Google is also shipping a pre built skills library from common workflows like ingredient breakdowns, gift selection and macro calculations with options to customize any library skill by editing the underlying prompt. On the privacy and security side skills inherent Chrome existing Gemini safeguards including automated red teaming and confirmation prompts for sensitive actions like sending email or calendar events can occur. Save skills sync across signed in Chrome desktop devices, making this more of a persistent personal workflow tool than a one off browser feature. But it is limited to desktop and there's no mention of separating pricing beyond existing Gemini and Chrome access. What could go wrong? I mean the red flag of course is the word macro. Well I'm also going with the Every macro has always been security driven. [01:12:17] Speaker D: Yeah but the synchronization. How many people do you know that are signed into your personal email on your work computer and on your home computer and now. Oh let me go click this button because I'm automatically want to be able to see every, you know, this trip I'm flying for going somewhere and great. Now that synchronizes to your work computer which does God knows what under the hood in reality. That sounds terrifying. Ryan should be having a panic attack right now. Where's your brown paper bag? [01:12:47] Speaker C: The syncing is a little separate than having multiple profiles. But. [01:12:51] Speaker D: But what average consumer, what average employee knows to have multiple profiles? [01:12:57] Speaker C: I think a lot of people have multiple profiles. [01:12:58] Speaker A: I have multiple. [01:12:59] Speaker C: I think a lot of people don't use the syncing because the syncing is like all your bookmarks and all that kind of stuff and so I wonder. But I mean having that sync to your work computer is already a different problem. Yeah, I don't know. I. I'm trying to figure out whether I like this or not. Right? Like because I can think of some things that are are kind of cool. And I'm trying to get around the, you know, the silliness of like, just executing things without really, without really knowing what's going on. That's usually how security problems are. [01:13:27] Speaker A: What's wrong with a skill that you've somehow saved on your personal computer at home that was a virus that then gets synchronized to your work account, your work Chrome that's logged into your work space and then sends all of your personal email from your work account to this hacker's account? What could go wrong with the skill that does that? I just don't know what you're doing. What you're not concerned about. [01:13:50] Speaker C: Well, yeah, I mean, it's. The sinking is where, like, that's, that's already an issue. Right. Like, that's what that means is you've signed into your browser at work with your personal profile and have enabled syncing for all your personal bookmarks and all those things. And so, like, it's already going to be a concern there because you can do the same thing with cloud code or cowork. [01:14:10] Speaker A: Yeah, I just. This is in your browser, which is just such a, you know, JavaScript. Like here, just open this website that has this bad JavaScript on it that affects your computer. Like, I just. This one makes me leery. [01:14:23] Speaker C: Yeah, you're right. The, the, the, the access at the DOM layer is, is terrifying because that's. You get into weird hidden prompts and websites. [01:14:30] Speaker A: I mean, then it says, you know, it says it's going to have prompts for security, you know, for doing safe, you know, secure things. And so it's going to prompt you. So then people are just going to prompt fatigue. There's like UAC where they just, you know, approve, approve, approve. And no one's actually thinking, like, should I be approving this or not. [01:14:44] Speaker C: Right. [01:14:44] Speaker A: So that's definitely a challenge. Yeah. So I, I like the idea of it. I like the idea of not having to repeat my skills. I just, I have questions. And if this ends up with, you know, work basically saying, look, you can no longer log your personal Google Chrome profile into your browser, that's gonna make a lot of people unhappy. So as a, as Matt said, I [01:15:07] Speaker C: mean, I think that, you know, I think corporate ownership of browsers is, Is going to happen. [01:15:13] Speaker A: I mean, it's happening already and just [01:15:14] Speaker C: be a stand and just. It's. I mean, it's already happening, but it's going to be, you know, table stakes for, for any organization. [01:15:20] Speaker A: I mean, the secure, Secure browsers. [01:15:22] Speaker C: Yeah. [01:15:22] Speaker A: You know, everywhere. [01:15:23] Speaker C: Too many weird extensions, too many weird things you can do with the browser. And then also just the universe, how close it is to the user interface. [01:15:31] Speaker D: Definitely coming. [01:15:33] Speaker A: There's a bunch of startups doing this. [01:15:36] Speaker D: Yeah. And it's for security. And you know, before that you just do a man in the mail attack. With MTLS being so prevalent nowadays, less and less things you could do a man in the mail attack for. So, you know, having a secure browser that they can. Your security team can watch what you're doing and see that you're not exfiltrating data, etc. Etc. It's just. It's where the world's going either way, whether we like it or not. [01:16:00] Speaker A: Yeah, it's the reality. All right, let's move on to our one and only Azure story plus Azure Microsoft released Agent Framework 1.0, merging semantic kernel and autogen into a single SDK. After maintaining them as incompatible parallel frameworks, Autogen will now receive only bug fixes and security patches, meaning developers on either framework face meaningful migration work to adopt the new unified tool. The Azure Agent stack still spans multiple distinct services, including Agent Framework for Pro Code Development, Copilot Studio for Low Code, Foundry, Agent Service as the managed runtime, and the Microsoft 365 agent's SDK for teams distribution, with each service having its own documentation deployment model requiring enterprise teams to make platform decisions before writing any agent. Logic Agent365, a government and compliance control plan for monitoring agents at enterprise scale, reaches general availability on May 1 at $15 per user per month, adding another procurement decision on top of your existing build and runtime layers rather than consolidating them. By comparison, Google Cloud and others AWS strands Agent SDK has a thin framework that pairs clean with agent cores that's managed runtime. Both competitors offer a more direct path for local development to production without requiring lateral platform decisions. This article was. Sorry, I should have clarified. This is an article bashing Microsoft from Forbes, where basically Microsoft's agent stacks are confusing developers with rival or while rivals are trying to make it easier. So basically the general feedback is that Azure has made this way too complicated, where everyone else is trying to simplify the code development process to get from development to agent production in a streamlined, simplified method. [01:17:31] Speaker D: So Microsoft make things harder and more confusing? Never. [01:17:35] Speaker C: And then make it worse by trying to simplify. Yeah, no, that sounds like Microsoft like parallel frameworks that are incompatible. Now it's just going to be incompatible [01:17:43] Speaker D: with itself and then in a couple years it'll just be like Google where they just renamed it the original name of things. They'll split back into two, it'll get too big and they'll fragment it out and split it into two and then [01:17:56] Speaker A: they'll make an ultra version and ultra premium version. Ultra premium version. [01:17:59] Speaker C: Yeah. [01:18:00] Speaker A: So, yeah, makes perfect sense. All right, gentlemen, I think that's it for another fantastic week here in the cloud. [01:18:06] Speaker D: We made it. [01:18:07] Speaker A: Woohoo. [01:18:08] Speaker D: And predictions. Plug this. [01:18:10] Speaker A: And predictions. [01:18:10] Speaker C: Yeah, great. Well, bye everybody. [01:18:13] Speaker D: See ya. [01:18:15] Speaker C: 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 on Oracle, who knows? Check out the CloudPod.net for our newsletter. Join our Slack Message us on socials or leave a review. That was awesome. [01:18:48] Speaker D: That was fabulous. [01:18:50] Speaker A: Yeah, you're welcome. That's our. That was some work I've been working on in the background. [01:18:53] Speaker C: So I love how we get like showed all this like live. [01:18:58] Speaker A: Yeah, I mean I did share it with Jonathan in the Slack like, like two weeks ago and you guys weren't around to. Yeah, Jonathan liked it too, so. But yeah, very cool. We don't always air it at the end of our episode, but we do have an after show, so it makes perfect sense today. I mean the world has watched hoping that there wasn't anything that went bad with Artemis 2. And you know, I'm glad to report that the astronauts are safely back on ground. So now we can talk about all the things. Yeah. Without jinxing the mission. [01:19:26] Speaker C: Yeah. [01:19:27] Speaker A: Which is great. And so the first one that came up was basically an article. Where did this come from? This was from communications of the acm. Basically how NASA built Artemis is two fault tolerant computer, which I found to be a fascinating article. Basically the Orion capsule run ran eight CPUs in parallel across four flight control modules using a fail silent design where faulty processors drop out rather than transmit bad data. The System can lose three of four modules within 22 seconds and still operate safely on the remaining module. The architecture enforced strict determinism through TRIME triggered Ethernet and RNC653 scheduler, ensuring all processors see identical inputs and produce identical outputs, which is a notable contrast to modern agile and DevOps practices, where this level of architecture discipline is increasingly uncommon. NASA uses dissimilar redundancy for the backup system, meaning different hardware, a different operating system, and independently written simplified software. Specifically to prevent a common software bug from taking down both primary and backup systems simultaneously. The verification process Relies on supercomputer scale, fault injection and Monte Carlo stress testing to simulate full mission timelines. With catastrophic harbor failures introduced, which offers a practical model for how cloud initial teams might approach resilience testing at scale. And the broader industry implications that software takes over functions previously handled by mechanical or manual controls, Whether in spacecraft, autonomous vehicles or industrial systems. The engineering patterns developed here around file salient design and layered redundancy become increasingly relevant relevant outside of aerospace. And I, you know, we talked a couple weeks ago about Outlook and the astronauts having the Outlook problems. In that case, Outlook is actually their personal side. The mission control computers are completely separate from those. And so this is just fascinating, you know, architecturally, what it's doing, how it's failing, how each processor is getting its own identical set of instructions and outputs. And if the outputs don't match, they basically kill the, kill the chip until they can debug it. And then even the fact that they're writing software for two different, completely separate systems. Systems, that's wild. Can't break it. How you design a software SDLC for this has got to be absolutely fascinating. [01:21:30] Speaker D: Well, it's not just software, it's the hardware SDLC too of you have. I think maybe I'm wrong, but the way I read it, it was also two different processors, which means it's the entire stack up of infrastructure all the way through of physical hardware, all the way through the software. So you're having multiple vendors involved and, and people wonder why this stuff's so expensive because they're. Was it 250 million miles away from 250,000, whatever the number was, you know, away from home on literally a slingshot around the moon to get back. I get why they want that level of fault tolerance because the last thing NASA and the world wants to see is dead astronauts orbiting the earth in a exponential decay. Like that's not going to go well either. [01:22:15] Speaker A: Yeah, I mean that's, yeah. I mean if you look at the technology from like the space shuttle to what we're talking about now in Artemis and Artemis is designed to go to Mars long term. So it's got to have even more fault redundancy because it's going to be, you know, subject to a lot more radiation, a lot more, you know, space effects that we typically don't have in our system. So having the system of eight computers basically handling all of this makes a lot of sense. And I feel a lot better with my self driving car if it had at least two computers, just two doing this. Because you know, there's there's been horror stories about, you know, EVs and things, you know, wrecking into things due to bad sensors or bad input data or, or all kinds of things. They're just normal hardware failures that we tolerate on Earth but you can't tolerate at space level. So yeah, it just overall fascinating and you know, like it'd be fun to be a software engineer on a project like this because it's a whole different. It's the same thing like building heart monitors or building health devices. Like you have a higher level of SLA because if you that up, you kill somebody. In this case, you kill, you know, three astronauts or four astronauts. Yeah, which would be terrible. And, and you know, instead of a national moment of pride for the America which we desperately needed in this recent political time, I would have been a, you know, a national tragedy like some of the other. And I, you know, I witnessed one shuttle implosion. And so, you know, it's, it's a bad scenario when those things happen. And I was very nervous about how things would be handled if that happened. So I didn. [01:23:40] Speaker C: Yeah, exactly. [01:23:41] Speaker A: But yeah, it just, it's cool. So we'll keep an eye out for this amount of articles. You know, they had a bunch of technology they had took along with them. They had iPhones, they had, you know, their personal computer stuff which had the outlook problem that we all laughed about a couple weeks ago. And definitely we'll keep an eye on these things as they go through them. But yeah, Apollo astronauts had a 1 MHz computer and 4 kilobytes of erasable memory. And this computer, you know, this is 8 CPUs and how knows how much memory in this contrast. So just probably all of it. [01:24:09] Speaker C: Hopefully all of it. [01:24:10] Speaker A: Hopefully. I mean NASA might not be able to afford memory right now. So hopefully they bought it for the price spike. [01:24:16] Speaker D: Well, that's the thing with most of the. I remember when was it Opportunity? Was it Curiosity or Opportunity? And what was Mars Rover? [01:24:25] Speaker C: Yeah, Curiosity. [01:24:26] Speaker D: You get it right? Was it Curiosity? Well, there was a couple there at once. I remember anyway, I remember people saying like, what do you mean this camera is only 2.1 megapixels? Because at the time they were like 20 or 30, whatever the number was. And it was. So much of this stuff was built years ago. Like you just said, stress test harden, you know, cosmic rays, radiation, all these things that most people don't think about when you're on Earth. They were built years ago. So luckily most of the stuff's priority built. But at the same point, it needs to be hardened to such an extreme level that, you know, it's gonna work. It's not like you can send out a spare, you know, anything to Mars easily. It's. That can be something simple to do. [01:25:07] Speaker C: Yeah, I mean, NASA seems to be changing that. Right. Like, the James Webb telescope was developed over the course of two decades. Right. But the new Roman telescope, which is scheduled to launch later this year, it's only been in development for a couple years, and the whole mission ethos was on time and on budget. And so it's kind of an interesting play by NASA. And, you know, it makes. It does make sense to me considering, you know, the political headwinds I. I'm sure they're going to face and trying to get all of that settled. But it is kind of fascinating. Makes me want to, you know, get out of sass. [01:25:45] Speaker D: It was spirit and opportunity. It was bothering me. I knew it wasn't right. It was spirit and opportunity. Opportunity, yeah. And curiosity was the other one that was there at the same time. [01:25:55] Speaker A: Okay, sorry. There was three. [01:25:57] Speaker D: It was bothering me. [01:25:59] Speaker A: I mean, unfortunately, AI might get you out of sass. I mean, that's. [01:26:02] Speaker C: Yeah, yeah. No, it's a. It's fun. I mean, it's. It is like you said, it is a very different world when you develop on health applications or space applications or. Or anything in the practical world, which [01:26:17] Speaker A: is kind of cool. [01:26:18] Speaker C: I have a buddy who works for. For a health device company, and his funny anecdote was that he had to remove all references because it's in healthcare, and so there can't be any fatal. Fatal layers anywhere in the code. [01:26:31] Speaker A: Well, that's. I mean, that's. I don't know. I was like, that's awesome. Yeah, I. I love this. It's sort of like, you know, when we had to remove Master from, you know, GitHub and switch to main, I was like, I. This makes sense to me. I know why we're doing this, but, yeah, this is a much better story. So. All right, gentlemen, we'll see you. Well, I'll see you in two weeks, Ryan and I. Matt, good luck with Jonathan next week, and we'll see you after. Hopefully. Me winning or one of you guys winning. I don't know. [01:27:00] Speaker D: Somebody will win. [01:27:01] Speaker C: I mean, you can win the predictions. I hope I win in Vegas. [01:27:04] Speaker A: That's fine. Yeah, I'd like to win in Vegas, too. [01:27:07] Speaker D: Well, I'm not going to Vegas, so I guess I have to root for winning. [01:27:11] Speaker A: All right, later. [01:27:12] Speaker C: Bye. [01:27:13] Speaker D: Bye.