This transcript was generated automatically and has not been fully reviewed. Timestamps are approximate and wording may contain errors.
00:07 Welcome to The Cloud Pod, where the forecast is always cloudy. We talk weekly about all things AWS, GCP, and Azure. We are your hosts, Justin, Jonathan, Ryan, and Matt.
00:18 Episode 376, recorded for October 13th, 2026. Google Gemini agents, Gemini, Gemini confused. Whoa. I don't know what that means, but we'll find out later.
00:28 Hi Gemini.
00:29 Good evening, Matt and Ryan. How you doing?
00:31 Gemini, Gemini, Gemini, Gemini, Gemini. Hi.
00:34 Yeah, it's like, it's like, uh, what's Bloody Mary? You say it enough times, it like comes and attacks you.
00:39 No, it just spends all your tokens.
00:41 Yeah. If you say it enough times, Google will hopefully, you know, take over first place in the AI race they started, but they're, they're basically losing at this point. So yeah, whatever. Good for them. Good for them. Well, how's, uh, how's it going other than, uh, I've been sick all week. How are you guys doing? Sorry about that. I'm doing okay.
00:59 Yeah.
01:00 When you travel 6 out of the last 4 weeks, sickness comes with it. It's just part of the territory. And then I have, I was home just long enough to be sick and get better and then go travel again.
01:09 Uh-huh.
01:10 Rough. Yeah, it's been rough. Uh, but, uh, it's all right. Uh, well, we have a bunch of stories to get into this week, so we should do that. First up, uh, new Rimini Street survey data confirms the ongoing VMware exodus trend with 90% of 300 surveyed organizations now actively exploring alternatives due to licensing costs, and 54% citing Broadcom's termination of perpetual license support as a reason for them moving away from VMware. The cost increase reported says the Broadcom Azure continued to vary widely from 100% to 300% for most customers, up to 1,000% in some extreme cases, reinforcing earlier reports that rather than interesting new pricing shock figures. And I have to say, congratulations Broadcom, uh, now becoming the new Oracle. That's really well done. Uh, I mean, I've also heard of them going after companies for Spring Boot licensing issues if they were using the VMware version of You know, all kinds of fun things they're going after companies for, trying to squeeze every penny possible out of these companies. So, uh, pretty rough. Now, I do point out this survey does come from Rimini Street, which does provide a third-party VMware support contract. Just keep that in mind. But, uh, still it's true. And we actually have some other evidence of that is Topgolf, uh, the very popular driving range business, uh, apparently had an article in Ars Technica where they saved over $11.3 million after leaving VMware.
02:26 Wow.
02:26 It's a pretty good amount of savings by moving to Proxmox or whatever they end up doing. I didn't actually see what they moved to.
02:31 Yeah.
02:32 But, uh, you know, save some money.
02:34 It's a lot of Topgolf that we would've had to spend there, guys.
02:38 Yep. It's, I mean, it's, uh, they must have just really, they must really just be banking on the lock-in of users. And I'm sure if most of 'em are, you know, going from, you know, it's a 3x cost, I'm sure they feel put over a barrel on that one. But, uh, yeah,
02:56 it's a large project to move also, you know. Yeah, small businesses or medium-sized businesses that have, what, 10, 15 servers, and sure, it's an added cost, but their IT departments are too overwhelmed, so they're just like, yeah, well, here you go, here's the extra 5x, what are we gonna do about it, versus, you know, spending the time and effort to move to something else.
03:16 Oh, and that's kind of the thing is if you You know, if you didn't know the price increase was coming and then you get the bill, you don't really have a lot of time. So, you know, if you had 15 hosts, maybe you could move quickly. But if you have a large fleet of virtualization, uh, you're definitely not going to be doing it then.
03:31 Well, then if it's like disaster recovery, you have to prove all that. If you're SOC or ISO or any compliance methods, like it just becomes a, a big hassle. But I did see that if you don't renew, they will immediately send you like a cease and desist like the next day if they think you're still using it without license.
03:46 Oh yeah. They think, well, because the people, there's people are arguing rightfully so that if you are perpetual licenses, you should be able to continue to use your perpetual licenses without support. And they're trying to come in and say, no, no, that's not the case, which I don't know if they've actually ever gone to court on that argument. Uh, but that's the cease and desist they basically sent you is that you're not authorized to use your perpetual licenses you paid money for and think you own as part of their terms. But, uh, yeah. Cool. Or the new Oracle of, uh, software, apparently Broadcom. So anyways, move off if you want to. Uh, they, they moved to VergeOS, which I have not heard of. Oh, either. But, uh, yeah, I know most, a lot of companies are moving to Proxmox and, uh, apparently there's also like Red Hat has a virtualization thing. And then of course Hyper-V, uh, if you're a Microsoft shop, OpenShift is, well, that's just Kubernetes with a fancy layer, uh, you know, fancy dress clothes, you know? So I, you know, those are options people are doing. Moving on to AI is how machine learning makes money. We finally have a new Claude Haiku 5.5. It's positioned as the cheapest, fastest, and most capable small model to date with pricing around 75% lower than Haiku 4.5 on average, which, you know, Haiku 4.5 was already super cheap. So that's crazy. Haiku 5.5 is the first Haiku class model with an adjustable effort setting, letting developers tune for cost versus intelligence. And it's designed for high-volume tasks like summarization, classification, database queries, and as a sub-agent alongside Sonnet 5.5., and Opus 5.5 for coding work. Anthropic cuts on its 5.5 cash repricing by 50% from $0.20 to $0.10 per million tokens, reducing overall cost on digital workloads by about 20% since cash fees make up a large share of token consumption. New monthly API credits are rolling out to Claude Max and Team subscribers with $100 for Max 5x and $200 for Max 20x and up to $500 for pooled for Teams accounts. Haiku 5.5 is also available on AWS, Google Cloud, Microsoft Azure, and anywhere else that supports Claude models. I mean, it took a year though for this to come out. They've released how many, how many Opus and how many Sonnet models between the last, since they released 4.5 of Haiku? Like it's, it's a lot.
05:49 And it was a year before that also from, was it 3.5 to 4.5? So in one year from now we'll be talking about 6.5, I have a feeling.
05:58 I wonder if that's demand or just because it gets prioritized behind Fable and Opus and Sonnet. I don't know.
06:04 Behind, it would be my gut cuz they don't make as much money on it, and it makes a lot less headline news when, you know, it's not, you know, locked down by the US government.
06:16 I mean, I'm, I'm definitely intrigued to try Phi Phi with a thingy model cuz like Haiku for like the show notes summarizer was not great initially cuz again, it, it takes very specific instructions, but you need a, and you need a little bit of thinking. You don't need a lot. So I'm, I'm actually gonna try some comparisons between Haiku Phi Phi and what we do with Sonnet for show notes. Um, I bet I get pretty close. I don't know if I get close enough that I would replace what we have, but, uh, I am interested. I am intrigued enough to try. So your little follow-up bot can, uh, your little follow-up bot can task that to me. Yep.
06:49 Actually, I thought I was supposed to add one to the show. Maybe it didn't.
06:52 I haven't seen that work yet. You told me it was supposed to,
06:54 but it worked last week when you weren't here that week.
06:58 So I didn't see it. So it never happened.
07:01 I know. Well, maybe you should have listened to us, Justin. Probably better that you didn't.
07:05 Maybe.
07:06 I've been using actually quite a bit where in my AI development workflow where I've told it to drop a lot more down to haiku in the last couple of days. And I've definitely, I don't know, I don't have data to say it, but it feels like it's burning a lot less and I'm not getting that much more on the other, like, QA and security reviews that I have set up in my workflow, I'm not seeing that many more findings than I was before with doing Sonnet 5.5. So I think it's actually been pretty good, at least for me with my workflow. So it's worth trying if you haven't yet.
07:47 Yeah, I don't have anything high volume enough or something simple enough in my workflows. Like, I'm not— I think if I was building more of a a user-facing app that was going to use AI on the backend, it'd probably be pretty good, but haven't really played around with it a whole lot.
08:02 I started a new project last week. It might have burned all my tokens in 3 days for the, for the week and then got mad. So now it was on pause until today when everything refreshed.
08:12 Well, I will report back. OpenAI is rolling out GPT-6 in ChatGPT with a new intelligent UI feature that generates interactive interfaces, buttons, forms, charts, and diagrams rather than plain text responses, choosing a format based on the question type. These are very similar to artifacts in Claude. The system uses a library of native streamable components plus a compiler that renders the interface progressively as the model generates it, avoiding a wait for the full response before anything appears. GPT-6 also introduces interleaved thinking and answering, letting the model start responding before it finishes reasoning. And per OpenAI, GPT-6, uh, instant begins answering 44% sooner on average than GPT-5.6 and GPT-6 Extra High. Safety updates include stronger resistance to multi-turn jailbreak attempts and updated training around cyberattack, biological, and violence-related misuse, building on safeguards referred to internally as Astra.
09:01 This is neat. I haven't really played around with artifacts all that much other than, you know, kind of accidentally would asking a question, it generates artifacts. And so this seems like a pretty useful feature for people using, you know, Codex and other ChatGPT apps.
09:17 I had to add to my Claude MD, please don't generate an artifact unless I asked you specifically for an artifact. Because it started getting to the point where it was like, it was just generating on every time. Like, I don't, I, I appreciate them. They're great in the right context. Sometimes I just want to be able to copy paste some text and I don't need this full picture breakdown. But in the, on the website, the link that goes to, you can see some of the examples of like the bike, the bicycle where they can pull out different components, highlight them in the picture. You know, that's the kind of things you can do with this feature. So it's neat. It has a lot of great capabilities. And again, just, you can see how this is a valuable feature to a lot of people. So, and you can create little mini apps almost with it.
09:53 So I will say I do like the artifacts and like the work environment. I know that's kind of a tangent to what we're actually supposed to be talking about here, but you know, it's useful when, you know, I'm building something. I'm like, hey, go check this out. Or like, you know, go check out kind of what I'm thinking and use that as the way to share. Like, here's what I'm developing or here's what I'm thinking for this project as like, you know, an in-between versus like dumping, like having it link up to Confluence and You know, having sending someone a Confluence link, it's kind of nice in that way. 'Cause then I can send it and have like, oh, I need to make a change. I can real time just continuously update in Claude. So that one URL kind of stays there even if you do make updates. So it does have some nice functionality in that way. And like multiple people can provide feedback on it and you know, it kind of becomes like a Google Doc almost the way I've used it.
10:41 Yeah. I mean, I, I, I don't like, don't dislike them. I just was mad that it. Decided that everything I ever asked for should be now be an artifact. And I'm like, I mean, I, I'll ask you if I want the artifact, but, um, I do like them and they're kind of nice. And even, even be able to create, you know, simple like, hey, I, I did a bunch of, uh, refactoring over the weekend. You know, so it, I had to just write a paragraph summary, but I could have done a really cool like, you know, artifact of that with some graphics and some code, you know, repository data. But I, you know, wasn't necessary for this use case, but you know, I could see the value of like, hey, I need to send something fancy to my boss to say what I did over the weekend. It's a nice little fancy way to do that. Databricks Apps on behalf of user authorization is now generally available, letting apps act under a signed-in user's identity. So Unity Catalog permissions, row filters, and column masks are enforced automatically without reimplementing governance logic in app code. Apps can mix two authorization modes: app authorization, which is service principal for shared or app-owned operations, and user authorization, forwarded user token for permission-sensitive queries. Choosing the right identity per request path. Now, how do you guarantee that though? Scopes like SQL restrict query limit apps to read-only SQL rather than full SQL access, reducing the attack surface, and scopes act as a capability ceiling, not a data access grant, and the user still needs underlying Unity Catalog permissions. Workspace admins can allowlist which API scopes developers are permitted to request, defaulting to all scopes but deniable or disabled entirely, giving the organization centralized control over what apps can do on the user's behalf. Sorry. Yeah, that's a cool idea, but, uh, I wanna guarantee that it's not gonna accidentally use the wrong identity when it's not supposed to.
12:13 Well, it's, I mean, I don't know that it's anything different than like doing OAuth in terms of choosing the identity. The difference is just an additional sort of key and, and validation that happens so that the actor that's acting on behalf of is, is also validating its identity. And so think about it in the terms of AI agents using using of that, it's, you can now, you know, have an AI agent be its own thing, but use the service principal permissions with the reduced scope. And we're going to see a lot more of this as people figure out how to do non-human identities permissions and including the, the three-way sort of token validation. It's going to be very popular.
12:54 Which does exactly what they said, which is pushes all this app, you know, pushes requirements to the app to basically figure out like, well, is Ryan supposed to be able to see the salary data or not see the salary data? You know, if this, if this particular chat agent talks to the HR database. And so, yeah, so being able to use a shared service role for like basic activity, basic use cases out of the thing, like, hey, I wanna, I wanna know who reports to, uh, Matt in the org chart. That's not something that needs my permission. But then if I say, okay, now tell me their salaries, now it should be using my local permissions to basically say, is he authorized to do that or not do that? Right now to do that, it's all built into the application tier. And so the app has to now have a whole bunch of user credential understanding, which is a lot of overhead and toil. So this does make more sense, and this is why this is going to be the future direction. But like, when you really start thinking about some of these use cases of like the agent and which identity it's using, like there are opportunities for data cross-pollination in the memory space that are somewhat dangerous. And so there are some interesting use, you know, attack vectors that could come out of these type of services. So I think there's some, some things to be concerned about. But over time, you can definitely see this getting hardened and be a much better place. But I, I don't know if I'd run out day one and go like, yes, this is ready for production, put it on my corporate website. But I mean, I think you definitely, with the right testing and the right things, this is the right direction. Well, if you were not excited about Claude or GPT-6 or Gemini and all of its amazingness, Meta has released their new full frontier models, the Microsoft Wonder— or sorry, Muse Wonder and Muse Wonderful, their newest agentic AI models built on earlier Muse Spark foundations targeting expert domains like law, finance, health, and coding tasks. There's two features that kind of stand out above the others, Active Memory, which retains relevant context across conversations over time, and Active Focus, which allows the model to persist through difficult tasks, trying alternate approaches, recover from errors, and check in before taking significant actions. Safety measures include training the models to avoid sharing private information without permission, refusing harmful requests, and support security teams in finding code vulnerabilities, all evaluated under Meta's Superintelligence Scaling Framework. Availability is immediately through the Meta Model API and Muse Code with broader rollout planned for Muse subscribers. Position this as infrastructure for agents handling real-world tasks like reviewing leases or managing insurance claims.
15:06 It's interesting because usually the memories and that sort of management's in the harness level, and so it's— it sounds like these are built directly into the models themselves, which is kind of fascinating. It's something I haven't seen before.
15:20 Yeah, it feels like it's a different approach that they're taking on this, and
15:28 It's, yeah, I, I don't know enough about it yet to really comment one way or another, but I do, I, it is interesting that again, I feel like they could have really owned open models and they, this is continues to be a closed model. So that's my only one disappointment about it.
15:43 Yeah. Well, hopefully they keep releasing open models as well, which it seems like they have recently.
15:48 I feel like they're done.
15:49 Didn't they recently?
15:50 I thought they stopped.
15:51 Thought there was one since, yeah, 'cause they, I can't remember now.
15:56 Well, I guess, I guess open weight, open model is Muse Glimmer, which was released in August. So that one is open weight, open model, and that one's designed to run on laptops. But, so yeah, I guess, I guess they have. That's fair.
16:09 I'm still not certain that I want Meta owning all my chat conversations. Just saying.
16:14 Yeah. Of all the companies who I'm not gonna give my data to, I'm not gonna use a model on principle, is this one.
16:19 Yeah.
16:20 Uh, everyone's really excited about their OpenClaw competitor, which I'm forgetting the name of right now. And I, I just look at it, I'm like, yeah, I'm never gonna use that. I just don't trust you with my data at all.
16:31 Isn't their OpenClaw data just Muse?
16:35 Uh, maybe. I don't remember what it's called.
16:37 Yeah, I think it is. I think it is. I think it's just Muse.
16:39 Yeah, that might be. But yeah, I just, you're not, you, you can't pay me enough to give them my data. Just like I won't use Grok for my, for an agent for myself either. Like I'll, I'll use Groq for searching Twitter cuz it's really good at that. Uh, but I, I have zero desire to use Groq for anything in my personal data. Sorry Elon, you're not getting me that way. Also, I don't use Starlink. Yeah. So other than unless I'm on a United Airlines flight, but I'm on VPN. Cause it is pretty, cause Matt's been saying how amazing United Starlink service is. And I was like, yeah, everyone I'm on is on the stupid satellite internet. It's costing me $8. Then I was finally on a flight and I was like, okay, this is pretty good.
17:15 It's pretty good. It's been a while since I was out. I finally was out last week and I was like, it's so nice. I can actually do things.
17:23 Yeah. Like, I mean, night and day difference. I had, I'd just done a, I was on Connection through Denver, so my first flight was the legacy satellite and I was like, you know, had all kinds of problems, kept disconnecting, you know, couldn't do a bunch of things I was trying to do with my agent. And then I go on the next flight and it's Starlink and it's like, oh, I haven't had a single dropped packet the entire time I've been on this flight. It's crazy. So, wow. That is messed. Yeah. Well, Satya Nadella posted on X that all AI models should be treated as potentially compromised by default rather than as trusted, as opaque black boxes whose outputs are simply accepted or rejected. Key recommendation: an emergency brake capability, meaning authorized personnel should always be able to pause or shut down a model mid-task, with more advanced models requiring standardized containment technologies. Other proposals include timely incident disclosure, independent audits, verifiable data, and tamper-proof human-readable evidence logs for model actions and decisions. These ideas align with broader industry discussion on AI governance and safety, though Nandell's emphasis on assuming compromise and requiring kill-switch-style containment is a more specific technical stance than many others. I want to know what he saw that made him post this tweet.
18:28 That's— well, I hadn't thought of that, um, because I was more bothered by the language, just because I feel like this is a sensational headline for a problem we've been talking about for weeks, right?
18:38 Like, it's— I mean, there are always sensational headlines. I mean, anytime Claude, Anthropic CEO, speaks, I'm like, oh, here comes a sensational headline of like jobs being destroyed or whatever.
18:49 Your, your model or your agent doesn't have to be compromised for these, these, uh, proposals to make sense, right? You want, you want to have the ability to stop a rogue process or program or anything from going off task, and you want to be able to do, uh, verify the data and Whether it's, you know, the, whether it's a script running on your laptop or it's an AI agent in some harness somewhere. Like, I don't really, like, I get it. The, we need to talk about the concerns because it does, AI definitely does introduce different, you know, different scale of problems, which, you know, we've seen with the OpenAI stuff. But all models are compromised. Like, it doesn't, you're just trying to get looks. I don't like it.
19:33 Yeah. I mean, like, how would they be compromised if they built the, you know, if you're gonna say OpenAI's model is compromised and it's designed and built inside of Azure, does that mean Azure's compromised? Mm-hmm. You know, small things. Maybe careful when you think about the chain of trust. Matt, you have real-time follow-up on the tracking feature.
19:55 So the tracking feature works Almost all the way. So it did the 3 researches like we wanted it to do. So one of which was around Outposts that Ryan Lucas wanted to get a lookup apparently, determine Outposts hardware generation and migrations actually work or not. And it found that there's no in-place hardware upgrades that you have to move from Gen 1 to Gen 2 for Outposts. The problem was it decided that we had taught it, it has a logic problem where If it finds it, it closes it versus always adding it to the notes. So it found it, it researched it like it was supposed to do and said, okay, I'm done versus pumping it into our show notes for us to talk about.
20:38 Hmm.
20:38 That's good. The second one of the three. Hey, I got 95% of the way there. Um, it's not like we have a great QA environment. We just tested prod. It's fine. The second one was around—
20:50 I mean, I test on my laptop, but whatever, you know, you test in prod.
20:53 I tested prod. It's fine. It's just bolts.
20:56 How dare you?
20:58 Justin, just reach across.
20:59 He'd be insulted.
21:01 Yeah, he is insulted. Don't worry. I'm sure I'll get some sassiness from him later. The second one was for Jonathan. Episode 345, find pricing of Bedrock Core policy. It looked up and found the pricing of it. And it goes through all the pricing, which is less interesting to talk about live. Again, that one was below the threshold of where it was going to automatically say it was good. Um, where I guess I think I marked it as yes, this is the right information, but again, it didn't pump it to the, to the notes. And the last one was around Firecracker GPU that Ryan Lucas, again, over here, we're supposed to talk research about. Ryan had a lot of follow-ups from when I did this. So, you know, Which was determine whether Firecracker will support GPU segmentation like NVIDIA Mix, which had an 82% confidence. Firecracker does not support GPU at all today with, with no PCI/VFO passthrough. And it's actually sitting there waiting for somebody to click yes or no on that one.
22:10 I think we should limit, 'cause that's going back. A long time.
22:14 It's only this year is all we ingested.
22:17 Okay.
22:18 And it's still going through the backlog. So we'll research 3 at a, 3 at a time. And if it doesn't get any information on 3, it will do up to 2 more. So essentially up to 5, it'll research. So we don't end up like researching them all at once. So currently there are 33 items in the open tracket.
22:39 With my goldfish of a memory, this is going to be a rough task for, for me to remember the track of things. I'm like, what did I talk about?
22:46 I mean, it adds you in the channel, but 17 things at you at this point. So, you know, you successfully ignore everything in the production channel. So,
22:56 you know,
22:57 it's fine. So Justin, it works 98% of the way there. It just didn't post properly. Which I blame Google Docs API, even though I don't know why, what the root cause is, but it's much easier to blame their fun APIs that you like so much.
23:10 I mean, I always blame their API. That's
23:12 what I do. Yeah.
23:12 Yeah.
23:13 It's, it's rough. It's the easiest way to go. Well, I'm glad it's working. So I'm, I, how'd you guys introduce that feature to the listeners? But I wasn't here that week.
23:22 Oh, to torture me, obviously. Like, it was a feature built specifically to torture me. I figured it out in advance, ruining the surprise, and I am very happy about that.
23:31 Nice. I, I mean, I have like 20 or 30 follow-up items as well. I just need to close some of them 'cause I'm not going to do them, but there are, I will, I definitely like the idea of having a reminder to do some of them. 'Cause like some of them are, I was like, oh yeah, I did wanna do that. And so I went and did it. It's like, okay, that's good.
23:46 Yeah. I'll definitely look. I haven't looked yet. I, I do have a tendency to wander a little bit, so I wonder if my speech patterns are, uh, particularly subject to being treat it as like, oh, I should get back to that. I'm just making excuses to stop talking.
24:02 Could be. Well, that'll teach you. That'll how you got rid of ums, which is like a trigger for yourself.
24:07 So it's true.
24:09 Uh, all right, well, moving on to, uh, cloud tools. Gremlin, uh, which is one of the many chaos engineering platforms out there, like Chaos Monkey, launched Foresight AI, an agentic add-on to its chaos engineering platform that identifies potential failure modes, recommends and runs tests, and applies fixes and verifies reliability without requiring deep distributed system expertise. From every engineer. The system uses 4 agent personas. The analyst gathers environment details and recommends tests. The tester runs the reliability tests. The operator provides remediation guidance. And a technical program manager tracks reliability, health issues, weekly reports, and managed commitments with assigned dates. The key differentiator is the failure atlas, a proprietary dataset built from millions of fault injection experiments across tens of thousands of real systems over the last decade. And Gremlin states customer data is not used to train the underlying LLM. The tool approach relies on actual task execution rather than LLM inference alone, meaning the LLM is used for summarizing and explaining results while decisions are based on data from reruns of tasks against the Gremlin platform. I mean, first of all, the fact that the Gremlin platform has been around for a decade to have that much data makes me feel old. Yeah. But second of all, this is cool. I like that because one of my big complaints about like failure mode event analysis, which is a pretty popular way to do some of these type of events, is that it's only as good as your imagination. And while I have an imagination of an evil operator from hell, Most of my peers don't. And so I always come up with like some really good ones and they look at me like I had 5 heads and now I can just have Gremlin do it. So that's great.
25:33 And it's the system, right? The system does it. They'll actually listen versus the guy with 5 heads.
25:39 Yep. All right. AWS S3 vectors is now supporting metadata pre-filtering for higher recall on filtered searches. Uh, metadata pre-filtering, which resolves filter conditions before running the similarity search rather than evaluating them in tandem, improved recall on scope queries by up to 5x on highly selective filters. The feature targets multi-tenant applications like RAG systems, GenAI workflows, and document search where queries typically target a narrow subset of an index, such as one customer's record out of millions of total vectors. Implementation requires switching an index from classic to enhanced mode via the update_index_mode API, and existing vectors don't need to re— be readjusted and queries and text stays the same, making adoption straightforward for existing deployments. The new starts_with operator enables prefix matching on paths, URLs, and hierarchical keys, usable for scoping searches to folder structures or document ID patterns, joining existing operators like Equality, Range, and Set Memberships. No additional cost for pre-filtering. Customers pay standard S3 Vectors pricing for storage, PUT requests, and queries, and the feature is available now in all commercial regions. I was hoping this would be something we could maybe add to Bolt, uh, because he uses Amazon S3 Vectors for all of its work. And then, you know, it really doesn't help us actually at our size because we don't have enough data in the vector database to make it worthwhile, which is a bummer. But, uh, still, if you have a large RAG workload and you're struggling with some of the narrow searches, this is a good feature for you.
26:54 Yeah, it seems like a cool concept, you know, doing any, any pre-filtering on that metadata to save like a more expensive lookup is pretty cool. It seems like a nice feature that they're just adding in as like a, you know, along the existing feature, which is neat. I like it.
27:12 All right, well then, yeah, no, it's good. AWS Batch now publishes job metrics to Amazon CloudWatch. These now get published natively under the AWS Batch namespace, eliminating the need for custom instrumentation to monitor batch workload health. Metrics include state transition counts and duration tracking segmented by job queue name dimension for pure queue visibility. This addresses a common pain point for batch processing teams who previously had to build custom monitoring solutions to track job failure rates and queue performance. Available now to you in all regions where batch operates.
27:42 Well, batch operates is a very— it's doing a lot of heavy lifting as far as I'm concerned. Batch is not a real service. It was never finished. I don't know what pet project someone decided to go live with, but it is terribly hard. And I think it's a service looking for a job. I don't really understand why you would need to—
28:04 I've done it for like migrations and things like that. Like, do we need to move all the data from here to there? Or, you know, like one-time, one-off things that you need to run at massive scale. I've never used it like reoccurring within a production workload. It never was like a feature of a service. It's always been like, oh, I need to do this thing, let me go do it this way so I don't have to orchestrate it myself.
28:29 Yeah,
28:30 I mean,
28:30 I mean, you do still have to orchestrate it yourself though.
28:34 Like, it's a little bit more done for you.
28:36 I mean, when was the last time you actually played with batch, out of curiosity?
28:39 I mean, it's been a couple years. Yeah, but I, I, I
28:43 Cause I mean like the main feature of Batch is to orchestrate and schedule your jobs that then you then run on CloudWatch or something on, on, on Docker or run on Lambda functions. I mean, that's my experience with it is it does that relatively okay. It doesn't do really complicated orchestration, but it does basic orchestration.
29:00 Yeah. I mean, I moved on from Batch to State Machine real fast, um, just because it was much easier to use, um, and had more feature parity and batch, you still, I mean, at least at the time I used it, you had to put all of your logic inside that container.
29:15 Yeah, exactly. Which, so your, your use case was not meant for batch. So batch is like, hey, I need to try, I need to process 100 million rows that have to do some, some very specific set thing over and over again. And I need to do it on a very scheduled basis. That's, that's the batch use case. It's a, is a mainframe use case before it's anything else than that. And you had, your use case was more No, you have parallel, you have fan-out parallelization in State Machine. That's exactly what it does.
29:41 Like,
29:41 yeah, but not,
29:43 yeah, but fan-out and parallelization only goes so far and you hit other limits because of it. And then like, it's like if you wanna do a complicated thing, like inside, like, so I've done it where it's like, okay, I need to take something, do a bunch of manipulation to it, and then kind of put the file somewhere else and then trigger some You know, and like where it's like, here's my business logic. I need a repeat.
30:09 Yeah.
30:09 Like, yes, state machines could work, but this was good for like, I already have the logic coded up in a container and things like that.
30:19 So, I mean, it's, if you need to run on your own VPC, there's, there's difficulties with it and all the logic needs to exist within the thing. So it's, it's, I don't know how much it's better or not. Like,
30:30 I mean, I, I can tell you, uh, now having a newfound respect for batch processing at this scale, I will tell you that while this is not the best solution for it, it is a solution for a pretty big problem that a lot of companies have, including both of our day jobs. But yours, your, your experience with it is much smaller scale than mine is. Yeah. Uh, so I'll tell you, uh, yeah, this is a thing. Um, You know, I have super high volumes. It's definitely valuable, but you know, again, if you need something more, you wanna make more decisions, you wanna do things, I would tell you batch is the wrong decision for that.
31:05 So, yeah. Okay. Because I thought that was the whole point was the decisions and having it run a same set of business logic over and over and over again.
31:14 Mm-hmm. It's, it is truly about running
31:17 batch jobs,
31:18 crazy amounts of volume in batch mode into different APIs or services or functions or whatever that's highly repeatable. Not— it doesn't need to have decisions made based on different data types. It's literally, here's a billion of this record, of these records that need to all be processed as fast as humanly possible into this database so that we can then do whatever our next step is. And it sounds like a really simple, dumb use case, and for most companies it is, but for companies that do massive amounts of things that are all very similar, uh, like a bank. Very important.
31:51 So, but if you run into any kind of concurrency limits on your dataset, uh, the whole thing falls down.
31:56 But you, that's my point is these systems
31:59 batch organize
32:00 it
32:00 for you.
32:01 Yeah.
32:02 I mean, you could like say you're doing inserts into databases, like, yeah, but that would be an RDS limitation, not batch.
32:08 No, that's what I mean. But you can't put any kind of logic inside the batch that would handle that use case. And that's not cool.
32:13 I think batch does do retries, I thought.
32:16 It does do some retry, but again, it's also, it's, it's also not a database like you think of a database either. It's going into IMS, uh, which is a mainframe database language, which is just typically flat file based. It does, and it is the most scalable database in history, which is why mainframe still exists at a lot of places. So anyways, we can talk about this. We can nerd about this for a while. I'll spend another time. But, uh, yeah, I'll tell you that I also thought your, what you wanted to use batch for was the right answer. And I will tell you, I have new data that tells me I I'm wrong. So yeah, I've now seen it at scale that I'm like, oh, oh yeah, this is a different problem than I was thinking it was. So, okay. Uh, AWS has released Transbox in developer preview, an open-source sandbox, uh, licensed under Apache 2.0 that pairs OS-level isolation with fine-grained policy enforcement for AI agents, addressing the risks of agents running in YOLO mode with unchecked access to files, shell commands, and APIs. Available to you at github.com/strands-agents/box. The system combines two components: OS-level containment, which is currently a macOS solution for hard boundaries, and the Dogwood policy engine for granular rules about what agents can do within those boundaries, including temporal rules based on action history, like rates limiting Slack posts to 3 per 10 minutes. Policy enforcement happens outside the agent process at 4 interception points: network egress, Python interpreter, shell interpreter, and MCP broker. So rules apply consistently regardless of which tool the agent uses and cannot be bypassed by the agent itself. Pesky agents that are compromised, per—
33:44 Yes, exactly.
33:45 Uh, notable technical detail is credential handling. The egress gateway injects real API keys and AWS credentials on outbound requests, meaning agents work with placeholder tokens and never see actual secrets, reducing credential exposure risks. Well, that's a pretty nice feature.
33:58 Yeah, no, I like this. You know, like, I like Docker policy validation and how it works and You know, this is something I struggle with in a day job. You know, when you set up a lot of like cowork or enterprise, you know, sandboxing, you literally have to define the allowed actions or, you know, hard things. And so you can't really define anything in like in a policy that would make sense that, and could be a little bit more adaptive. And so I don't really want to, you know, dictate all the different commands that are okay to run in the sandbox and getting, you know, credential information into the sandbox is incredibly hard with a lot of these things. And I'm hoping that, you know, this solves some of those problems, which I think is pretty cool.
34:40 I feel like it's a step in that direction, but I still feel like the problem is pretty hard to solve for. Like, yeah, I feel like we're getting there. We're getting, you know, agent-based identity, identity like we've talked about, you know, and on the past of the show. And we're getting, you know, this injects the credentials in real time in, so it's only good for X number of time, you know, I feel like we're getting there, but I feel like there's still more to still do to make the security person happy.
35:08 Oh, for sure. There's, I mean, the compliance and audit tracking, it's got to be really hardcore and mapping the identity, which, you know, no one really knows how to fix yet, you know, is still a problem. So I agree. I think there's a lot more work and I think it'll, it'll continue like because of Of course they will.
35:26 But these are all great steps in that same direction, you know, of getting us to where we need to be. You're not gonna boil the ocean day one. This is gonna be, you know, 7 steps forward, 3 steps back, and, you know, maybe a nuclear bomb that explodes along the way, aren't ya?
35:40 Yeah, I mean, but this type of thing, like when you talk about agent kill switches, like this is exactly the type of policy you can define in Docker, you know, 'cause it can use the metadata or the data, the context around the request to, to enforce a policy rather than just saying this rule, yes, no, which is cool.
36:00 Well, in June, Amazon had its first Prime Day of 2026, and so they're here to brag about the amazing scale of their sales event. And hopefully they'll have lots of money in their earnings to justify all this spend, which they don't pay for. So, you know, they do, but not directly like we do. So, you know, the impressive numbers are as follows. DynamoDB processed over 59 trillion requests across the 4-day event, peaking at 192 million requests per second while maintaining single-digit millisecond latency, illustrating the scale of AWS managed databases handle under extreme load. Their serverless compute saw heavy use, with Lambda handling 2.3 trillion invocations per day, and ECS Fargate launched an average of 158.3 million tasks daily, a 47.7% increase over last year, showing continued customer adoption of container and function-based architectures for variable workloads. Security and reliability tooling scaled significantly as well, with GuardDuty monitoring 14.08 trillion log events per hour, up 59% from the year before, and AWS running over 44,000 fault injection service experiments, 6 times 2025 volume, to validate resilience before the event. Networking and messaging services posted notable growth too, with CloudFront delivering 2.1 trillion HTTP requests, up 5% from the and ECS receiving a peak of 213 million messages per second, reflecting the interconnected microservices architecture behind Amazon.com. AWS Graviton powered up to 49% of the EC2 compute used by Amazon.com during the event, reinforcing AWS's push towards low— towards custom silicon for cost and performance efficiency at scale. And AWS also promotes its Countdown premium service for customers preparing for similar high-stakes customer events.
37:38 Yeah, these, these numbers always continue to astound. Like it's so such a large number and I can't imagine the sort of business-wide coordination that putting something together takes, which is really impressive.
37:51 I have friends that work at Amazon and it's full-on war room setup for the multiple day events with full teams on call for it. Cause you don't want to be that team that causes a problem. You know, that does not feel like a good thing. The one metric I don't think that you said that I found interesting, mainly because I was looking up the number of zeros, was the 2.15 quadrillion metrics observed per day. It's a multi-day event. So 6 quadrillion metrics and at 10 cents per metric, that adds up real fast if you actually pay for all that. So,
38:27 well, those are custom metrics, right? 10 cents per metric.
38:29 Oh, that's true. Yeah.
38:30 Yeah.
38:31 Yeah. Well, I mean, you, I guarantee you they're using custom metrics.
38:34 Oh yeah, yeah, yeah. Just not for all of them.
38:37 Yeah.
38:38 But it's the big numbers.
38:40 Yeah. Yeah. Well, I, uh, glad to see another successful one. I think we have another one coming up again, which is here in September. I think it just passed us actually. So their, their fall one.
38:49 It was earlier October.
38:51 October.
38:51 October.
38:52 Yeah.
38:52 Yeah.
38:52 It already passed. Yeah.
38:53 It just passed. I, uh, I bought nothing during either one this year. I have a, just didn't have anything to buy.
38:59 Your wallet thanks you.
39:00 Yeah,
39:01 yeah, my wallet and my cardboard recycling guy, which is also me. Uh, AWS IAM Identity Center now supports network access controls, uh, for Identity Store. This lets admins lock down the Identity Store API and SCIM API to specific networks, addressing a common security gap where identity management APIs were accessible from anywhere with valid credentials. Admins can require Identity Store API traffic to originate only from approved VPC endpoints or source VPCs, while separately restricting SCIM API access to specific IP ranges, such as an external identity provider's published range. A use case for this would be the organization uses Okta or Azure AD or similar provider for SCIM provisioning, can now whitelist those providers' IP range— wish good luck finding them— while forcing internal custom apps and provisioning workflows through private VPC endpoints only. This feature is opt-in and off by default, configured via the update identity store API. Though, through SDK or CLI with no additional cost. So this is my time now to plead Amazon to, hey, work with all your partner companies like Okta and Azure AD to, you know, provide a managed firewall rule for this so I don't have to figure out what their IP addresses are in their terrible documentation. Because again, you just have to publish a, you know, a rule that I could then plug into a security group or into this policy. I'd be really happy if you could do that. So,
40:19 and then you have to maintain it too. So that means you have to watch it and update it. And, you know, I've done things like that with Terraform for providers, but then you have to make sure you run your Terraform every so often to see if there's drift in that, if it pulls, or you have to monitor the, was it the JSON file that both AWS and Azure I know have? I assume Google has one too. Has all the IP addresses of all the services and deal with that mess. You know, so it's a great feature. It's an extra lockdown step, but you gotta be on top of maintaining it if you set this up.
40:55 And it's absolutely required for some compliance frameworks, which is the only reason I know about it at all. And
41:00 so,
41:00 like, it's a silly control, especially when you're mapping out like 4 trillion IP ranges, you know, IP ranges going, yeah, it's anywhere from here is fine. Like basically most of the internet. I do think it's important, especially for the services where you can have the identity provider sort of come from your own IP space. I think it's a pretty good use case there.
41:21 Yeah, it's good if you're running your own on-prem still or something like that. I could see being useful.
41:26 Yeah, I mean, I, again, it checks a box. I don't have to worry about it anymore. I'm pleased with that part of it. But, uh, yeah, it's, it is a little bit of a security theater of like, oh well, you know, we, we restricted the IP address. Okay, what Someone can spoof IP addresses. It's possible. So if they're really that committed to hacking you, it's possible.
41:43 Mm-hmm. Defense in depth. They still have to have the identity credentials and all this stuff. But yeah.
41:49 Announcing AWS Lambda Web Functions, which means clearly Amazon is starting to get concerned about CloudFront. So Lambda now splits into 3 form factors. Event functions, which is the Lambda that you love and know forever. MicroVMs, which are built on top of Firecrackerization and have been around for a while. And now new Web Functions, uh, which are now available to you. Web Functions let developers run standard HTTP servers like Express, Fastify, Next.js directly on Lambda with no code rewrites or adapters using a single CLI command for deployment, aws lambda-web deploy. Each execution environment supports up to 128 concurrent requests, and billing is based on active CPU time rather than wall clock time, so I/O wait during database or API calls doesn't accrue significant compute costs. A multi-region deployment and latency-based routing are built in. With a single command, can replicate the app across regions like us-east-1, eu-west-1, and ap-northeast-1, with traffic automatically routed to the lowest latency region via global URL. Integrate with CloudFront using Origin Access Control for private origin security, enabling a pattern similar to S3 origins, combining CDN cache static assets with serverless dynamic compute. Available to you in 5 regions: us-east-north, Virginia and Ohio, us-west-oregon, and Europe-Ireland, Asia-Pacific, Tokyo, wrapping up the 5. Runs Node.js 24 on ARM64 with VPC connectivity support. Pricing details are available to you on the AWS Lambda pricing page if you wanna see them. And as I mentioned, CloudFront also supports this as an origin, so you get both.
43:13 Awesome. I missed the fact though, when I read the first article when it was released, the multi-region support. That's kind of nice that you don't have, you can just set it up to deploy in that way. You know, it's like you're writing something simple, here you go, set it up, CloudFront in front of it. You don't have to worry about a regional outage. It's a single command versus before it was always, multiple deployments and everything else like that that you always had to take on, which was always a little bit of a heavy lift.
43:37 Yep. This also is a pretty big shot across the bow at Vercel, I would say.
43:41 For sure.
43:42 Yeah. So Vercel, CloudFront, Route53, kind of trying to go like, you don't need all this server infrastructure, you know, you can just run on serverless, run your Vercel app with us, run your Cloudflare web app, you know, and get all those functions very similar. So yeah, I'm glad to see this as one more option. I'm kind of curious how I might use it for the Cloud Pod actually. Um, I gotta do some thinking about, I've already been planning to go off of the WordPress backend. Um, so now I'm like, hmm, I wonder if I could use this to make a much lighter weight version of a static app in some way. Cause I do need some dynamic stuff, like to allow people to comment on the website, to allow people to buy stuff. There are some e-commerce comments. So there's, there's a couple of edge cases that are on the podcast website that isn't just the RSS feed to everybody for the episodes. So I've been racking my brain about how I wanted to handle all that mess. And so this gives me one more option in my playbook. So excited about
44:30 it. Uh, the first thing I thought of was that, uh, there's a couple of different use cases where I had to run like a single container on a node somewhere so that I could do stuff like run nginx proxy to URL rewrite to redirects that were more dynamic. So for, for any kind of Lambda-backed API. So this is, I replaces all that use case, which is pretty cool.
44:49 Matt, you have any use cases for this one?
44:52 I definitely feel like simple, small apps is kind of where I'll do it, you know, Probably more like personal projects, but I really just, most of the time when I do stuff, I just try to go, you know, CloudFront to S3 as much as I can. I'm building pretty much everything because it's free.
45:08 Yeah. I mean, I've been driving more and more towards static apps. So like my personal blog is now a static website. You know, my, The Cloud Pod I'm trying to make into a static website. It's getting like, you know, A, I don't want to, I'm tired of patching WordPress. For sure. As much as I like WordPress, it's, it is, you know, constantly trying to be attacked on the world. And, you know, I get firewall WAF reports every day of like all the places trying to brute force our web, our website. I'm like, you clearly don't think we have valuable data, or you think we have valuable data, which we do not. But, you know, it's still annoying and it's something I have to secure and worry about and keep me up at night. So I'd like to not worry about those things anymore. So it's definitely something on my radar. Uh, and Amazon ECS is now adding live task count metric and removes the charge for the running task count metric. And so to break this down for you just a little bit, uh, the live task count gives you a complete picture of active ECS tasks across activating, running, and deactivating states, including tasks still draining connections during scaling events, making it more useful than running task count for auto scaling and capacity planning decisions, which is what you would like. Uh, live task count is free and automatically available without opting into CloudWatch Container Insights. Which is apparently what you used to have to do for running task count, which is annoying because CloudWatch Container Insights is very expensive. Running task count, previously a paid Container Insights metric, became free for all users starting on October 1st. I'd like a credit, please. Reducing monitoring costs for existing ECS customers who rely on it as a stability signal. So nice to see this move over to the right place. The fact that it needed Container Insights, I always thought was dumb. And I use this metric probably all the time. So.
46:43 Yep. Yeah, I just basically assumed that Container Insights was required for ECS workloads because of this. And so, and in fact, I struggled to think of any other metric that I was getting out of Container Insights besides this. So I'm really glad to see this.
46:57 I definitely turned it on.
47:00 There are some good insights on there, but the price makes you realize those insights aren't worth it. So. Uh-huh.
47:05 Yeah, I put it on, on like my personal cluster and then saw how fast my bill went up and it was like, Okay, all these things that I wanted that were useful, yeah, I don't want them anymore. Not for my personal random stuff that's running on a spot that dies because I'm too cheap to deal with it.
47:22 That's when you're like, how do I send this to a Grafana node that I run on my own box? That was my experience with Container Insights. I was like, oh, this makes me really want to use Grafana on a container. All right, let's move to Google. Google is releasing the MCP Toolbox for Java SDK version 1., which will give you an MCP protocol to enterprise Java environments, letting AI agents connect to databases and data sources through a single standardized protocol instead of a custom integration for each model database pairing. Key version 1, uh, additions include decoupled client authentication with async credential resolution, automatic pruning of server-side bound parameters so LLMs can't manipulate sensitive fields like tenant ID, and runtime warnings when credentials travel over plaintext HTTP. The SDK integrates with Spring Boot and LangChain4j using annotations like @MemoryId to handle conversation state across HTTP sessions. Reducing the need for custom intent parsing logic in conversational agents. Google demonstrated the SDK with a transit booking example using AlloyDB, showing how an agent can combine unstructured data, like policy lookups, with structured queries and have transactional booking in a single conversational flow. Deployment is designed for Cloud Run with the MCP Toolbox server and Spring Boot agent scaling independently. The SDK is open source and available via Maven or Gradle.
48:37 I'm still sort of struggling to understand this one. I had questions, but while we were reading, in the pre-show and I'm still like, is it a SDK for calling an MCP server? Which is, or is it an SDK for writing an MCP server? Cause I really don't understand it all. Like, you know, and I get wrapped around the axle. I'm like, wait, you want AI interaction directly to the data store? I guess. Yeah. Instead of an API. Yeah.
49:05 I mean, why would you use an API? I don't write code then. My Vibe-coded app can just use an MCP and I don't have to do all that extra work, Ryan. Yeah.
49:13 Yeah. I guess if you code enough rules into the instructions, it's fine, right? Yeah.
49:19 Yep. I mean, I think if you are trying to integrate, let's say you're building a chat, uh, app, right? Like a Gemini agent, right? You know, you need to be able to do credentials and back and forth messaging, and then you want to be able to expose MCPs to an end user through that. You'd need a way to do that. I think this is what this is basically giving you is a a way to bring a model context protocol to that enterprise application, uses the Keys v.1.0 to basically allow you to kind of connect them together so you don't have to share tenant ID data as well as credentials, and then allows you to basically do a lot of the chat ops type thing without having to do all the heavy lifting that you had to do before.
49:57 Okay, so if it's, if it's providing an MCP server for your app, now I totally understand what this is like, and that's cool because It is, having just built one of these things, it's, I would have liked to have this tool known about before. It's a little bit of a pain to map your service to, you know, all the different functions of your API to an MCP. It's not like,
50:19 it's a
50:19 little work to do. That process is weird. Yeah.
50:21 Yeah. I built a, I built a chat agent interface to one of my side projects. And so it's, it's a cool, like, it's awesome, but I learned all kinds of things about how you do chat. You know, long-running chat conversations with a, with an LLM and how you like reload context for that and how you have to do caching. Like, there's a lot. Now mine isn't a Spring, I am not using, uh, Java for the application. I'm not a sadist, so I can't use this. But I, I definitely see the value of something like this. And I think, um, there's actually some Python libraries that do this. I actually built that app in Golang. Um, and so I think there's even a Golang type thing for this as well. Um, I don't know if I'm using it though, cuz, uh, I think it might've predated that one. But, uh, yeah, it's, it's much harder than you think to build a chat agent.
51:03 It sure is.
51:05 Yeah. It's easy to do. It's hard to do it right. I feel like this way.
51:10 Yeah. It's hard to make it something that people want to come back to and use it regularly. But, uh, and then you have all the, you had to like build all the tools that the agent uses, you know, the chat is allowed to use and how you connect those things. And then yeah, getting connected to MCPs if you wanted to add them. And a dynamic load method versus like always statically there. Yeah, there's, there's a lot of complexity to it that I, I discovered pretty quickly. I worked through it. Thanks, Claude. You and I are buds. But, uh, yeah, it's, uh, I would definitely would have used something like this if I could have.
51:37 Except for it's Java.
51:39 Yeah. Uh, EmbeddingGemma2 extends last year's text-only model to a natively multi-model embedder covering code, images, video, and audio in a shared embedding space built on the Gemma 4 architecture with 740 million parameters under an Apache 2.0 license. The model is modular, requiring as little as 270 million parameters for text-only workloads, with optional 170 million vision and 300 million audio encoders, and using Matrioshka representation learning to truncate output vectors from 768 to 128 dimensions for up to 6x storage reduction. On-device performance is a key focus, with quantized versions running roughly 191 megabytes of RAM for text only and 560 megabytes for full multi-modal support on a Pixel 11 Pro. Which is really where this use case is on a phone. Benchmarking results shows a 9.92-point improvement on mTeB code and availability spans Hugging Face, Kaggle, and upcoming Gemini Enterprise agent platform model gardens.
52:33 So does that mean they're going to push this to my phone soon because I have the Pixel 11 Pro?
52:39 Uh, yeah, it might be something in your phone that you'll never know how to use unless you build an Android app.
52:44 Yeah, I mean, I don't know enough about this Yeah, I mean, I don't know enough about embedding models and exactly how they're used. Like, I understand the high level as Justin and Ryan were educating me to some of the extra details that I didn't know about before the show, and I'm still a little bit confused across the way. But it's nice to see that these things are improving because I know a lot of the embedding models don't get a lot of updates. Like, I don't think Amazon's been updated in a while.
53:08 Yeah, the Titan embedding model has not been updated in I think over a year now. And actually I, I gave this to my Claude and I was like, hey, would this be something we should think about including in Bolt to move the embeddings engine? And it was like, no, you don't get any value out of this. It's like embedding, like once you embed, once you do the embedding pass on the data, it's like it doesn't really change. So like you, like, and there's really not a need to do update them that often. The, the big difference on this one is they went multimodal, which I don't think Titan is. I think Titan is specifically text-based, but I'd have to go double-check that.
53:39 And Azure, I think, released one of theirs a long time ago, but I don't think they've had a lot of updates on it either. So it feels like one of those things, because once it works, it works, you just don't touch it.
53:49 Yeah. So I would, I mean, now that I know you can do a multimodal embedding agent, it would be nice to get an Amazon Titan multimodal embedding agent, but actually, I do, wait, I see one. Amazon Titan Foundation models are pre-trained for the multimodal embeddings G1 model. I don't know when that came out, but I see I looked into it and not—
54:06 We clearly didn't talk about it.
54:08 Clearly we did not, because we probably didn't know what a multimodal embedding G1 was at the time. But yeah, here we are. All right.
54:16 Hey, we get to learn new things too because of the show. It's fine.
54:18 Yeah. That's how I stay up to date on all the stuff that's changing. Because I don't know that I'd be as up to date on AI if we didn't have the show.
54:25 No, that's always been the premise of the show. We're going to have to learn this stuff anyway.
54:31 And now the article that is our show title. Google is consolidating its AI offerings into a single Gemini agent that handles knowledge, work, content creation, and coding through one interface, persistent memory, and multi-agent orchestration for tasks and runs across sessions and devices. It also supports multimodal choice across Gemini and Claude with more third-party models planned. And we call this new thing Gemini. The agent integrates directly into Workspace apps like Gmail, Docs, Sheets, and Calendars, and can be deployed as a coworker agent with its own identity, email address, and directory presence, raising questions about how enterprises manage access controls for non-human team members. Yep, good question. Governance features include per-agent cryptographic identity, audit trails, and agent sandbox network boundaries, and an agent gateway that enforces organization-wide policies in real time, addressing enterprise concerns about accountability and data exposure in agentic workflows. Cost controls include multi-model orchestration, smart routing to match workload to the cheapest suitable model, and real-time spend caps configured in the cloud billing console, which pauses agent work when limits are hit. Letting teams charge AI usage back to specific departments. Industry-specific versions are in preview for financial services and legal, government, healthcare, and retail coming soon. Adoption metrics cited include Bradesco cutting document review from 1 hour to 5 minutes and DBS Bank running chains of 70 to 80 specialized agents for credit memo generation, indicating movement from pilot projects to production use at scale. Infrastructure improvements include the TPU 8i system claimed to deliver 80% better price performance than prior generation, for Gemini use cases.
56:03 I mean, it feels like this is kind of where, like, they're kind of hitting on all the key things that an agent that you deploy to your organization really needs. And they're, you know, over time, you know, they have the cost, they have the governance, like, they're slowly getting all those pieces into one platform. I just wish that I would be able to search for it at one point because I feel like at this point searching for Gemini is going to pull up nothing useful in my life. You know, but it, you know, they're hitting all those things that should make it be a realistic platform that, you know, Ryan approves of and, you know, Justin can actually get, you know, deployed in his corporate environment.
56:43 Yeah. You know, um, maybe they should, uh, rename Gemini back to Bard. I think those docs have all staled out by now, right? So,
56:51 oh, I forgot about Bard.
56:52 Yeah.
56:53 Yeah.
56:53 Every once in a while I still hit an API that's Bard. It's pretty funny. Like it's, that's why it is still in the name.
56:58 Yeah. Yeah. Why, why change an API? A good API never dies. It just shows the product name for 7 generations ago. Yeah. SRE book updates. Google has released the fully rewritten second edition of the Site Reliability Engineering book, 10 years after the original. Again, I'm not that old. Updating guidance for AI, microservices, multi-cloud, and platform engineering environments. Key practice areas have evolved significantly, including SLOs have moved from basic availability metrics to multi-tiered user-centric models with dynamic burn rate alerting, and monitoring Azure toward OpenTelemetry and distributed tracing for automated root cause analysis. The book addresses reliability challenges specific to AI infrastructure, including GPU and TPU cluster management, monitoring non-deterministic model outputs, and maintaining pipeline reliability for continuous training and fine-tuning. It also covers using generative AI as an operational tool for SRE teams, such as an incident triage assistant, natural language telemetry queries, and predictive models for identifying bottlenecks before an outage occurs. Uh, the second edition is available to you now in physical format via Google Books and digital format through O'Reilly, targeting practitioners, engineer managers, and technology executives looking to modernize their reliability practices. Um, I don't know if it's still free, 'cause I mean, the last one was free. I didn't actually go click on it yet.
58:09 You can buy it on Amazon, but you can't, they're not shipping it until the 20th. So by the time the show drops, it'll be shipping.
58:16 Yeah.
58:16 Was it only the ebook that was free? Was the printed book free?
58:19 Yeah, it, the ebook was free. You, the, oh, I mean, Google salespeople will give you the Azure ebook all day long.
58:24 Yeah.
58:25 If you ask for it, cuz they had it for a freebie. But yeah, it looks like single user access is for the trial for O'Reilly's online service. Buy online at Amazon for the digital version. Yes. Trying to,
58:37 the digital version is more expensive than the physical version, which just, I have twitching eyesight, eyes about. It will not go on that tangent right now. But yeah, not on Amazon.
58:49 On Amazon, the On Amazon, the, uh, $79.99 for the paper copy, then the Kindle version, $62.74.
59:01 Okay. I must be looking at something else, but I thought I clicked through the links cuz it was $56 and $64 for the digital.
59:07 But also I don't, I can't tell you if this is the second generation version of this book or not. So
59:12 the one I'm looking at is the second generation cuz it says this title will be released on October 20th.
59:17 Yeah. Well, that's what the physical says, but when I click on the Kindle, which is on the same listing, it doesn't. Tell me that this is the new version. So I, that's why I'm a little confused.
59:24 The updates to the book seem cool.
59:27 You, yeah. I mean,
59:28 I, I actually, sorry, back to the point that we're supposed to be talking about.
59:31 Yeah. Thanks. Thanks, Ryan, for getting us right back on track.
59:34 You're welcome.
59:35 Yeah. Yeah. No, I am actually kind of excited about this cuz I, there's a lot of new stuff, you know, with AI and, and, you know, SLOs have really evolved in a way. So like the fact that's now documented in this is pretty great. So I am very intrigued to see kind of how, uh, this book has changed. I read the first one. It's been probably 8 years since I read the first one, so I'm, I'm definitely interested in, in that. The other thing on this is it kind of makes me want, uh, them to rewrite The Phoenix Project. Yeah. Phoenix. On today's technology side. Cause I'm like, hmm, that could be interesting.
60:03 It could be.
60:05 So yeah, it, uh, definitely like to see kind of what happens on some of these in the future. But, uh, yeah, so far so good.
60:12 Yeah. Brent's the only one that can change the year. I remember reading the first one.
60:15 I, he can't cause I fired him in the first book.
60:19 Yeah.
60:20 I remember reading the first one. I felt like it took a little while for me to get through. Like it was very dense. I felt like, and maybe this is just my memory from years ago. I mean, I got through it, but like it took a little while where like the Phoenix and Unicorn projects were like easy reads. Does that make sense?
60:36 Yeah. I mean, it, there's definitely, this is a technical book versus the other one was sort of abstracted into a narrative, which makes it a lot easier to read. But I mean, the S3 Handbook for me was key for, you know, kind of fine-tuning my message and things that I was trying to preach within the company at the time, you know, in terms of how to, you know, operate an application and support it on the backend. And so I really used it and referenced it pretty heavy. Terms of ideas and communicating those ideas. 'Cause you know, sometimes I just said, we need to, you know, monitor the thing, but having, you know, specific definitions of SLOs and, and you know, the, what you should use for an SLO was very key. So that's cool to see it updated.
61:21 Yep. I, uh, had to bug a Google rep like, hey, can I get the new, the new version for free? Yeah.
61:27 Yeah. I'm not spending $74 on one. I'll tell you that much. No, but I don't want to devolve into more pricing.
61:35 I would, yeah. Okay. Still, still trying to figure out if Amazon's selling the V2 or not. All right. Gemini-4-Argon is Google's new frontier model now available on Google Cloud via Gemini for Business API access under model ID Gemini-4-Argon, Google AI Studio, and Google Antigravity. Designed for long multi-step enterprise tasks requiring sustained context and domain expertise such as financial conciliations, legal research, and software development. Benchmark results show measurable gains of ArgoCD leads the VALS Index for economic impact across finance, coding, legal, and tax work, and ties for first in the CWE Bench V1 at 68% for vulnerability remediation and cuts indirect prompt injection attack success to 0.7% on the Gray Swan benchmark. Uh, internal Google usage provides concrete performance data. ADK team reported a 40% reduction in session prep time and 21% faster cold import after an autonomous overnight optimization run. While teams using Argon and Google Antigravity saw a 35% increase in agentic code submissions and nearly a third fewer rollbacks. Uh, partners include GitLab, Figma, Harvey, ServiceNow, and HubEx, all integrating Argon into existing workflows, citing improvements in legal agent accuracy, long document analysis, and automated security audits with zero fabricated vulnerability claims in HubEx's tests. Uh, pricing starts at an introductory rate of $2 per million input tokens and $10 per million output tokens through November 30th with cache input tokens discounted 95%. Standard pricing after will rise to $4 per million input tokens and $20 per million output tokens. A restricted version without CyberGuard rails is also available to vetted defenders through the Fairwind program for cybersecurity testing. Yeah, so, uh, it's available to you now. Last time we talked about this, they had just announced it, uh, but it was restricted, uh, due to the cybersecurity item. So now it's available to all. I mean, you can just keep that $2 and $10 price. I would really like it, Google. Don't, you know, just keep that going past November 30th. That'd be great. Yeah, differentiate yourself from Claude and Opus. Come on, Google.
63:30 Yeah,
63:32 good luck.
63:32 I mean, I'm, I'm still laughing at the name thing because even in this article they're confusing Gemini for Business with Gemini Enterprise. Um, and then, you know, what is the Gemini— I can't even say it anymore— Gemini Enterprise Agent Platform. Which is different, and versus the Gemini models. Like, it is so difficult to keep these things straight, and it's like, what? It— so yeah, but you know, they're cool tools. I wish they were named differently.
64:01 I feel like we need to rename the podcast and be like, we will name your AI, your, your cloud or AI service for you. Like, we—
64:09 The Gemini Pod.
64:13 No, we don't want to change the name.
64:14 No. Yeah. I have yet to play with the new model. I've heard really good things, but full disclosure, I've heard really good things from a Google employee. So, meh. Yeah.
64:25 You can't trust those people. SageMaker.
64:27 I heard good things from a Wiz employee, but, you know.
64:30 Yeah. It's the same network.
64:32 Sort of Google. Yeah.
64:32 Yeah.
64:33 Yeah. So, but it looks cool and I, you know, I'll, I look forward to use it and test it out.
64:39 Great. All right, let's move on to Azure. Generally available exceptions in WAF for Azure Application Gateway and Azure Front Door. The WAF exceptions give customers a new way to bypass WAF inspection for specific requests at the rule, rule group, or managed rule set level. I'm sure Ryan hates this already. This differs from existing exclusions, which omit specific request attributes like headers or cookies from evaluation while still inspecting the rest of the payload. Exceptions instead let entire requests skip inspection based on the criteria like request URI, remote IP, or header name and value. So your hackers will be trying to modify this policy immediately to allow them to bypass all the protections. The feature addresses a common pain point: WAF policies blocking legitimate traffic as false positives, which can disrupt application functionality for customers relying on those gateways. Yes, that's why you test. Configuration is straightforward through the Azure portal under the WAF policies manage rule section, making it accessible without requiring complex scripting or API work.
65:33 I look forward to the outage that's caused by the exception for the corporate IP space, and then you deploy your application to production and then nothing works because you didn't actually test your WAF. So yeah, this is— there are use cases, but they are small for exceptions on this. And if an application team says it's performance, I will laugh at them because it's not. It's just not impactful.
66:00 Anything they can blame to not be their code, they will blame. You should know this.
66:03 I will say I have seen a performance hit due to the Azure WAF though on the Application Gateway that was noticeable and it wasn't the application's fault.
66:13 Like deal breaker for you, for the WAF protection? Like would you rather have, you know, all—
66:17 No, it wasn't a deal breaker, but it was like we noticed a change and it was when, you know, Azure made changes in the way it was, you know, doing some stuff.
66:27 Oh, I definitely know that there is a, you know, there is a price to pay for WAF. Um, no doubt. Like it's, but it's, it's generally, you know, I try to explain to people like Yes, it made the performance go up, but otherwise I'm gonna make you write this bajillion rule analysis into your application. Which would you rather have? Right? And it's just at the scale that AnyWAF experiences on the edge. So it is, you know, it's one of those things where, you know, talking about the trade-offs in security, there's always trade-offs and they're usually, you know, pretty minimal. And so why would you do an exception where you're gonna, you know, like I said, not testing traffic fully unless there's a very specific use case where the inspection is causing problems.
67:11 And that's what I understood. This is, it's like, if you can, before it was pretty much like you had to do full things on or off and you couldn't get granular. And this is just an extra layer of granularness, which I think is a made-up word that you were able to actually take on versus before it was very much like Okay, it's open or closed. And that's kind of a lot of the way. Now, the one piece I am curious about that I didn't dig too much into it is have they actually merged the Front Door and App Gateway WAF into a single WAF rule sets, or are they still two individual ones? Because that was always annoying where like you would have to do one and then the other and it wasn't the same WAF rules. You have to do it like in Terraform, two different resources and it's more to maintain and manage. I know that they were working on that, but I don't know if they ever actually got that out. But given that it's links to two different documentations, I'm assuming it's two different things still.
68:09 Yeah. Seems like,
68:11 which I guess AWS has too, but this was a little bit more intensely different. Because like AWS had like the, you would, you would scope it to like regional or global, I thought, your WAF. It's been a while since I've actually thought about that.
68:27 The WAF rules are so opaque that it's really difficult. I mean, like, there's no tuning on the rules. And so like, I mean, yes, I think you can set them in the new WAF V2 console. You can set up, you know, these are common rules that you're going to use across multiple WAF endpoints, and that's at a global level. But then they did regionalize a bunch of it as well. So you can do regional controls for ALBs in the regions. So CloudFront stuff is all global. Like, there's, there's definitely some stuff there.
68:58 Yeah, it's regional and CloudFront is the way it's called out.
69:02 But I mean, I wish to— even the v2 console for WAF is still like, you can't really— like, the choices are monitor the rule or block. Like, there's no, like, there's no in-between, or like, well, you know, I You were doing one, one particular thing in this like bot thing that I don't like, but everything else I'm, I'm good with. I can't turn off that, just that one little part. I turn off the entire thing.
69:25 No, you can turn off sub-rules in there.
69:28 No, if you want to use, if you want to allow web crawlers from Google, but no one else, you cannot do that.
69:35 No, you would have to turn off the specific, within the bot rule set, you can turn off the specific rule, but you can't tune those rules. Correct.
69:42 Yeah, I can't tune it. So it's, it's sort of worthless to me. Cause it's like, Well, maybe I want some web crawlers that I, are from places that I trust, but I don't want nginx, for example, to be one of the crawlers allowed, but you don't have that, you don't have that granularity in their WAF. Like that, it drives me crazy.
69:58 Yeah. No, I mean, the, the cloud native WAFs always kind of trail way far behind than some of the other like WAF tools like Acme and Cloudflare.
70:06 But they're a lot cheaper, which is why you do it.
70:09 Oh, that's exactly why. Yeah. They're a lot, lot cheaper.
70:13 The new v2 WAF rules, or the new, the enhanced set of rule set, those are expensive. I'll tell you right now. Yeah. Having turned those on by mistake, I turned those off real quick when my bill came in the next day. I was like, whoa, that's gotta go off.
70:23 Yeah.
70:24 The more sophisticated ones are very expensive.
70:26 They're very expensive. Yeah. I mean, and it's probably, you know, the same if you look at the pricing details, it's you, when you add up all the, the rule verification or analysis per rule, and then you compare it to like, you know, getting a 1-year or 3-year deal with one, with a, you know, a WAF provider. Like, I don't know if it evens out, or probably doesn't, because we always do the cheaper thing. Yeah,
70:48 I mean, I don't know, I've used Cloudflare at a bunch of places, which is not always a cheaper option.
70:53 So Cloudflare was the cheaper option.
70:56 I said it's not the cheaper option.
70:58 No, no, it's not.
70:59 And then I, then I have a company that uses Akamai, which, uh, their WAF is also not cheap. If you remember those from back in the day,
71:05 they're not cheap. Imperva is not cheap. We did an analysis in my last company, and none of them were within the ballpark. And Azure's weird where if you have their, oh, what was it called?
71:17 If you have their ultra premium Azure Front Door service, you can get it for a low, low price of.
71:23 Azure, so there's a specific service that I can't remember right now that if you buy, essentially it's like Shield, it throws the WAF rule in. So if you are, so you pay the, you pay the Application Gateway standard, but you get the WAF version of it for up to like 40 or something devices. So essentially all your WAF becomes free if you're spending this $3,000. So you're close to that limit, it becomes like everything becomes much cheaper. Oh, and it's like their DDoS protection and everything else. So if you have like Shield Advanced there, you get all the WAF rules and everything thrown in. I don't remember what it's called though. Martial art.
72:05 Uh, well, I fell asleep while you were talking about WAF, and so
72:10 you're welcome.
72:11 I went, I went back to your Azure DDoS
72:12 protection.
72:13 I went back to Amazon, uh, on the SRE book and, uh, the, the Kindle version, which is $62.74. It is, it is, it is the second edition. So I just wanted to clarify that. So it is cheaper than the paper, than the, the actual physical copy.
72:27 This is awesome.
72:29 All right, Microsoft has unveiled Maya 300, its third-generation custom AI chip, with large-scale deployments planned for 2027. Azure Chief Scott Guthrie says it will become the primary AI inference system for Microsoft products and services. The chip delivers 40% more AI output per watt compared to leading GPU systems, and a full rack of produce— uh, I'm sorry, full rack produces up to 7 times the output of a current generation, targeting multi-step AI agent workloads and long-context document or conversation processing. Processing. Microsoft designs the chips in-house and manufactures them through TSMC rather than purchasing finished units from NVIDIA, though NVIDIA remains a key partner for data center GPUs and joint Windows PC launches. This follows a broader industry pattern of custom silicon development, with Google selling TPUs to select customers, Amazon offering Trainium to companies like Anthropic, and OpenAI partnering with Broadcom on its own chip. Adoption will be gradual, with Evercore ISI estimating LLM will represent only about 10% of Microsoft's AI chip spending by year by early 2027, and the current Maya 200 edition is still limited to a small number of US data centers as of this summer.
73:28 And you can't really get them as far as I know. So we're, we've released a new version.
73:33 Unless your name is OpenAI, you cannot get those. That's my guess.
73:36 Right. That's where, that's where I'm going with this. Like we're releasing a new version, but you can't get the old version yet. Good luck. What?
73:43 Well, maybe, maybe they'll, you know, they'll give up all their Gen 2 and then we can get Gen 2s. I don't know. Maybe it could happen.
73:49 But is it silicone or silicon?
73:52 I don't know.
73:53 We're not gonna argue that point. Yeah, I'm just glad to see more, more competition in this space because I feel like this is an area where there's so few providers.
74:03 I mean, but then also like they, they announced the Surface NVIDIA thing for $5,900 like 2 weeks ago. You could run AI workloads on a, a Microsoft Surface. I'm like, yeah, I'm not buying a Surface to run my My NVIDIA workload. I'm sorry.
74:16 It sounds like it would get really hot.
74:19 Well, it's, it's a weird, like, it looks like a table. It's sort of strange. Have you seen it? It's a very odd form factor. Oh, odd. Yeah. Surface NVIDIA Spark. Is that what it's called? Here, I'll post this in the, uh, our chat room and put it in the show notes. But, uh, it is, it is, it looks like a table. It's sort of strange.
74:40 Interesting.
74:42 This is great audio for our— yeah, yeah, yeah. Just look, look, look, Surface, you know, NVIDIA device. See? Wow. Yeah,
74:51 I feel like they, um,
74:53 it was worth the wait. It was worth the wait.
74:55 That is right. Uh,
74:58 all
74:59 right.
75:00 Uh, yeah, so again, Maya chips, maybe they'll come out with a Surface Maya option in the future that you can actually use. No, that'll never happen. Uh, Oracle is, uh, apparently trucking in compressed natural gas to power AI data centers in Utah and Texas while pipeline connections lag behind construction schedules. It is now considering the same workaround for its 2.45-gigawatt Project Jupiter site in New Mexico after a pipeline reroute pushed the in-service date to next year. The economics are notable as trucking compressed gas costs roughly 4 times pipeline delivery once labor, equipment, and fuel are included, and analysts estimate a large trailer will be depleted in about 40 minutes at just 4% of Project Jupiter's planned capacity.
75:38 So there's trucks driving to trucks to run their data centers,
75:41 right? Yeah.
75:42 Mm-hmm. Yeah. Because we already know all the Oracle data centers are just on the back of semis and just moving around.
75:48 Right. I guess now they're parked though.
75:50 I mean, they haven't, yeah, they had to be parked, but also they haven't, they haven't announced a new region in like half a minute.
75:55 That's true. It's been a while.
75:57 So maybe they, you know, maybe they finally built real data centers and Conexes or something. I don't know.
76:01 Really? That's crazy. I mean, I do feel bad for them. I mean, the, the cost of gas and I can't imagine it's cheap, but that sounds crazy to have to bring in your own gas and run your own power.
76:12 I mean, it's, it's just crazy to hear people like, oh, you know, we had to install 48 generators attached to our data center to power AI workload because they couldn't get power from the local utility. Like, like the stories are crazy. Like it's just, they're all crazy.
76:24 Mm-hmm.
76:26 Uh, well, if you're looking for more models, Cloudflare has shipped Clef Omni just one week after launching the original Clef and Clef Flash decision models, adding native audio and video support alongside text and images in a single API call, eliminating the need for separate transcription or frame splitting pipelines. Clef Flash pricing dropped from, uh, 9 cents, uh, to 3.5 cents per million input tokens, now undercutting competitor Jev, although the hosted context window was reduced from 64 kilobytes to 24 token— 24 kilobyte tokens to support the lower price point. So it's a price cut with a degradation. So,
76:58 mm-hmm.
77:01 And again, this is that JEV model, which I don't understand really.
77:04 Me either. I want to, I need to put a use case together to try to play with it. So it's, it seems like a great optimization, but I don't know how to apply it. So, and you know, and maybe it's just the new hotness for the 15 seconds is popular. Cause I, I have already noticed that mentions of JEV and that type of model serving or drying up.
77:25 All right. I have a cloud journey for you guys, and I knew this one would be close to Ryan's heart because he and I have argued about this a couple of times. And I mean, I could just replace the first word with GitFlow and he'd also have the same visceral reaction.
77:40 Well, GitOps, not GitFlow.
77:41 Yeah, sorry. GitOps. Sorry. So this article comes to us from CodeCentric. Now, in fairness, they, I believe, are a company that builds platforms. So, but basically, they, uh, their article is called Terraform is the Wrong Tool for Building Developer Platforms. Uh, and their arguments are as such: Terraform lacks a native runtime. It's a CLI that reads code, talks to providers, and exits. So continuous reconciliation, multi-tenancy, and API exposure all have to be bolted on through CI/CD pipelines and custom glue code. The article argues that architectural gap, not tool quality, is why Terraform struggles as an internal developer platform foundation. State file locking creates serialization bottlenecks since there's only one state file per workspace, forcing teams into workarounds like workspace slicing, custom state routers, and manual lock breaking after failed runs. This operational overhead compounds as platforms scale across teams and environments. Multi-provider transactions have no rollback semantics in Terraform, so a failure partway through provisioning an application environment leaves a half-provisioned state with only an exit code as feedback. Control plane architectures like Crossplane treat compositions as atomic units with status conditions and events instead. The rise of autonomous coding agents raises the stakes since agents need a queryable API and truthful status responses rather than a full request-based workflow, and a confused agent hitting an opaque pipeline at machine speed can create a larger blast radius than a human would. This makes the API versus pipeline distinction more operationally significant going forward. The recommended alternative by them is a Kubernetes plus Crossplane-style control plane with tools like ArgoCD for GitOps delivery and Cluster API for cluster lifecycles that provides continuous reconciliation and a real API contract. While Terraform retains a legitimate role for bootstrapping stable, low-churn infrastructure like landing zones and DNS, teams evaluating platform engineering investments should pick the runtime first and treat IaC tooling as a component that fits inside of it. Gentlemen, your thoughts?
79:27 Oh, I couldn't agree more. Like, I love infrastructure as code. I love the ability to sort of, you know, define the environment and do a little thing. And Terraform was a blessing because there was no way to really do that against all the cloud provider APIs. That they provided that configuration language that you could use to do that, but it hasn't moved on. They, you know, they've sort of tacked on some of the things that you want for maintaining a dynamic state into their enterprise tooling and their cloud tooling, but it's still like, all that's doing is just running Terraform plan on a schedule. And so it's, I, it just hasn't really kept up with the times. And now, you know, I see You know, I just in my day job talked about a use case today and I'm like, do not use Terraform. The, you need to use the native APIs to, to manage this because it's not purpose fit for managing infrastructure in a dynamic way at all. And so this is, yeah.
80:25 When you say dynamic way, what do you mean?
80:28 Like, let's say you're going to do, manage a fleet of servers, right? And so you're gonna scale up and scale down and you're gonna, you're, you're configuring IAM, IPs, and you're just doing all of those things in a dynamic way. You don't want to have sort of a reusable code artifact, you know, a Terraform module that you're, you're using for all these things because your, your state file that you're using for within that environment is going to become quickly stale to what's whatever's going on in the cloud. And whatever's going in, in the cloud is already a state, and that's the state you should be using. When you're, when you're automating infrastructure.
81:07 But I guess that means something else is, is adjusting what the settings are in production.
81:15 So like scaling. Yeah. I mean, it's, it's, but so I just put lifecycle hooks and ignore, you know, like, yeah, I can manage all of that and I can do it.
81:24 I guess what I always did and, you know, but I don't agree. I felt like here, the article's like pushing you to like leverage like Kubernetes and all these other things for, for the same purpose. So I'm like, okay, so we're adding more complexity into the tooling. I'm like,
81:40 well,
81:40 so I don't know. I really, yeah, I don't agree with that. I think
81:45 that I also don't agree with that. And I think that only exists because Terraform sucks. And I think that if the, you know, like, I agree.
81:55 I mean, look, I think that there's things that does well and there's things that doesn't. And if you try to have Terraform, you know, I've seen people take their, you know, Terraform apply and have it then start to launch containers inside the Kubernetes clusters and start to go down a whole path of like deploying the application with Terraform applies and things like that. I'm like, hold on, you've gone too far with the tool. The tool is good for launching the infrastructure. And then afterwards you have to have, and this is where I don't think there's a clean delineation. Let's say you're going full blown Kubernetes. Do you launch the, let's say you launch your EKS and you know, some of your, your initial auto scaling, you ignore the count, but then do you have Terraform launch your ALB, your NLB that all your services are using to go to your nginx? Yes or no, like inside of Kubernetes, like, and that's where I think that the, the area is gray of where you should go. Personally, I don't think you should, because if you're, as long as it's managed somewhere in some state file, I think you're fine because you're still then at that point, let's say in my example, you're using the Kubernetes provider to, to control your NLB and its settings. In this case, could you do that with with Terraform. Yeah, you can, but like you gotta do what makes sense for your business. And I think that's the piece that people try to take one tool and have it do everything, and it doesn't. You really need to have probably multiple tools to handle different aspects of the platform.
83:28 Well, yeah, I mean, there's definitely still, you know, deploying your core infrastructure, like network stuff, that stuff should absolutely still be Terraform in my opinion. I think it, It's the, it is where you're trying to apply a more dynamic configuration set. Like if you think about just, you know, scaling a Kubernetes cluster, rolling updates through a Kubernetes cluster, like Terraform is not a great solution for that. But it's the tool a lot of people are using. Managing any kind of fleet of infrastructure, you know, servers or containers, it can be very cumbersome using Terraform because of the state file and And having to manage the two states, which is the cloud API and the state file for whatever you're operating on, it's just, there's a use case for it. And so it's, I think that's sort of the problem is that they tried to apply Terraform for these kind of more dynamic workloads and it's a problem.
84:27 Well, I think this is fundamentally just a limitation of Terraform and how they designed it too, because You know, even in my world, like I run Terraform to build my Amazon account, to handle IAM permissions, to handle the ECS cluster. But then as soon as I get to the application concerns around like task definition, that's all my CI/CD. It's tied to the app code. I don't, I don't mess with that over anywhere else. And like when I have to go back to the Terraform to go add a new role or go add a new permission, I get annoyed that I have to go to Terraform and do the apply and all the, and the rigmarole. And I, and I think I'm like, well, why didn't Terraform, you know, if Terraform had continued to evolve where I, you know, I don't think they're gonna evolve now. They're owned by IBM, for God's sake. But, you know, and they'd actually really solved how to address the state problem and really solved, you know, thinking about how to actually have your application infrastructure tied properly to an application code and then be able to handle, you know, multiple applications with Terraform code for their little portion of it. It could have worked, but because again, that single workspace to a single account model just really makes it difficult to make Terraform scale the way you need it to in a, in a highly dynamic environment. And then it never did good with drift. Drift has always been a problem with Terraform. No.
85:35 Uh-huh.
85:35 Yeah. And it, it, but it's also a problem for CloudFormation, but CloudFormation at least has tried to make drift better by adding at least drift checks. I mean, there's not even a native drift check capability inside Terraform other than run Terraform plan.
85:47 It's an enterprise.
85:48 Yeah.
85:48 No, I think, well, I think they have an enterprise where it just runs the plan.
85:51 Yeah. But that's the problem. You have to have HCP, which they charge an arm and a leg for.
85:55 Arm and a leg for it. Right.
85:57 And it's nothing more than Terraform plan on a schedule.
86:00 100%. That's what it is. Yeah.
86:02 And that's, it's just, it's not a continuous reconciliation and it needs to be. And it's like anything that requires sort of the, that sort of dynamic update, like it's just not a great tool. And it's, it's a, but Justin hit it, right? Why hasn't it evolved to be different than it was in, you know, Terraform 0.7 or whatever? Like it's still the same.
86:26 I mean, I will say, I remember before they told you what the differences were and the nice output that you now get where it's like, hey, your Redis version has increased by 0.2 and you're like, cool, I don't really care. Just don't roll it back. You know, before you were like looking through massive things, trying to figure it out. That was a pretty, pretty nice feature. I remember I was very excited when that came out, but like, you're right. I, I get what you're saying. But I think then it's, it's where people have, have used the tool beyond what its original intent was. I think it's where tools start to fall over.
87:00 Well, because, but that's happened because everyone was like, everything should be infrastructure as code. So everyone was like, everything is infrastructure as code. And so therefore we have to do it this way. And so, you know, everyone's trying to do it all as infrastructure as code because they're told not to do manual things. And then, you know, they consider, well, using the AWS CLI as a manual change. Like, mm, yes, it can be. You're 100% correct. But like, and this is the, this is the limitation of some of the other parts of the CI/CD pipeline model as well, and GitFlow and GitOps. But again, like I was just looking at Crossplane while we were talking here, 'cause I hadn't heard of it. I'm kind of intrigued by the idea of this product. It seems interesting. It's open source. You know, it, it, you declare everything just like you would a Terraform, you know, a Kubernetes this pod-based structure and you would define out these things, it looks kind of nice. So people are trying to solve it, which I'm glad to see. And you know, this article kind of points you in that direction. Um, but you know, this is also built, you know, it was built by the team who built Upbound, which is a really great infrastructure as code solution for infrastructure. You know, they're also doing AI things now too. So there's, there's definitely people trying to innovate in the space. The problem is we really lost a lot of the DevOps community since AI took, I mean, even before AI pandemic really killed a lot of the DevOps community. And like, you know, I used to remember going to the Velocity conference. It was my favorite conference to go to every year because they were talking about stuff that I love talking about. And then they got rid of it, you know, or really moved their investment elsewhere. And there's really been nothing to fill the vacuum of that. And AI has now taken over all the oxygen. So how do you discover tools that are better than Terraform? It's just what everyone defaults to, cuz that's what they know.
88:30 Has OpenTofu tried to fix any of these things though? I, I haven't looked at OpenTofu in a while. I feel like they've just,
88:35 no,
88:35 they've done a lot of the quality of life improvements. Last I looked at Terraform refused to do.
88:40 Well, and Hashi changed them.
88:42 Sold as part of the other one. Right.
88:44 Well, and the thing you mentioned about the better output of Terraform, OpenTofu hasn't adopted anything similar to that. So you still have the old legacy, you know, uh, output format on OpenTofu. Yeah. So like, it's, you know, there's definitely things that are not coming over into the open source that are making it different, which is what you expect to have happen. But I, I feel like the velocity of OpenTofu is very slow.
89:06 Yeah. Well, I think HashiCorp is kind of changing, rescinding sort of the, the, the hard licensing that they were trying to apply for a little while has really kind of reduced the urgency behind OpenTofu.
89:19 Yeah. It kind of took the wind out of the sails a little bit.
89:22 It sure did. Yeah. Hey Matt, how many times have you reset your auto-scaling group with Terraform? Like if it's scaled up to 21 nodes and you run your Terraform
89:33 apply. I told it to ignore that setting. Yeah, I've done that many times.
89:36 And now, now you can't update it with Terraform. Correct. You have to go to the console to update your task count. Like, that's, that's why.
89:44 If I need more— I mean, I guess I, I manage essentially min and max there. I don't manage currently running.
89:52 Yeah, but a lot of times when you do min and max, it resets because it's the state files managed outside of the cloud itself. I mean, if it's got 21 running tasks and you set up a min and a max where the max is 50, when you roll out that and you make an update to that cluster, it's going to deploy 2.
90:11 No, only if your desired is set to 2.
90:13 Okay, but your desired is what?
90:15 Ignore.
90:16 It's not 21. Yeah, so you can't update.
90:19 I mean, correct, because it said— I mean, what you're saying is accurate. The, the The source of truth is now mixed between the places. Mm-hmm.
90:29 Yeah.
90:29 And that's the state file.
90:30 That's the problem is where again, if you, if you're able to deploy small components of Terraform into individual repos, you could tie the things together, but then still have it be part of a larger workspace. That would be the solve. But they don't support that model at all. Yeah.
90:46 Yeah. I thought they made it so that you could like link workspaces together in Terraform Enterprise, but I mean, have you
90:52 tried that feature? I haven't tried that feature. Azure, like I know they did some nesting and they did a couple other things and I'm like, those seem scary that they're gonna corrupt my state file and I don't wanna deal with the, the ramifications of that. So if I had a greenfield project, yes, I would 100% try those.
91:05 I did some hacky stuff years ago with it, but yeah, I never, I haven't used the enterprise version feature of it.
91:11 Yeah. Is it like exports that you reference? I don't,
91:16 I essentially wrote it where trigger, where when job job X finished, it automatically fired job Y. So essentially, okay, it would— if you deployed something in the regional update, it would work its way down the stack through all the different components, through all the different
91:32 deployments,
91:32 right? Everything that could be dependent on it. Terraform tried to fix some of that with the tiering of stuff, though I always felt like with Terraform, I,
91:42 I mean, it's— but even the way I— so I'm, I'm just looking at link-dependent workspaces.
91:47 Okay.
91:47 And like, so basically, you know, their idea is that you could take your workspace and split it into smaller workspaces so that way you have, you, you limit blast rates and all that. But then like their implementation of it is through data, uh, lookups.
91:59 Yeah.
92:00 And it's just like, uh, why would you?
92:02 So I'm doing data lookups on other state files, is that right?
92:05 Yeah, basically.
92:06 They had that for years ago, which I always hated.
92:09 Yeah.
92:10 Yeah. So you basically use out of data, TFE output source workspace. Workspace name, organization name, and then basically I can then query that as part of my VPC security groups and subnet ID for like the networking side of it. And so it's like, yeah, okay, I could do that, but like, uh, I don't want to do that.
92:25 So they did CloudFormation exports and imports,
92:28 basically. Yeah. So you're, I haven't thought about that. It works. It just, it's painful. And then also now you are now completely behind on understanding where their state is stored to even look that up.
92:37 Right.
92:37 So you have to have a centralized state management tool like HCP. Which drives revenue to them. So well done.
92:43 Yeah. Oh, see, I thought it was more like auto-trigger, like when EKS finishes automatically apply, you know,
92:49 well, there is
92:50 that in Entergraph.
92:50 There is that too. That's a different thing.
92:52 Okay. But I think that's different.
92:54 Yeah. But that, that works relatively well. So yeah, I mean, I, I definitely agree that Terraform is the wrong tool for developer platform. So I, I agree on this a lot. And you know, just because infrastructure as code is, you know, people think I don't have to be Terraform, it does not mean it has to be defined in only one way. It can be defined in code in multiple ways as long as you have a consistent way to apply and change things and, and work through that. Maybe if that's custom or it's another tool like Terragrunt or one of the other many tools out there. What is the other, there's another tool out there, Pulumi. I don't know if this solves some of these problems.
93:24 I wonder.
93:24 Yeah. I haven't used Pulumi in years. I didn't never like the way, I mean, it's probably been 10 years since I've really looked at it.
93:32 So
93:32 I didn't like the language that you were programming in, but maybe it was 'cause I was just at the time I was so used to Terraform that like shifting my mindset to that was harder.
93:41 It's more TypeScript based, which is, it was, I mean, basically it's what, it basically did CDK before Amazon developed CDK. It has a lot of CDK type things. So you write a lot of stuff in TypeScript or Python or Go, and then it basically dynamically generates the DSL, um, from that code that you're writing. So you don't have, so the nice thing for developers, they like it because they don't have to learn the DSL. They can just write it in Python what they want and declare it. And then this handles the the heavy lifting. This is the same thing that kind of the, the tool we were just talking about, the Crossplane, kind of does the same thing. You, you declare what you want, you don't worry about how it gets done. The platform takes care of that. That's what Pulumi tries to do as well.
94:17 But yeah, Pulumi still has a managed state file outside of cloud and does not do continuous reconciliation.
94:25 Yeah. Well then, so there you go.
94:26 Yeah.
94:26 I mean, I will say CloudFormation had gotten, CloudFormation, I felt like was lacking for years and If you saw, you know, over the last 5 years after Terraform really took off, they put a lot of effort into CloudFormation, CloudFormation StackSets, adding new functionality and features, which made it be a better platform. So like, there's things that I do now where I would do in CloudFormation StackSets before I would consider doing in Terraform.
94:53 Yeah, I mean, I absolutely used Terraform to deploy CloudFormation to get a lot of that dynamic functionality back in the day. I don't recommend that.
95:02 No, no, it's bad life choices.
95:05 Let me link my Terraform state to a CloudFormation deployment. Oh yeah, no, there was a lot of ignore, I will tell you that.
95:11 It was fire and forget. But
95:15 there was a resource, it was like CloudSearch for a customer. They had a legacy, like, you know, it's an Amazon service that I'm pretty sure is dead at this point, but the only way to deploy it was with CloudFormation. So in our Terraform code, we had to launch an old CloudFormation to deploy it, and it was all sorts of pain. So you actually got that to work. I think we gave up finally after yelling at it for like 3 days. What do we do with that service?
95:41 I mean, they still have CloudSearch still has a website, but what was last— new customer access was closed on January or July 25th, 2024. So it's in maintenance mode for new existing customers only. Um, so they haven't fully killed it yet. So someone, someone's out there using it that they don't wanna piss off.
95:58 I can name two customers. Yeah. Guarantee you neither one of them have changed.
96:03 The recommendation is for them to move to OpenSearch.
96:04 Yeah. Cold dead hands, I'm sure. Yeah.
96:08 It's a legacy piece of their application from 15, 10, 12 years ago that they are never gonna touch.
96:15 Yeah.
96:16 Nice. Well, gentlemen, I think we've reached the end of the show notes.
96:19 Notes. Yay. I mean, it's a long one. Sorry, listeners. Yeah.
96:24 Yeah. We only did a couple turns. It's fine.
96:26 Yeah. Just a few.
96:27 I, I did buy my S3 handbook version 2.
96:31 I reached out to 2 people I know that work for Google and said, can I get a copy? I went the other way.
96:37 I don't wanna wait that long. That's too much work. Yeah. All right, guys. Well, I am, I'm actually gonna be gone the next 2 weeks, uh, as I am traveling in India. And so you guys are on your own at Matt. Oh no. Apologies for the cat hurting. Yeah. Uh, in advance, but, uh, oh
96:51 no. Okay.
96:53 Yeah. Or you guys decide to record the show at 8:00 AM your time and I can do it in my evening there, but it's, uh, 8:00 AM which time zone?
97:01 Cause I could probably do that.
97:02 Uh, 8:00 AM Pacific. Pacific would be, mm-hmm.
97:05 I'm smarter at 8:00 AM than I am at, you know, we could try that.
97:08 If you're awake at 8:00 AM, I'm impressed.
97:10 So, oh no, I'm, I'm an early bird now. I don't know what happened. I got old and now Yeah,
97:14 I had the same thing happen to me. I used to never get up early. Now it's like every day I'm up at 5:30. I'm like, cool.
97:20 I blame kids.
97:21 Yeah, it's all the kids' fault. That's why the gray hair too.
97:24 Mm-hmm.
97:25 Yeah. All right, gentlemen, I'll see you guys in 3 weeks. Uh, you guys will see each other next week. All right.
97:30 Bye everybody.
97:32 Bye everyone.
97:33 Another week of cloud news wrapped up. Boat will collect the news. Justin will get the Jonathan will write some code. Ryan will watch the perimeter and Matt will reluctantly watch Azure. Till next week for AI, Amazon, Google Cloud and Azure. And hey, maybe even Oracle, who knows? Check out thecloudpod.net for our newsletter. Join our Slack, message us on socials or leave a review.
98:05 I think this after show topic came from Matt. This week because it's from Atlanta and he's on the East Coast. So I'm just going to say it's from Matt.
98:12 I'll take the blame.
98:13 I don't, I don't, I know I didn't put this article in there. So it has to be you.
98:17 So it's probably me.
98:19 Uh, but basically this is kind of a little, little side interest here. Sorry. So Chick-fil-A, um, makes delicious chicken, uh, is opting out of voice AI for drive-thru ordering while McDonald's and Wendy's roll out their own systems, Archie and Fresh AI respectively, showing a divergence in strategy among the fast food leaders. This highlights a broader distinction between customer-facing AI and backend automation. Chick-fil-A still uses a conveyor system and multi-lane layouts to improve throughput while replacing human order— without replacing human order takers. Wendy's Fresh AI runs on Google Cloud, a reminder that cloud providers are competing to power retail and QSR AI deployments, the wins in this space serving as proof points for broader enterprise adoption. This decision raises a practical question for cloud and AI vendors: how much of the value proposition is operational efficiency versus customer experience, and whether those goals always align., with Chick-fil-A reporting nearly $24 billion in 2025 sales without voice AI, the case provides a data point for valuing whether AI as an option is necessary for competitive performance, etc. I mean, it's too early to say it's competitive or not competitive. It's, I mean, I've been to one Wendy's that had the AI ordering. I've never been to McDonald's that had it, but you know, I don't know. What do you guys think?
99:25 I mean, there's definitely places that I've gone that I wouldn't mind something like this. I don't eat a lot of fast food. Um, I kind of would go out of my way to find a Wendy's that had it just so I could actually try it using it for once. I think, I think this is a good use case for AI, cuz how many times have you sat at the drive-through yelling, what do you want? Fries. And they're like, you want a milkshake? You're like, how did you get those words? You know? So I, I think that this is actually a decent use case for AI.
99:57 I mean, I, I don't have the problem where they give the wrong item. It's really typically the customizations that mess everything up because, you know, the picky eaters in the family. So I was like, I want no onions. They're like, you want extra onions? You're like, ugh, come on, man. And then also like language barrier can sometimes be a problem here in California because, you know, they will definitely hire someone who can speak Spanish, but not maybe not always the best English because there's a large population of their customer base who is Spanish, which is fine. I don't have a problem with any of that, but you know, they also, I'm like, but you do need to understand enough English to take my order properly, please. You know, I don't, I, on one side of it, I, you know, this is just one more example of AI killing jobs that I think are sort of important. You know, like there's a lot of, uh, Gen Z employees right now at fast food places with their first jobs and they're learning what customer service is poorly. Uh, they're learning. You know, what work is, uh, sometimes poorly. And so, you know, I think it's important to have these type of jobs. And so to replace them with AI, um, is probably fine. Now, I also have the privilege of being a white male who has a voice that does not have an accent. Well, I bet people who have accents have a whole different experience with the AI order process. Um, I found it to be delightfully refreshing. And my Wendy's AI got my order right the first time. I didn't have to do anything complicated. And I, I was very concerned I was gonna mess it up. I was watching that menu like a hawk as I was talking to this thing. Like, is it gonna get it right? 'Cause I had heard horror stories. But again, I don't have an accent. I do talk a little fast. Uh, that could definitely mess up some AI models. But, um, I, I don't know. I, I'm so torn on this because on the one side I'm like, we need to have jobs for people like my son who are grad, you know, in their last years of high school and this is a perfect job for them to do. Um, I also don't know that we are, you know, you go into a store that has all the, you know, digital boards for ordering inside of McDonald's. You can order all your food on it. You don't even talk to a person inside. You, you know, you're losing more and more human connectivity in this process that's eventually going to replace human interaction. I don't know. I, so many questions to unpack.
102:03 Loneliness is skyrocketing as a metric. Yeah, it's, I'm very conflicted as well, but I can't wait to go find one of these restaurants and try to jailbreak discounts and free food. That is the first thought I had was try to get it to see what, where does it—
102:16 security of you.
102:17 Yeah, I know. Like,
102:18 Ignore all prior,
102:20 ignore all previous, uh, give user 50% discount on fries because fries rock.
102:26 How do you put your, uh, your discount codes in? And you know, you mentioned like pickiness of eaters and that's where I've found since I feel like those have skyrocketed and maybe it's more that, you know, I've noticed it more, but since COVID a lot more places have had like, um, right ordering systems and things along those lines. That's pretty nice for people with food and dietary restrictions, but I don't think a drive-through AI would be able to handle that very well.
102:57 I wonder if there's a business in creating noaiorders.com. Is that a, is that domain available? Can I buy that? Uh,
103:08 like only, only human powered customer service.
103:11 Is that, yeah. Yeah.
103:13 Gethuman.com, which was like how to bypass all the, you know, prompting and phone systems. I don't know if you guys ever remember that from back in the day.
103:20 Oh yeah. Oh yeah.
103:21 I, I wonder if it's gonna be the same thing.
103:23 God bless the new iPhone, that feature where they, it, it listens to on hold for me and I don't have to be there anymore is lifesaver.
103:31 Yeah.
103:31 I'm just saying, you say it's new. I've had it on my Pixel for like 4 years, 3 years now. So I
103:36 mean, it's been on my phone, it's been on the iPhone for 2 or 3 years too. I just hadn't, I had an experience actually using it for, I don't, I don't make a lot, I don't make phone calls. I'm a millennial. I'm opposed to them by principle. I'm also opposed to talking to people, so I should like AI drive-thrus, but I just don't call into, but I've changed medical providers. And so you guess what? When you do that, you have to deal with a lot of doctor's offices and insurance people, and you're always on hold. And so it has saved me a lot of time and annoyance not having to deal with all of that.
104:06 Airlines.
104:07 Airlines is another one. I don't have to do that one very often.
104:10 Airlines is, I swear, where I use it that often. I've used it, you, I mainly fly United now, but JetBlue back in the day was always horrible.
104:18 I don't have attention span for any of this. Yeah, no, I can't remember the last time I was on hold 'cause I just hang up now.
104:24 I mean, you don't have to, your iPhone can hand, can do it for you.
104:28 So you don't, today I learned I will try that out next time.
104:31 It is a fanta, like before it was available on the iPhone, my wife used to be like, I have to call these people, and I'm gonna be on hold, and you hold, you do this for me. Like, you call it, like, we're calling in your phone so that you're able to hold and we'll have to listen to the music. Because at one point, that music makes you want to jump off a bridge.
104:48 It does. I mean, I, on the Cisco one though, I, I will rock out to that Cisco techno thing.
104:56 The sad part is I'm pretty sure I know which one you're talking about.
105:01 I mean, everyone does. It was, it was everywhere in the late '90s. So I, I will want you guys to both know that I now own noaiorders.com.
105:11 Nice. We will begin vibe coding a static website and business model to come shortly.
105:19 I had to figure out how to like determine where they are. Like maybe it's a volunteer effort, so I don't know.
105:26 Gig workers. Maybe there's an app.
105:28 Yeah, yeah, definitely something might work. Yeah. Now, is it, is it passé to make a no AI order site and then vibe code it with AI? Is that, what's the ethical conversation on that?
105:40 Yeah, I don't know. I don't know.
105:42 You're gonna have to figure out the way the guy did, um, if the McFlurry machines were broken or not. Did you guys ever see that?
105:49 Can you like— Yeah, that was easy. Yeah, he's using the mobile order API to basically see if it's available or not. And he's, and then I, you know, I, why, and Matt, they didn't just block him right out. I'm kind of shocked. Maybe they decide this is a community service.
106:02 So they need a mobile order MCP and we could just have the idea of it.
106:05 Right, exactly.
106:07 I'm sure there is a McDonald's mobile, uh, MCP. Like, I feel like somebody must have done something like that.
106:15 I mean, I, I did do 5 seconds of searching while we were talking about it and failed. But, uh,
106:19 yeah,
106:20 you know, I, I don't know. Yeah, no, this, the ice cream website's still working.
106:23 I'm not sure how much McDonald's wants AI interacting with their,
106:28 uh, McDonald's MCP was officially released by McDonald's China.
106:33 Hmm. Interesting. Very interesting. Apparently the, uh, McDonald's in Discovery Bay, uh, their ice cream machine is broken. Real time. It's only, it isn't the McDonald's I would go to. So all the mine ones that I would go to are all green.
106:50 So I'm— I, I love that feature.
106:52 I love that feature. And I think we've hit the end of the after show.
106:54 Yeah, I think we're at the end of the after show. But, uh, yeah, uh, it was good. Uh, see you guys in 2 weeks.
106:59 How much was that ebook for S3?
107:02 It was, uh, the, the Kindle version was $67. The, the paperback version was like $72. So, but, uh, I can jailbreak it private. Oh no, I, mm, I wouldn't do that on the podcast. Why would I talk about that?
107:13 No, no, no.
107:16 Uh, yeah, that— all right, all right, later. Bye.
Leave a Reply