324: Clippy’s Revenge: The AI Assistant That Actually Works - Sort Of

[00:00:00] Speaker A: Foreign. [00:00:08] Speaker B: Where the forecast is always cloudy. We talk weekly about all things aws, GCP and Azure. [00:00:14] Speaker C: We are your hosts, Justin, Jonathan, Ryan and Matthew. [00:00:18] Speaker A: Episode 324 recorded for September 30, 2025. Clippy's Revenge. The AI assistant. That actually works. [00:00:26] Speaker D: Sort of. [00:00:27] Speaker A: Maybe not sure. Good evening, Ryan, Matt and Jonathan. How you guys doing? [00:00:32] Speaker C: Hey yo. [00:00:33] Speaker A: Doing all right? [00:00:34] Speaker D: Doing well. How about you? [00:00:35] Speaker A: Well, I'm doing great. Cause I am going on vacation in two days and that is unfortunately mean that you guys are in charge of recording next week without me, which is always a crapshoot. But I. I have confidence you guys have delivered every time for the last least two years. So this joke is getting kind of old, but still 50, 50 shots gonna happen. So. Yeah, I give it 60%. [00:00:56] Speaker C: You just jinxed it too, by the way. [00:00:59] Speaker A: I probably did. I probably did. But the nice thing is I brought Madden. And Matt's more responsible than both of you too. So Matt rallies you guys together. See, that's. That was the trick. That is true. When Peter was here, Peter, there was no hope because we could barely get Peter on the show. So now I have Matt, Matt's number two. I'm getting the show recorded for the week. [00:01:20] Speaker D: Yeah, but now that we have Bold Bot, it's actually much easier because the problem was before we had. Which was a pain. But you've automated that, you know, so we're in a much better spot. [00:01:30] Speaker A: That's true. You still get the articles to the bolt bot, but yes, you can do that part. All right, well follow up. If you remember, a year ago Google filed a formal complaint with the European Commission about Microsoft's anti competitive cloud licensing practices. So the 400% price markup Microsoft imposes on customers who move Windows Server workloads to non Azure clouds. And Google decided that they need to write a blog post about what's happened a year later. And I can tell you as a customer, nothing. Nothing has happened. [00:02:00] Speaker D: See, you should be on Azure, Justin. [00:02:02] Speaker A: Right. But apparently the UK Competition and Markets Authority found that restrictive licensing costs UK cloud customers 500 million pounds annually due to lack of competition, while US government agencies overspend by 750 million yearly because of Microsoft's licensing tactics. Microsoft recently disclosed that forcing software customers to use Azure as one of three pillars driving its growth and it's implementing a new licensing change is preventing managed service providers from hosting certain workloads on Azure. Competitors multiple regulators globally include South Africa and the US FTC are now investigating Microsoft's cloud licensing practices. The CMA finding that Azure has gained customers at 2 to 3x the rate of competitors since implementing restrictive terms. A European center for National Political Economy study Stress ending restrictive licensing could unlock £1.2 trillion in additional EU GDP by 2030 and generate £450 billion annually in fiscal savings and productivity gains. Yes, bring on the Microsoft savings. Unfortunately, in our political climate, you just bribe the current administration, Microsoft will get away with it, so. But maybe the rest of the world will get some reprieve in costs. [00:03:05] Speaker C: Yeah, it's always such a big bill and big true up and it's, you know, like it cloud computing originally, like it was nice, you know, because if you were using, you know, managed ami, then it was sort of built in that. But then now you don't really have any option to get savings and it just gets problematic from there. [00:03:23] Speaker A: Well, I mean Microsoft cracked down on bring your own licensing, which was a huge loss for most of us. But yeah, it's, it's painful out there if you need to learn Windows workloads, which is why you just boot to Linux. Solves all your problems. [00:03:38] Speaker B: All right, well, I'd feel happy about these, these complaints that Google are making if they actually reciprocated the, the deals they make for their customers in the EU in the us you know, so, so when they're giving egress away for free for EU customers, but they're not doing that for the U.S. so yeah. [00:03:56] Speaker A: It would be nice if they did that. [00:03:57] Speaker B: Yeah, get off your high horse a little bit. Maybe Google and like, yeah, but it's. [00:04:03] Speaker D: So easy to take shots at Microsoft because they own the entire stack, so they can easily just charge what they want, so they're a little bit of an easy target. [00:04:12] Speaker A: I mean, I'm also really tired of the whole like we added AI to our products, so now we're going to add money to our licensing costs. Like that's really annoying too. And like, I don't want your AI though. But nope, you have to pay for it anyways. I'm waiting for our Riverside to do that to us because I keep adding more of my AI features and I'm like, when are they going to charge me more for these? Know what's coming. We love Riverside though. Please don't up our prices. [00:04:37] Speaker B: Do they have the AI podcast host option yet? [00:04:42] Speaker A: They have something close, but like they have a bunch of really cool AI editing features. Not that we use them because we don't have time for that, but if Elliot ever said to him, screw you guys, we have options. Not that I want to do any of those. So Elliot, please don't go away. [00:04:57] Speaker D: Please don't do that. We barely get it out of this. [00:05:02] Speaker A: Same thing with Heather on the show side, I'm like, I don't want to use the AI. We do it in a pinch when I need to. Or for some of the TCB talks recently we've done it because I just didn't have time or the budget to pay Heather. But you know, he's like, it's not as good as when Heather does it. [00:05:18] Speaker D: So I much prefer humans are still useful. [00:05:22] Speaker A: All right. AI is how Emma makes money this week with Vibe working, Microsoft's introducing Agent mode for Office apps and Office Agent in copilot chat, leveraging OpenAI's latest reasoning models and anthropic models to enable multi step iterative AI workflows for document creation. This represents a shift from single prompt AI assistance to conversational agentic productivity where AI can evaluate results, fix issues, and iterate until outcomes are verified. Agent mode in Excel democratizes expert level spreadsheet capability by enabling AI to speak Excel natively handling complex formulas, data visualizations and financial analyst tasks. System achieved notable performance on spreadsheet bench benchmarks and can execute prompts like creating financial reports, loan calculators and budget trackers with full validation steps. The Agent mode in Word transforms document creation to interactive dialogue where copilot drafts content, suggests refinements and ask clarifying questions while maintaining words native formatting and this enables faster iteration on complex documents like monthly reports and product updates through conversational prompts rather than your manual editing. That's a good way to get AI slop. By the way, workslop we talked about last week and then finally the one that I'm actually most excited about, Office Agent and Copilot Chat creates complete PowerPoint presentations and Word documents through a three step process, clarifying your intent, conducting web based research with reasoning capabilities, and producing quality chat content using code generation. I don't really care about the word part of it, but the PowerPoint part and I'm excited about so here. Here's what I here's what I want. Make me pretty slides go. [00:06:47] Speaker D: Yeah, the old copilot in PowerPoint prior. [00:06:50] Speaker A: To this was terrible, terrible, terrible. [00:06:53] Speaker D: I tried it once. It reformatted everything and like the words were behind pictures. I was like just undo. Undo quickly. [00:07:00] Speaker A: Yeah, if you only get access though if you have Microsoft's Frontier program for Microsoft 365 copilot licensed or personal family subscribers with Excel and Word Agent mode available on the web with desktop coming soon, an Office agent currently used only in English. This is why I could I looked for this today on my my Office apps on my Mac and I said I just turned it only one to Windows but apparently it's just because it's only on web so I should try it out there. But yeah, good to know. [00:07:24] Speaker D: Wait, the PowerPoint ones on web two? [00:07:27] Speaker A: Yeah, web. There's web apps for all of them. They're not not as good as Google Web apps, but they pretend to be. [00:07:33] Speaker D: I didn't realize that PowerPoint had one. I I guess I've never used that. I always just open the app I. [00:07:37] Speaker A: Guess by in teams I have a default to use the apps because I hate the web interfaces for all of them. [00:07:43] Speaker D: That's what I do. [00:07:45] Speaker A: Yeah, they all have the option to use a web interface if you'd like. [00:07:48] Speaker D: To learn something new. [00:07:51] Speaker C: Yeah, for a long time if you were a Mac user you had no choice, right? Because half the features were on. [00:07:55] Speaker A: I mean if you want to use Visio, that's still your only choice for Mac because you know, decades later they still haven't ported that product to Mac. [00:08:03] Speaker D: Just go with lucidchart Dry IO and them are equally as good. [00:08:08] Speaker A: I like the Omni Graffle but it's so expensive so yes, Lucid chart's nice. Well Claude Sorry? Well, Anthropic has launched Claude Sonet 4.5 which achieves a 77.2 on the SWE bench of verified, positioning it as a leading coding model with ability to maintain focus for over 30 hours on complex multi step tasks. The model is available by API at $3 for input tokens and $15 per dollars per million output tokens matching the previous Sonnet 4 pricing, which is nice. The Cloud Agent SDK provides developers with the same infrastructure that powers cloud code, enabling creation of custom AI agents for various tasks. Beyond coding, this includes memory management for long running tasks, permission systems and sub agent coordination capabilities. Computer use capabilities improved significantly with 61.4 on the OS World Benchmark, up from 42.2, enabling direct browser navigation, spreadsheet manipulation and task completion. The cloud for Chrome extension brings these capabilities to max subscribers. New product features include checkpoints and cloud code for progress saving and rollback, a native VS code extension, context editing with memory tools in the API and direct code execution with file creation for spreadsheets, slides and documents in the cloud apps. Early Customer results show 44% reduction in vulnerability intake for security agents, 18% improvement in planning performance for Devon, and zero error rate on the Internal code editing benchmarks, down from 9% previously. The models operate under ASL3 safety protections, while improved with improved alignment metrics. There's also a couple other little things they released as well, including a new update to the SDK, as I mentioned, and a new Imagine with Claude feature, which I described to Jonathan earlier as Microsoft Bob for AI. [00:09:49] Speaker B: Imagine, it's cool to play with. How to play with that. It's only available for the next couple of days, I think three days, and then they're going to take it out of preview again. I mean, it kind of builds apps that look like they belong in Windows workgroups. But it's kind of neat though, because it really is getting to the no code. The AI provides the interface and the AI provides all the business logic driven through prompts only. So I think it's kind of where we predicted we were going to end up eventually. I think this is just a sneak peek of what might be coming. Um, but Sonnet 4.5 is awesome for with Claude code. I had to go back last night and revisit some of the work old Claude had done on my projects and I'm like, analyze this project and you know, tell me what you think and make it better. And it's like, oh, I don't know who. Don't know who touched this last, but we don't want to call them again. [00:10:45] Speaker C: That's my favorite. I love what it makes fun of itself. [00:10:48] Speaker A: Yeah, I did notice clouds on it. 4.5 is a little less friendly. About before, I was like, that's a great idea. Or you're absolutely right, you should do that. It's doing a little less of that, which I actually appreciate because I'm like, it's just annoying. I'll put contacts I have to pay for. So it's a little bit more direct than the old version of Claude, which one of the things people liked about Claude is that it is a little bit more personable than some of the others. So I hope they haven't overtuned the wrong direction on that. But so far I've been pretty happy with Claude's 4.5 stuff already as well. [00:11:17] Speaker B: Yeah, when I gave it those projects, I think the biggest thing that it did, the most important thing that it did, was to look at the architecture that had been built and then realize how it could have done it better. And it was like, you know, code reuse, building modules to do things for Terraform instead of building separate files for each, you know, Kubernetes deployment is an example that came up. So it kind of shrunk the code base by about 50% and it still worked. I let it go in kind of YOLO mode overnight and it worked this morning, which is pretty impressive. [00:11:57] Speaker C: That is impressive. Well, I've been using Sonnet 4 pretty much exclusively for coding just because the results I've been getting on everything else is a little bit hit or miss. But I definitely don't, I definitely don't let it go off because it'll go off on tangents and so it's kind of neat to see. Like, it'd be kind of nice. [00:12:20] Speaker A: I'm hoping to get into the cloud code for Chrome Research Preview because that one also looks very cool. [00:12:27] Speaker B: I have that. It's very neat. [00:12:29] Speaker C: Is it like browser based ideas clawed. [00:12:34] Speaker B: In an extension in Chrome and it's pretty limited at the moment to only the tab that you currently looking at, so you can't do things across multiple tabs yet. [00:12:43] Speaker C: Oh, so it can take action in the browser. [00:12:45] Speaker D: Gotcha. [00:12:45] Speaker B: Yeah. So it can do things on ab. [00:12:47] Speaker A: That's cool. [00:12:48] Speaker B: It's. [00:12:50] Speaker A: Also, I've been playing with some browser mcps for some web development work that I've been doing and that's been really nice because, you know, having Claude be able to go use the MCP to access the web browser to see what it's doing versus me taking a screenshot of it and pasting it into the thing, it's not a huge time saving, but it is kind of nice because it also looks at the dev tools, looks at all the other things that it uses to debug the app. So me as a not front end developer, is suddenly a better front end developer because of cloud code, which I appreciate. So I'm hoping this research, this Chrome tab thing is kind of similar and gives me some of the similar capabilities as the mcp. But the MCP is also quite good if you're looking for something like that. Encoding. [00:13:30] Speaker B: Yeah, I think the problem we're going to run up against, which I've already come across now, is Robots txt. Everyone's updated their Robots TXT files to block AI page scraping, basically. So even if it's not for training, if you ask a model to hey, can you fetch this LinkedIn profile and tell me what you think of this person? Robots Txt has been updated to disallow that completely, which is kind of frustrating because I'm, I'm browsing the site, I want to use Claude to analyze the content of the site, but when it makes a request it's considered A bot. Even though I'm literally sitting here saying do this for me. [00:14:11] Speaker C: Yeah, but they can't sell you advertising or whatever. [00:14:17] Speaker A: Well, it's interesting because hashiconf was last week and I was trying to find the details I want to talk about on the show. And the, the problem that I have with hashiconf is that they root all their blogs to Stack as part of the IBM acquisition. And then even for RSS readers they consider an RSS reader to be a bot. So they won't. They just like, they've completely broken RSS functionality, which is like how I live on the web. So I had to send a note to the person who wrote the blog post about our amazing new stack, our amazing new stack site. And I said, hey, just so you know, I think you've overtuned the robots txt to not allow, you know, simple RSS readers to be able to access your feed. And I don't mind clicking on your website if you just want to give me a snippet, but I'd like to at least have the feed. So he said he was going into it but he hasn't got back to me yet. But I'm hoping, hoping he fixes that soon. And speaking of that, I think it actually was some articles we tried to put into the show, this from that that didn't make it in because I think Bolt got blocked. So you know, there you go the way the world works with ChatGPT and all these things. And then you know, even in the why Jonathan was out on a secret slide mission we I was dealing with problems like I couldn't browse, I couldn't use Claude to parse OpenAI's website blog. So I had to like add to Bolt the ability to, if it's an OpenAI website, to use ChatGPT API instead of using the Claude one. So I've been this, this whole bot blocking thing is definitely a problem. Yeah. That I run into occasionally on some of the corporate blogs that we, we follow for the cloud. [00:15:51] Speaker B: Yeah, I mean it's the hypocrisy really. You've got Microsoft and OpenAI and you know, they're making money hand over fist over with these products but they're going to block you from using other the competitors tools to read information from their sites. It's not, it's not in the interests of the users I don't think, or the customers. [00:16:10] Speaker A: Well, and the Beauty of open LLMs or LLMs that are widely distributed across multiple clouds, we now get to give you the pleasure of now learning all the places that Cloud Sonnet 4.5 is as we go through show us today. So first up, Databricks is supporting Cloudson at 4.5. [00:16:25] Speaker D: Yay. [00:16:26] Speaker A: And also Snowflake Cortex AI is also supporting Anthropic Cloud Sonnet 4.5. Weird. I don't know they all got a press release at the same time. Databricks also is giving us a SQL Server connector for lakeflow Connect in general availability, providing fully managed data ingestion from SQL Server to the lakehouse with built in CDC and chain tracking support, eliminating the need for custom pipelines or complex ETL tools tools. The connector addresses the common challenge of SQL Server data being locked in transactional systems by enabling incremental data capture without impacting production performance, supporting both on PREM and cloud SQL Server environments through a simple point and click UI or API. And I really just leave this here because one of the show note titles that we did not go with this week was SQL Server finally gets a Lakehouse instead of a Server Room and I was wishing that SQL Server would just go out to pasture already instead. [00:17:14] Speaker D: Needs to go upstate. Go to the farm guys. [00:17:18] Speaker A: So if you're a databricks shop, you're welcome. I mean they're just copying a feature that Snowflake already had, to be honest. [00:17:24] Speaker C: But I mean it's this has been a challenge for a while for getting getting data out of these transactional databases so that you can run large reporting jobs on. So I like any kind of easy button that moves you out of that ecosystem. [00:17:37] Speaker A: Agreed. Well let's move on to AWS and guess What? Bedrock supports Cloudson at 4.5. [00:17:45] Speaker D: No way. [00:17:46] Speaker A: Yeah, that's how I use all my cloud code connects to Amazon directly now because I just use their I just use the API for Bedrock to run cloud code. So I was mad because it wasn't working. And then I remembered oh yeah, in Bedrock you have to go enable the new model one by one. So if you are trying to use Bedrock and it's not working, remember to update your model access to allow you to access Claude 4.5. Amazon ECS and Amazon EC2 auto scaling are now supporting IPv6 only workloads, allowing containers to run without any IPv4 dependencies while maintaining full compatibility with AWS services like ECR, Cloudwatch and Secrets Manager through native IPv6 endpoints. This addresses IPv4 address exhaustion challenges and the fact that you're paying for those IPv4 addresses even if you don't want to. And the implementation Requires minimal configuration changes for you. Simply using IPv6 only subnets with your ECS task and the service will automatically adapt without needing IPv6 specific parameters. For the EC2 side of things, you used to have to do IPv6 and dual stack configuration alongside IPv4. Addressing the growing scarcity, which is now at least both options. Allows gradual migration from IPv4 to IPv6, reducing risk during transition, and allows you to turn off the IPv4 as well there, so you can stop paying for that. This one's available to everyone in a commercial AWS region except for New Zealand. Sorry, New Zealand. The feature integrates with existing VPC configurations and requires no additional charges beyond standard EC2 and networking costs. So there you go. [00:19:15] Speaker D: I don't know. [00:19:15] Speaker A: It's up in the New Zealand though. That's a weird limitation. [00:19:17] Speaker C: Yeah, I mean, it's a fairly new region. [00:19:19] Speaker D: Yeah. I'm really curious why they're still studying. [00:19:21] Speaker A: It. Must be like, they're like, we don't support IB4 at all there. We never enabled it. Maybe that's the secret of New Zealand and I didn't catch when they announced it. [00:19:29] Speaker C: Yeah, or it's just capacity. [00:19:31] Speaker A: Right. [00:19:31] Speaker C: So it could be simply. [00:19:34] Speaker D: I mean, it is amazing how, how fast that, you know, IPv4 cost does add up in your account, you know, especially if you have load balancers, multiple subnets, and you're running multiple ECS containers and public subnets for some reason, you know, or anything like that. Like that, that little bit adds up quickly once you start running tens to hundreds of containers and really can, you know, take a good chunk out of your bill. So it's nice that to me, they should have made sure they had all these features before they started charging you for them. But, you know, they needed to make money somewhere along the way. [00:20:09] Speaker A: So I get it, you know, it's hard these days in this tough economy. You have to justify every investment you want to make with revenue or cost savings. So, you know, they had to make revenue before they could give you cost savings. Amazon easy to allowed AMI setting is a new parameter for enhanced AMI governance. These four new parameters, marketplace codes, deprecation time, creation date and AMI names, give organizations more granular control over which Amazon machine images can be discovered and launched across your AWS account. The Marketplace A marketplace code parameter addresses a common security concern by allowing teams to restrict usage to specific vetted Marketplace AMIs. While deprecation time and creation date parameters help enforce policies against outdated or potentially vulnerable images. The AMI name parameter enables enforcement of naming conventions, which is particularly useful for large organizations that use standardized naming patterns to indicate compliance status, department ownership and or approved software stack. These parameters integrate with AWS declarative policies for organization wide governance, allowing central IT teams to enforce AMI compliance across hundreds or thousands of accounts without manual interventions. It's available to you at no additional cost and represents a practical solution to the challenge of shadow it. [00:21:18] Speaker C: Yeah, I mean this is a bane of my existence for decades now. Like I remember trying to do this with SCP and doing conditionals on parameters. [00:21:29] Speaker A: That did not exist. [00:21:31] Speaker C: Trying to get features like this. So it wasn't just pulling some rando OS image from someone and someone's public AMS AWS account. Like you just have no idea what you're running in a lot of these cases. [00:21:43] Speaker A: So it's, I mean even there was some other things they tried to do with marketplace where you could you create a private marketplace for your company so you can only put in the approved apps and ami. Remember all that which is all a lot of noise and it's like this is much cleaner I think than trying to create a custom marketplace. I have to manage separately from the main one so I much prefer this option. [00:22:05] Speaker C: Yeah, no, this is. It's definitely easier to just define what you need and then give it the rule set. [00:22:11] Speaker D: I thought the marketplace for your organization was more like here's our golden image as an easy way to share. So you have to share it with the organization. [00:22:20] Speaker A: So you could do that as well. Although they'll tell you that's not there but they built it for. So yes, you can do that. [00:22:26] Speaker D: That's what I thought it was built for. [00:22:27] Speaker A: You can do that but they'll tell you like that's not why we built it though. They really bought it. So you, if you had a Red Hat subscription that you bought at an organizational level that you could then distribute it through a private marketplace to all your sub accounts so they could get benefit of your subscription. That's why they built it. [00:22:41] Speaker D: Oh, interesting. I guess I just avoid license the. [00:22:45] Speaker A: Whole using it on for your own AMIS was not, not appreciated by the product manager. When I talked to him the first time about it I was like well can I just use this to publish ami? And he's like well you could but we really don't want you to do it that way. Yeah, okay. They want you off these amis. Yeah, that's a good feature request though. We should get that to somebody else and like cool, cool. So never going to get it. That's how that's going to work. [00:23:06] Speaker B: I think this is kind of where Google's actually still ahead of Amazon in the flexibility around IM and these images. I think it'd be nice if you could just add any tags and use any. You put a policy together based on any kind of tags you like and sort of being restricted to image names, marketplace codes. [00:23:26] Speaker C: I mean the trick is if you're talking about public ami, so like you can. Now you could set these on stuff that were. That may not have a tag and you don't have control of the tags. [00:23:36] Speaker B: That's. That's fair. That's fair. I'm thinking, I don't know, I mean, I guess I've worked in places where we have to build our own amis anyway and then. And it'd be super nice to be able to have more control. Like you can, you can deploy this image as a new instance, but, but you can't deploy this image as a new instance. However, if it's an auto scaling group, you can scale up with the same image you've already configured for. But I don't know, it'd be nice to kind of have more flexibility around, around the control about what, what people can do with which images. I like. I like the idea of disappearing images though, that once it's six months old, you just can't see anymore. You have to use a new one. [00:24:10] Speaker C: Yeah, I mean that's my, that's my favorite feature in Google is that. And it doesn't destroy your workloads that are still running. You just can't deploy it. [00:24:20] Speaker A: You guys tell me all the time that Google's AMIs really AMIs, so. Which I still don't fully understand. [00:24:25] Speaker C: Well, they don't call them amis, they call them compute images and it's just a terrible name and. Or VM image. I don't know. I don't even know the name because it's so generic and boring. So I just call it amis. [00:24:33] Speaker A: Like all Google services generic and boring. That's how we like to describe them. [00:24:37] Speaker B: Yeah. [00:24:39] Speaker D: Did AWS just release a feature of like tracking where all the AMI is used across? [00:24:44] Speaker A: Yes, they did. We talked about it a few weeks ago. [00:24:45] Speaker D: Yeah, I was trying to remember if that was the AWS or Azure thing that they were talking about. [00:24:51] Speaker B: So just wait six months and they'll all have the same features anyway. [00:24:57] Speaker D: Sometimes it's not six months. [00:24:58] Speaker C: It's weird. [00:24:59] Speaker A: Sometimes it is, sometimes it's not. [00:25:03] Speaker D: I feel like by the end of the podcast and my next day at the Office is trying to remember the right acronyms based on the people I'm talking to. So I have to go back to all Asher versus like here we talk aws, like all them. So by Monday, by the next morning, I'm always like, I don't know. It's a, it's a vm, it's a, you know, scale set, it's a mig, whatever you want to call it. It's a damn auto scaling group. So it goes up and down. [00:25:24] Speaker A: I use Amazon terms all the time, like, oh yeah, well I just need an S3 bucket. They're like a GCS object bucket. I'm like, oh yeah, yeah. I just. You're just. Oh, when you attach a drive to it, that's. That's ebs. And that's partially because everything is GCS for objects and drives in Google, which is confusing, but yeah. Anyways, easy to Auto scaling now allows a forced cancellation of instance refreshes by setting wait for transitioning instances to false in the Cancel instance Refresh API, enabling immediate abort without waiting for in progress launches or terminations to complete. This feature addresses emergency scenarios where rapid roll forward is needed, such as when a current deployment causes service disruptions and the team needs to quickly abandon the problematic refresh and start a new one. The enhancement provides better control over auto scaling group updates by bypassing lifecycle hooks and pending instance activities, reducing downtime during critical deployment issues. Available in all aws regions, including GovCloud, this feature integrates with existing auto scaling workflows. And I was like, this isn't really that big of an issue. And I was like, well, but I've had a really big auto scaling group. This could be a really big problem. So like this is a problem if you have large scale that if you have like five web servers, you probably don't care, but if you have hundreds in that auto scale group, this could be a big lifesaver for you. So appreciate it. [00:26:37] Speaker C: I've definitely been in scenarios and it's usually trying to recover from my own mess up, right? [00:26:41] Speaker A: Like, no, it's always when I mess. [00:26:43] Speaker C: It up, it's trying to refresh through it. [00:26:45] Speaker A: I'm like, it's doing instance refresh. Cause I pushed out a new image and then I realized, oh crap, I have a major defect. Yeah, that's. That's me. [00:26:52] Speaker C: Yeah, I'm waiting for that problem. It's terrible. It's taking years off my life. So I'm happy to see this. [00:26:58] Speaker A: And then it never fails. Jonathan or Matt or Ryan Notice the website's down. And I'm like, yeah, yeah, sorry guys, like fucked it up again. [00:27:05] Speaker D: Yeah, we just keep you on your toes. [00:27:08] Speaker A: It's fine. Two in the morning. Also, I get a message from Jonathan like, hey, the website's down. I'm like, why are you awake? Go to bed. [00:27:16] Speaker D: He has a Route 53 health check. Yeah, we've got synthetic checks, calls his cell phone. [00:27:21] Speaker A: I'm sure you guys do. [00:27:22] Speaker D: To wake him up to text you. [00:27:23] Speaker B: I know we can't afford things like those health checks. That's like 50 cents. [00:27:26] Speaker A: Yeah, health checks are expensive and it. [00:27:29] Speaker B: Must suck being a product owner. Sometimes you come up with these great new features and then the very next new feature request you get from a customer is, yeah, we like those features, but now we need a way to not use those features in an emergency. [00:27:40] Speaker C: Yeah, exactly. [00:27:41] Speaker A: Yeah, we like it most of the time, but this 1% of the time we really hate it and we need it to stop. Yeah. It's sort of funny to me that they didn't even bother just writing this into the console. Like we gave you an API call with some really weird cryptic name, waiting for transitioning instances to false and cancel instance Refresh API. That's all we're going to give you right now. Let's not put it in the GUI for you. [00:28:03] Speaker C: If they put it in the gui, it would just. [00:28:05] Speaker A: Oh, it would cause nothing but support. [00:28:06] Speaker C: Cause nothing but. [00:28:07] Speaker D: Yeah, they're like, no, I want it to be done faster. Hit the cancel button. Be like, it's done, but now I'm. [00:28:13] Speaker C: On different images, but now everything's broken. [00:28:16] Speaker B: Yeah, actually I like that about the allowed AMIS thing because the configuration from that, it's not like a really nice cloud formation example or something else is. Like you have to create this JSON document and upload it to us. That is not the AWS of two or three years ago. [00:28:31] Speaker C: No, it really isn't. But yeah, but I'm okay, I'm here for it. [00:28:35] Speaker A: That's fine. I don't like the console and you like JSON, so it makes sense. [00:28:41] Speaker C: Yeah, I do. [00:28:45] Speaker B: There are a lot of cloud cost management tools out there, but only Archera provides cloud commitment insurance. It sounds fancy, but it's really simple. Archera gives you the cost savings of a one or three year AWS savings plan with a commitment as short as 30 days. If you don't use all the cloud resources you've committed to, they will literally put the money back in your bank account to cover the difference. Other cost management tools may say they offer commitment insurance, but remember to ask will you actually give me my money back? Achero will click the link in the Show Notes to check them out on the AWS Marketplace. [00:29:23] Speaker A: Well, in a feature that caused your host to debate this for quite a while, Amazon ECS managed instances bridge the gap between serverless simplicity and EC2 flexibility by providing fully managed container compute that supports all easy to instance types, including GPUs and specialized architectures, while AWS handles provisioning, scaling and security patching. The service automatically selects cost optimized instances by default, but allows customers to Specify up to 20 instance attributes when workloads require specific capabilities, addressing the limitations that prevented customers with EC2 pricing commitments from using serverless options. Integer management includes automated security patches every 14 days using Bottle Rocket OS, Intelligent Task Placement to consolidate workloads onto fewer instances, and automatic termination of idle instances to optimize your cost. Pricing consists of standard EC2 instance cost plus a management fee for the Cloud Pod 1 single node ECS node. It was $0.03 for that management fee and it's available to you in six regions including US East, US West, Europe, Africa and Asia Pacific specific with Support for console, CLI, CDK and CloudFormation deployments. Sorry, no JSON. This addresses a key customer pain point where teams wanted service operational simplicity but needed specific compute capabilities like GPU acceleration or particular CPU architectures that weren't available to you in Fargate. I mean I love Fargate, but I don't like paying for Fargate so that's why I run our Cloud Pod website on an EC2 instance because it's way cheaper and so for $0.03 more a gig versus going to Fargate, this is probably where I would land if I didn't really want to manage the host. But like we talked on the hosts on the before show like things like agent installs probably aren't allowed and there's a bunch of other limitations of this so you get more flexibility, you get GPU access, but you still have some limitations. [00:31:02] Speaker C: But I bet if you can run that agent as ICAR container like it's it's you get the node groups like and that's not something that you really get with in Fargate, at least not in ECS Fargate. So that's it's kind of neat to to see this because it's all the benefits of like being able to manage at the the compute layer. But I don't have to patch it and I'll pay that premium for that for sure. [00:31:26] Speaker D: Where's the shared security model line Though, because like if I don't manage the host, then why do I need to run an agent on it? Are you saying run an agent on the. A sidecar agent on like your workload? [00:31:39] Speaker C: It'd be. Yeah, it'd be more for inspecting my workloads than it would be at the OS lady. [00:31:46] Speaker D: Yeah. For three cents if you're an enterprise, like enterprise by the time you add. You know, we were talking about this beforehand. An admin sysadmin to manage it or a cloud admin to manage this. Your security tool if you're lucky enough to only have to have one installed on your host. Depending on how much your security department hates everyone else in the organization looking at you, Ryan, right now and you know everything else. You know, 3 cents doesn't sound bad if you're. If you could convince your security department and your audit at that like while yes, there is a physical host there, I am not responsible for it, which could be an uphill battle depending on how sophisticated your team is. It could be a decent. You know, for 3 cents it's a good next step. [00:32:30] Speaker C: Yeah. I think there's an advantage is that the serverless technologies and containerized technology, especially on like Autopilot and Fargate have been long enough where the concept is there and it's at least something that I can hand wave to and explain to auditing teams now. Whereas before it was just black magic and they looked at me like I was crazy. [00:32:51] Speaker D: You have never had to do an audit on Azure before. I swear, every auditor I'm like okay. [00:32:56] Speaker A: And God willing I'll never will. [00:32:59] Speaker D: Yeah, I'm like every time I join one of these meetings I'm like okay, what cloud do you know? Okay, we know aws, but you know you're doing on Azure. Yeah, we are familiar with Astro. Okay, cool. What language you want me? Can you translate everything to AWS first? Sounds great. I'm just a cloud translator. Babelfish for cloud. [00:33:19] Speaker C: It works. Everyone speaks aws. [00:33:21] Speaker A: Everyone speaks aws. That's what we talked about earlier. Like I just. Everything's EC2 and our S3 buckets to me. Announcing AWS outposts. Third party storage integration with either Dell or HPE. This gives you access to the Dell Powerstore and HP Electra storage NPB 10,000 arrays. Joining existing support for NetApp and Pure Storage. Allowing customers to use their third party storage investments with Outposts through native AWS tooling. The integration supports both data and boot volumes with two boot methods. ISCSI Sandboot for read write volumes and local boot for read only volumes using ISCSI or NVME over TCP protocols manageable through the EC2 Launch Instance Wizard. This addresses two key customer needs, organizing and migrating VMware workloads who need to maintain existing storage during transition, and companies with strict data residency requirements that must keep data on premises while using AWS services available at no additional charge across all outpost form factors 2U and both rack Generations and all supported regions with AWS verified AMIs for Windows Server 2022 and Red Hat Linux 9 plus automation scripts on AWS samples. Second generation outpost racks can now be combined with Double the compute performance 2xb CPU memory and network bandwidth with customers preferred storage array providing flexibility Hybrid cloud environments so originally I thought this was you could buy your Outpost with the storage from HPE or Dell, but I now figured out that's not the case. But I guess I'm surprised that this was a thing that was a problem before because if it's on your network and it can connect to it through iscsi I would have thought it would define but apparently now it's more natively supported than Nitro Card. Is that how you guys read this? [00:34:53] Speaker B: Yeah, I think it's more that you can not have AWS provide the storage layer, but you can have them still support S3 and EBS and those other things on top of this third party storage subsystem. [00:35:06] Speaker A: Makes sense. I mean I appreciate having the access to different options other than just pure netapp. Dell and NHP are really the other two players in the market, so makes sense. I would still like to see Outpost capability on Dell or HP or IBM hardware without having to use the proprietary stack. But I think Amazon's still pretty proud of their Outpost design. [00:35:30] Speaker D: Azure Stack Hub. [00:35:34] Speaker A: Yep, not doing that either. [00:35:35] Speaker D: Not doing that either myself. [00:35:37] Speaker A: You're not tricking me into this Azure thing, Matt. You can keep trying all you'd like to. [00:35:43] Speaker B: I actually wonder if the reason they're doing this is not because of what they say, but because the outposts don't come in a sensible size form factor. [00:35:56] Speaker A: With the amount 2U or 48U is you know 42U is not a you know there's no middle like no just right size. [00:36:03] Speaker B: Well I mean I'm thinking for look at the globally there's dozens of data centers being built out right now for AI. True and they use an enormous amount of storage and they're sucking up all the storage they can possibly find. First they came for your hard drives, then they came for your NVME if you need, if you need storage by it now because it's going to, it's going to get expensive I think. But I mean, I guess you need a lot more storage than you do compute. And I think it's probably like the ratio of storage to compute in outposts were sort of normal use cases, not these very special use cases. So it's probably driven by customers saying, but I don't want to pay for all this compute when I'm not going to use it. [00:36:45] Speaker A: Moving on to gcp, Google's launching Flex Start VMS in General Availability, a new consumption model that queues GPU requests for up to two hours instead of failing, failing immediately. Addressing the persistent challenge of GPU scarcity for AI workloads, this appears to be unique among major cloud providers. Rather than competing on raw capacity, Google is innovating on the access model itself by introducing a fair queuing system with significant discounts compared to on demand pricing. The service integrates directly with Compute Engine's existing instance API and cli, allowing easy adoption to current workflows without requiring a migration to a separate scheduling service with VMS running for up to seven days uninterrupted. Key use cases include of course, AI model fine tuning, batch inference and HPC workloads that can tolerate delayed starts in exchange for better resource availability and lower cost. Particularly valuable for research and development teams. The soft start capability with automatic re queuing and configuration actions provides flexibility for long running experiments while managing costs effectively. I love this. [00:37:42] Speaker C: This is great. I mean it's, there's, you're still going to see a whole bunch of like data scientists and, and stuff spamming the workbooks but you know, trying to get this to run. But I do think that from a, you know, pure capacity standpoint like this is the right answer to some of these things because a lot of these jobs are very long running and you know, it's not really instant results and it's so painful when you're sitting there just like hitting the button and getting, you know, access denied every time until it finally goes through. So this is great way to solve that problem. [00:38:15] Speaker B: Yeah, it's like, it's like the next iteration of Spot really like the evolution of spot instances. So they don't really talk in here about you combine something like this with Spot Spot now and then, now you've got a, now you've got a fair queue, but if you bid higher you get to top of the queue faster. [00:38:31] Speaker A: I mean, wasn't it kind of always that way? Now with the spot Market, you bid higher, you get up to the, you get first capacity first. [00:38:38] Speaker B: Yeah, I mean I guess you could always, you could, you could always have implemented this by having a auto scaling group and then when there was capacity available, it would have spun up the machine and done. Done the thing anyway. [00:38:48] Speaker C: But imagine that. Yeah, imagine that. Would it either get expensive really fast or it wouldn't work as. Because you'd end up spinning up computer GPUs there and then. [00:38:59] Speaker B: Yeah, I think what would really help honestly is if you completely separated the GPU compute from the. From the CPU compute. Unfortunately, Nvidia have made it very difficult to do that with the licensing terms of the drivers. You have to have the cards, the machines that you're using the, that use a CUDA on. You're not allowed anymore to build an abstraction that would make it look like you're running CUDA locally but actually running it someplace else. Yeah, I mean for most use cases, you know, at home over a 100 meg network, it would suck. The performance it would be will be horrendous. But when you're talking about these cloud providers who've got 10 gig backbones connected to every single hypervisor, I would imagine it would work quite nicely. [00:39:43] Speaker C: Yeah, I've always thought it was strange that they sort of coupled those so tightly together just given how distributed computing works these days. [00:39:52] Speaker D: Right. [00:39:52] Speaker C: But okay. [00:39:54] Speaker A: GKE Autopilot is now available in standard clusters through compute classes, allowing existing and GKE users to access container optimized compute without migrating to dedicated autopilot clusters. This brings efficient bin packing and rapid scaling to 70% of GKE clusters that weren't using Autopilot previously. Container optimized compute platform sizes just 50 milli cpu 5% of one core and scales to 28 VCPU with customers only paying for requested resources rather than entire nodes. Addressing the common kubernetes challenge of over provisioning and wasted compute capacity. New automatic provisioning for compute classes lets teams gradually adopt autopilot features alongside existing node pools without disrupting current workloads. Solving the previously all or nothing approach that made migration risky for production environments. AI workflows can now run GPUs and GPUs with autopilot to manage node properties and enterprise grade security controls. Competing directly with AWS EKS Auto Mode and Azure AKS automatic node provisioning but with tighter integration to Google's AI systems available starting with GKE version 3.1.33.1 in the rapid release channel. Available to you now. So nice. [00:40:58] Speaker C: So now you can have not only like dedicated compute, but preemptible and now autopilot capacity, all in a single cluster, which is kind of, kind of cool. [00:41:07] Speaker D: So I feel like I need a chart, like, which one's autopilot, which ones? All these other ones. As a non Google person, I feel like I've got. Tracking all the, all the different specific features. [00:41:18] Speaker C: Well, preemptive is just their version of Spot. [00:41:20] Speaker D: Yeah. [00:41:21] Speaker C: So it's, you know, it's, it's a way to have sort of a cheaper compute layer that may go away, but you know, having that and then, but having still a dedicated core so your workload doesn't go crashing to the ground if the market goes crazy. But then autopilot kind of allows, you know, management of these things to go away. And so now it's expanding that out, which is kind of neat. Although expensive. [00:41:43] Speaker A: So everything's expensive. [00:41:46] Speaker C: Everything's expensive, yeah. [00:41:48] Speaker A: Google is launching Gemini CLI extensions for data cloud services including Cloud SQL, Alloy DB and BigQuery, enabling developers to manage databases and run analytics directly from their terminal using natural language. Go to prompts. What could go wrong? The extension allows developers to provision databases, create tables, generate APIs and perform data analysis through conversational commands, potentially reducing the time needed for common database operations and eliminating context switching between your tools. BigQuery extensions include AI powered forecasting capabilities and conversational analytics APIs, letting users ask business questions in natural language and receive insights without writing SQL queries. Key use cases include rapid prototyping for startups, data exploration for analysts who aren't SQL experts, and streamline database operations for DevOps teams managing multiple Cloud SQL SQL or Alloy DB instances. [00:42:34] Speaker C: Yeah, I mean, you're worried about the performance hit of some, you know, natural language query or database and whatever that's going to do. But compare that to me trying to learn SQL against this production data set. So, you know, it's, it's pick one. [00:42:47] Speaker A: I mean, I don't know, I'd almost prefer Select Star from Star versus some of the inner joins that AI might come up with. [00:42:55] Speaker C: Yeah, it depends on like if you're trying to debug business logic, you know, long term and. Yeah, probably, but performance and databases, it's always so hard when you hit those performance bottlenecks and you're like, what's causing this? [00:43:09] Speaker A: Could you imagine that Vertex has a new model this week? Really? [00:43:13] Speaker C: I wonder which one it is. [00:43:15] Speaker A: Cloudsonic 4.5, available to you via Vertex AI. Woo. It does integrate with the Vertex AI's agent development kit and Agent engine for building your multi agent systems so you get all the power of Vertex with your Claude 4.5 magic. Awesome. [00:43:33] Speaker D: Woohoo. [00:43:35] Speaker A: GKE Compute classes now let you define a prioritized list of machine families for auto scaling, automatically failing back to alternative VM types if your preferred options aren't available. Solving the challenge of adopting a new gen 4 machine like N4 and C4 while maintaining workload availability so this is a solution to a problem that Google has because they're terrible at capacity capacity planning. Perfect. Also this is a this is their version of AWS Fleets Compute. Flexible cuds provide spend based discounts up to 46% that allow follow your workload across different machine families. Unlike resource based cuds that lock you to specific VM types, the combination addresses a real adoption barrier, compatibility testing through gradual rollouts, regional capacity constraints with automatic fallbacks and financial commitment alignment by allowing discounts to apply across multiple VM families including both new and legacy options. They highlighted Shopify's success using this approach during Black Friday and Cyber Monday last year, prioritizing new N4 machines with N2 fallbacks to handle massive scale while maintaining cost optimization through their flexible cuds. This is also great for large GKE fleets and any other large auto scaling thing that can handle multiple CPU architectures. [00:44:41] Speaker C: Yeah, I mean it's just more options at that compute layer where if if you have a workload where you don't have to be, you know, super particular about the CPU capacity and memory ratio of the two like this is a great option and then having it be a failbox model I think is fantastic. [00:44:59] Speaker A: Right? [00:45:00] Speaker C: So you have your preference unit but you have a sort of get out of jail card and your performance may take a hit but you know it's better than it going down. [00:45:09] Speaker A: Agreed. BigQuery is now offering you two new AI tools for data analysis. Ask Data Insight enables natural language queries against structured data using Conversational Analytics APIs, while BigQuery Forecast provides time series predictions using the built in Times FM model without requiring separate ML infrastructure setups. These tools integrate with both Google's Agent Development Kit and Model Context Protocol toolbox, allowing developers to build AI agents that can analyze BigQuery data and generate forecasts. Just a few lines of code Positioning Google against AWS, Bedrock and Azure OpenAI the ASK data Insights tool provides transparency by showing step by step query formulation and execution logs addressing enterprise concerns about AI black boxes when analyzing sensitive business data, while BigQuery forecast leverages the AI forecast function to deliver predictions with confidence intervals. Key use cases could include retail Sales forecasting, web traffic prediction and inventory management, with a Demo showing Google Analytics360 data analysis particularly valuable for businesses already invested in Google Analytics. Both tools are available today in the McP toolbox and ADK's built in tool set. [00:46:12] Speaker C: Yeah, I, I mean this is really neat. And then the fact that it does show you the logic all the way through, which I think is super important. You know, like it's, you can ask natural language questions and it just comes back with a whole bunch of analysis and then what happens if that doesn't work consistently? How do you debug that? But this is basically, you know, building it so it's, which is, you know, how I learn anyway, so that's, it works really well when it's, it's spitting out the actual config for me instead of just telling me what the results are. And I still don't have to know anything, which is important. [00:46:43] Speaker D: I feel like it's always like the math teacher that's always like show your work. And that's kind of what I feel like whether these things walk you through it. Like it's showing me the thought process it made in order to either agree or disagree every step and give it, for lack of a better term, partial credit. Cool. You got the first three right and you went south on number four. So let's kind of go back to number four and revisit. [00:47:05] Speaker A: I mean it's my favorite and least favorite feature of OpenAI's new models right now on ChatGPT5 is the forking. Because this is what happens to me when I do reasoning. Because like it'll go down a path and all of a sudden it reasons in a way. I'm like, huh, that's a weird reasoning path fork. Why did you go down that path and what do you mean by this? And then I'm down that thing and then it does, you know, I go through that whole thing and then another and like next thing I know in like six forks down, like what was I doing? What day is it of the week? I've lost all, all tracking. It's, it's a great way to get rat holed pretty quickly. But I do, I do love it and I still, and I, I think I mentioned on the show when this came out first time that I wish they would add it to cortex. It is now in cortex. I saw it today when I was playing with cortex on something and I would love Cortex if it wasn't so darn slow. But it compared to anthropic cloud code. It is so slow, but it's it's a lot more methodical about how it thinks about things. So the output is typically better, but I feel less productive because I get less prompts done per hour using Cortex than I do or codecs using versus using cloud code. So I'm definitely hoping to see some improvements come back to cloud code because it was definitely feeling a little dumb in the last couple of weeks. Probably because they're getting ready for this Cloudsonnet 4.5 launch perhaps. But definitely my initial two days of using cloud code with the new Sonnet 4.5 has been very, very good. So hoping we're back to back to the old days of cloud's awesome and Codex isn't as awesome Moving on to Azure Matt, your time to shine Microsoft's announcing Agentic AI tools for Migration and modernization at their Migrate and Modernize Summit with GitHub Copilot. Now automating Java and. Net app upgrades that previously took months down to days or hours, Azure Migrate introduces AI powered guidance and connects directly with GitHub Copilot for app modernization, enabling IT and developer teams to collaborate seamlessly while providing application awareness by default and expanded support for PostgreSQL and Linux distributions. The new Azure Accelerate program combines expert guidance with funding for eligible projects and includes the Cloud Accelerate factory where Microsoft engineers provide zero cost deployment support for over 30 Azure services. So yeah, that's nice. I'm sure the Gentek AI will rebuild my 20 year old application perfectly on the first try. [00:49:22] Speaker C: Well, I mean it's funny because I'm jaded on this one. I'm like, well the dev teams are still going to find some excuse not to do this for some reason. I don't know why. But then it's a starting point. Well, and maybe it's not a starting point. Get these things migrated because you can't run them on these ancient frameworks that are full of vulnerabilities and because you're afraid to go through your code base and upgrade it. [00:49:48] Speaker D: So I mean it's also not even that. It's, you know, you're missing new features, new performance improvements, new everything on the newer versions of. Net. So if you're still stuck on I honestly can't keep track of the Microsoft. Net naming convention. But like if you're on 4.0 and you're trying to get to 8 or whatever and that upgrade path is like multiple steps along the way and if this can do and handle a lot of that upgrade and you have a complicated code Base. It's probably worth it to try it at least. What's the worst that happens? It bombs out and either you throw it out or you tell it to go take a smaller chunk for a step one. [00:50:28] Speaker B: Now, the worst that can happen is you think it works and you put it in production and then it fails on some weird edge case. [00:50:36] Speaker C: That'S going to happen. No, no doubt. [00:50:38] Speaker A: Like that's no doubt. [00:50:41] Speaker B: I honestly think there's very little reason to modernize old code in the way people are talking with posts like this. I think it'd be much better served to just analyze what it's doing and rebuild it from scratch. I mean, it's a machine that's going to run 24 hours a day anyway. You know, you don't need to give it vacations yet. They don't have human rights yet. There's no reason to not just rip it off and start again from the beginning. Otherwise you're just going to be sort of like reinforcing those bad plans. [00:51:16] Speaker A: The reasons backwards compatibility, I assume, is that you, you know, because if you just rip and replace, then you are deprecating all the legacy ways of working. And so now you've essentially broken people who've integrated with you and so you've broken backwards compatibility. That's sometimes a big issue for some companies. Like Microsoft is very particular about that versus Apple, who doesn't give a crap. And they'll break all the old stuff every time. [00:51:39] Speaker D: Score truth. [00:51:40] Speaker C: Yeah, well, and even within a business, right, you've got, you've got a whole bunch of teams trying to coordinate. And so it's like you, it's all good and easy for one team to say, I'm just going to completely rip this up and redo a new one. But like, you know, you got to think about the teams that are going to measure the performance, the teams that have to support the underlying infrastructure. Do they have the ability to just, just make the change at the same level? And so like it's, you know, I understand and I agree with you that it's, you know, we don't spend enough time just replacing things. You know, it's we and which we should. But I do like that there's options that are more of the middle ground. [00:52:15] Speaker D: Yeah. And it goes back to what you were just saying too. If you're on, you know, app services, well, you can rebuild the entire app, I guess you could launch a parallel app service and cut over to it, you know, but you still need to do the infrastructure, everything else in order to tell it's on net 8 or whatever the latest version is and whatever you're upgrading to and kind of going from there. So there's reasons for both sides of that. [00:52:41] Speaker B: People cling onto things for far too long. [00:52:45] Speaker D: Depends. If you have many things dependent on one thing, so you got to modulize it out. There's reasons. But I agree, people like to stick with what they know. [00:52:57] Speaker B: I mean, it does the hard thing and it forces people to think about the contracts they have with their customers and whether those are other services or people or other businesses. And as long as you can reproduce the functionality, whatever you want to convert some 25 year old Perl code to run in net 8, then go ahead. [00:53:18] Speaker D: Tim, Many more questions about that statement, but we'll bypass that. Yeah, I was trying to figure out the pricing, which is remarkably difficult because it just keeps giving me to other marketing pages. [00:53:34] Speaker B: Yeah, I kind of feel like whenever there's a question of pricing it should be why are you in the eu? Well, no. Well then you're screwed. It's the feeling of the distinct lack of consumer protections here at the moment. [00:53:47] Speaker A: Then, you know, employment protections and on and on. Microsoft's unifying Azure Marketplace and AppSource into a single Microsoft Marketplace, creating one destination for cloud solutions, AI apps and agents. With over 3,000 AI offerings now available for direct integration into Azure, AI Foundry and Microsoft 365 Copilot, the marketplace introduces multi party private offers and CSP integration, allowing channel partners like Arrow Cran and TD Synx, for example, to resell solutions through their own marketplaces while maintaining Microsoft security and governance standards. For Azure consumption commitment customers, 100% of purchases for Azure benefit eligible solutions count toward their commitment, providing a financial incentive to consolidate software procurement through the marketplace. Configuration time for AI apps has been reduced from 20 minutes to 1 minute, for instance, according to Siemens, while solutions now deployable directly with Microsoft products using MCP standards, which, I mean, I guess it's nice to have one marketplace to rule them all, but 3000 AI apps feels like a lot of AI slop. [00:54:48] Speaker B: Yeah, that's a lot. [00:54:51] Speaker D: Everyone and their mother has built an AI app, you know, so like every one of our vendors is like, we have an AI app, you need to go use it. It's like, do I, Is it going to provide me value? [00:55:02] Speaker B: Yeah, I think, I think the, the MCP services are kind of blooming outwards and, and there's, there's a lot of, there's a lot of junk in there. It's like, do you really need MCP service to tell you what clothes to wear today based on the weather. I mean maybe I do kind of useful, but I think like the more MCP services you want to depend on in your workflow, the more of your context you fill up with the instructions on how to use that service. So there's like one of two ways to go. One is a cloud provider like Azure will fine tune the model so you don't need to put that information in the context every time. And that will, that's, that's a pretty high value thing. Or the other thing is like they're just rest calls. I mean MCB provides this, this, this sort of secure framework and session based tracking and things. But, but really I think like we need like an abstraction for mcps. It's an, it's the MCP that rules the other mcps. That's what we need. I gotta call my turtle. [00:56:06] Speaker A: Yeah. Azure is giving us a feature I don't understand. Azure Compute Gallery now includes soft delete functionality with a seven day retention period recovery of excellent VM images and application packages before permanent deletion. Okay, I got that. I understand soft delete, but what's an Azure Compute Gallery? [00:56:30] Speaker D: It's a machine image. Yeah, it's an ami. It's an AMI gallery. So you have a gallery and what's nice about it is think of like ECR because we only speak AWS terms here where you have the versions of it in there, your hashes and you can have a latest. There are a couple sharp edges with image galleries where you can't define what the latest is. It always has to be the latest one. You can't do like custom tags as of like six months ago or anything else like that. But it gives you some availability where you can then also say in your auto scaling groups your maintenance window is from 1am to 4am every day and always grab the latest image in here. So essentially it's an easy way to do upgrades versus the way aws and you have to press and by press I mean type your cancel API command to stop the, you know, rolling upgrade of the system so you can do it based on your automatic of your image gallery. So it does have some nice integrations and it's a good way to share. But I think this also prevents the same issue that we've all run into. Okay, I've stopped sharing this Amy. Across accounts. Oh God. We just broke production somewhere. [00:57:38] Speaker C: Broken everything. [00:57:38] Speaker A: Yeah, exactly. [00:57:39] Speaker B: Yeah, that's one of the things like about Google Cloud is the image families and the way image families work for that very reason. It's just so seamless. [00:57:49] Speaker A: I'm glad to see that capability Azure experience Pale two major regional outages in September 25, Switzerland north suffered a 22 hour outage affecting 20 plus services due to malformed certificate prefix, while East US 2 had a 10 hour incident caused by an allocator service issue that created cascading failures across Availability zones. The East US 2 incident reveals critical architecture challenges in Azure's control plane design, including aggressive retry logic meant to improve reliability actually made the problem worse by creating a thundering herd it took hours to drain even after initial issues was resolved. Both incidents highlight gaps in Azure's incident communication systems. Automate alerts only covered a subset of affected services, forcing manual notifications and public status pages. Our saturated updates hours into the outages, leaving many customers uninformed during critical periods. Microsoft's response included immediate fixes like reverting the problematic allocator behavior and adjusting filing configurations, plus longer term improvements to load load testing backlog drainage tools and communication systems scheduled through June of next year. So until June next year, get ready for this to happen at least three more times. And these outages underscore the importance of multi region deployment strategies for your workload. [00:59:00] Speaker C: Yeah, it's interesting because it's, you know, like it seems like the region has such a correlation to the, you know, the, the level of pain in these things. So it's, it's kind of fascinating there. Like is it, you know, just a management issue? Is it just a technology issue? Sounds like bad times for all involved. [00:59:20] Speaker A: Yeah, outages are never, never great. But you know, it's interesting because they've had so many blog posts about all the things they're trying to do to avoid outages, et cetera. And they were very focused on the hardware level of like predicting when VM instances were going to die and why. That's definitely a problem and something. It's kind of like the basic building block of a major outage. If you're properly architected, it shouldn't be a big, big deal. But these software issues that become like Thundering Herd and bad retry logic and you know, distributed system issues and eventual consistency, like those are all really tough problems and I'd rather they send a big blog series about how they're making that better for Azure customers. [01:00:00] Speaker D: I would too. What did you appreciate about Azure and I feel like I've gotten some on AWS is I like that. I like the way they do their outages where they kind of post the time, the, you know, minute by minute of kind of what happened. And they also post here's what we're doing to actually fix it and show like dates and whatnot. Now I've never gone back and looked. I doubt they actually go back and say this was completed. You know, they said in June 2026. It was completed in May 2026. But it's nice that they're at least publicly putting dates on stuff now. You can't really hold them to it, but it's nice that they are information sharing. [01:00:43] Speaker C: I bet the lawyers can. I bet the lawyers absolutely can. [01:00:45] Speaker D: Lawyers can do whatever they want. [01:00:48] Speaker A: I have an Oracle story this week. So many, many moons ago, Larry Ellison was the CEO and founder of Oracle and then became the CEO and kind of still the founder, but didn't really talk about himself being the founder. And then he stepped up, stepped aside and became the cto. And then there was Safra Katz and other gentlemen who went to hp, then eventually got fired for some nonsense that he did to himself and the MeToo movement. All deserved his firing. But then I think he passed away. So RIP and so staffer Katz has been basically the CEO for a while, which has been working out really well. But Oracle said no, no, no, it's time to go back to the co CEO model because they've now promoted Clay McGurrick to executive vice President of Oracle Cloud Infrastructure and Mike Silia to Executive Vice President of Oracle Industries, signaling their continued investment in cloud infrastructure and vertical market strategies despite their distant third place position behind AWS and Azure. McGurrick's promotion after leading OCI engineering suggests Oracle is doubling down on their infrastructure first approach and Cecilia's elevation to lead Oracle Industries includes indicates a focus on vertical specific solutions, a strategy that could differentiate Oracle from its competitors. These executive changes come as Oracle tries to position OCI as the preferred cloud for enterprise workloads, particularly for customers already invested in Oracle databases and applications who want integrated stack benefits or better licensing terms. And the promotion suggests their organizational stability at Oracle Cloud during a critical growth phase. So congratulations gentlemen as being the new CEOs. As Safra Katz has been appointed to the executive vice chair of the Board of directors. Right? They were formerly president executive code for. [01:02:27] Speaker C: Put out to pasture or is it like I don't want to work anymore semi retirement? [01:02:31] Speaker A: It's I've been doing this for 10 years and I'm tired of your Larry and I'm gonna go work on the board. That's how I would see it. But I don't I actually don't know. I don't know her in any way shape or form and I just know Larry's has a reputation but I assume you know you. I don't know how long she's now been sole CEO. Let's see. Safran Katz, CEO of Oracle when she's been there since 1999. She was CEO from 2019 to 2020. Sole CEO 2019 to 2025 and before that she was co CEO for quite a while as well. So imagine she wants to go do something else with her time, you know, semi retire, get money on the board and not have to stress about making sure Oracle's growth keeps going and she just leaves on a high because they just had that huge, you know, huge earnings. So things are now in a much better place. [01:03:24] Speaker B: Yeah, I think her being becoming self CEO is one of the first stories we ever covered on the podcast. [01:03:29] Speaker A: It was a very early story I do recall. Well gentlemen, we've made it to the end of another fantastic week here on the show. So thank you all. Good to see everybody and good luck next week while I'm out. [01:03:41] Speaker D: One week of all four of us. Good job guys. [01:03:44] Speaker A: One week consistency till next week. [01:03:47] Speaker C: Yeah. [01:03:48] Speaker B: See you later guys. [01:03:50] Speaker D: See ya. [01:03:51] Speaker C: Bye everybody. [01:03:54] Speaker B: And that's all for this week in Cloud. We'd like to thank our sponsor, Archera. Be sure to click the link in our show notes to learn more about their services. While you're at it, head over to our [email protected] where you can subscribe to our newsletter, join our Slack community, send us your feedback and ask any questions you might have. Thanks for listening and we'll catch you on the next episode.