# 362: Mechanical Turk: The Fall of the Ottoman API
Duration: 56 minutes
Speakers: Justin Brodley, Justin, Matt
Date: 2026-07-17

## Transcript

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

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

[00:19] Matt: Episode 362, recorded for July 7th, 2026. Mechanical Turk: The Fall of the Ottoman API. Good evening, Jonathan and Matt. How you guys doing?

[00:28] Justin Brodley: Doin' okay, Justin.

[00:29] Justin: Good. How are you?

[00:30] Matt: You know, it's, uh, been a long day. We'll leave it at that.

[00:35] Justin Brodley: But is it Friday yet? That's the question.

[00:37] Matt: I wish it was Friday. That'd be nice. I'm ready for the week to be over.

[00:41] Justin: I'm just happy it's not 115 degrees at my house today.

[00:44] Justin Brodley: Just 110.

[00:45] Justin: It was 102. It feels like 116. And I'm from Florida and I said it was too hot to go outside.

[00:51] Matt: I mean, you do live in New Jersey. Don't you have air conditioning there? You're not in Europe.

[00:55] Justin: Correct. But you know, though it did make smoking meat easier because it was so hot outside. I didn't have to like raise the smoker up. So I made beef back ribs and like, it was the easiest thing ever to cook. It actually almost cooked too fast.

[01:10] Matt: Well, that, that sounds like a problem to have. You know, I'm glad you didn't die in the heat. That's good. Unlike the Europeans have been melting to death, unfortunately.

[01:19] Justin: Yeah, they did that. They did not do well. Yeah.

[01:22] Matt: Uh, so last week we talked about, uh, Amazon killing a bunch of services and the one I forgot to mention that I did want to talk about last week, uh, there was a bunch of follow-up articles about, so it worked out perfectly. So we have a follow-up article from Amazon killing some services last week. And the service I want to talk about is them killing Mechanical Turk. Amazon's Mechanical Turk. It's closing this to new customers, effectively ending the crowdsourced human labor platform that launched in 2005 and once served as a foundational tool for AI training data labeling. The service predates the current generative AI boom by nearly 2 decades, originally designed for tasks like image tagging, transcription, and data verification that computers couldn't handle at the time. Ironically, the rise of AI and large language models has reduced demand for the type of human-in-the-loop microtasking Mechanical Turk provided, as automated systems now handle much of that labeling work without them. Existing customers can reportedly continue using the platform, but the halt on new signups signals that Amazon is deprioritizing the service rather than investing in it further. I mean, if you ever actually used the service, you'll know that both the Turk interface that you actually did work and the setting up the jobs was terrible anyways. It was always very difficult to use and you were, you know, they never made it easier over the years. So I don't know if they've ever really been investing in it heavily, but it's definitely been around for a long time.

[02:30] Justin Brodley: Do you think they used it to train their own models early on?

[02:33] Matt: Oh, I think it was very involved in Alexa training. Yeah.

[02:37] Justin Brodley: Yeah. I played with it. I got paid a couple of dollars. I was looking at like counting cells on microscope slides or something. I just figured I'd try something out. It was kind of mildly interesting technologically, nice idea, but yeah, no, no place anymore.

[02:53] Matt: Yeah.

[02:54] Justin: I mean, technically it predates AWS because November 20, 2005, AWS was SQS in 2020, in 20, sorry, not 2026, 2006.

[03:06] Matt: This is technically an Amazon service because it's Amazon Mechanical Turk. It's not AWS Mechanical Turk. You know, it's interesting because there's been other companies that have come up before AI really took over, you know, doing ML training, but did similar things. Mechanical Turk in a much more reliable, user-friendly way. And those companies, I think, are also sort of going down the same path of like, they don't have a business anymore. Um, so I'm not, doesn't surprise me terribly that this is going away. It's just, uh, it was always a fun idea that they created the Mechanical Turk. And then, you know, people who aren't familiar with the story of Mechanical Turk, uh, you know, is from ancient times where a man was hiding in a machine, uh, playing chess. Uh, and that was the, the whole joke. It was, it looked like a machine, but it was really a human. And so it's kind of sad.

[03:48] Justin Brodley: Automated with people.

[03:49] Matt: Yes. Automated with people, which is, you know, sort of Lambda Spackle is kind of like a good segue from there to, you know, where we are now, which is now everything's done by AI. So things are changing quite quickly. All right, well, let's move on to AI is how machine learning makes money. Anthropic has added richer admin analytics, model-level entitlements, and spend alerts to Cloud Enterprise, giving IT and finance teams more granular visibility into how Claude is being used across groups and individual users. For those of you who are familiar with the old way, thank God. This is so, so much is needed here. The updated analytics dashboard now breaks down usage and cost by SCIM group and user, with Claude getting dedicated tabs that estimate productivity lift, cost per commit, and annual value using adjustable transparent formulas. Analytics API lets finance and IT pull cloud usage and cost data into existing tools like Datadog Cloud Cost Management and CloudZero, with filtering by date range, team, product, or model, treating cloud spend like any other cloud cost line item. Admins can now set model defaults per role or product so routines or work does not automatically route to the most expensive model, and spend thresholds alert fire at 75% and 90% of org-level limits to prevent mid-task interruptions. I mean, you could put it in Datadog or in Cloud Zero or, you know, Cloudability, but you could also just have Claude write you a dashboard.

[05:07] Justin Brodley: Yeah, yeah, hey Claude, build this Grafana dashboard for me. I wonder how they estimate your productivity increase. That's kind of an interesting one, 'cause, you know, if I went by Claude's numbers, I'm doing about 3 years worth of work every other day.

[05:22] Justin: I like how it doesn't automatically route to the, most expensive model. Like that's a feature they had to put in there 'cause you know, everyone just defaulted to Opus or I guess Fable for the week. It's, it's out and everything else, you know? So I guess it's nice that they had to put that feature in 'cause try to help people out. Yeah.

[05:44] Matt: I mean, anything they can do to make it better, 'cause token maxing is no longer popular if you don't know. Suddenly, suddenly everyone's upset about people spending too much on tokens and how do they get ROI outta their costs.

[05:56] Justin Brodley: Training, I think it's what's needed, honestly. Yeah. Yeah, use the expensive model and it does a good job, but it costs more, or use a cheap model and iterate over it 10 times and it probably still costs more. So I don't know, I think some training's needed on how to actually use them properly.

[06:13] Justin: You mean give your entire company AI and say, go use it and maybe educate them on how to use it properly?

[06:21] Justin Brodley: Who's got time for that? We've already promised features to customers.

[06:29] Matt: Uh, well, if you have a need to do all that work while you're on the go, you might have been missing the fact that you couldn't use Claude Cowork on the web or on mobile apps. Anthropic is pleased to tell you that they're expanding Claude Cowork beyond desktop to mobile and web, allowing users to start tasks on one device and check on completion from another. Beta access rolls out several weeks, uh, over several weeks, starting with Max plan subscribers. Cowork enables Claude to work autonomously across files, calendar, email, messaging apps, and connected tools until a task is completed. Anthropic usage data shows over 90% of Cowork activity is non-coding work, with business operations and content creation making up roughly half of all usage. Uh, and they actually had a whole other blog post to talk about what they actually found out, because they took your data, they sampled it, and they figured out what you're actually using Cloud Cowork to do. And so they take data from May 11th through the 31st of 2026 across more than 600,000 organizations, classifying activity into a 20-category taxonomy of work tasks, with business process and operations tasks accounting for 33.4% of usage, including reconciling spreadsheets, building onboarding checklists, and consolidating status updates, while content creation and copywriting made up 16.4% of the work, covering drafts, slide decks, and proposals. Uh, software development represented only 8.7% of Cloud Cowork sessions, which makes sense because it's just used Cloud Code. And the data shows roughly half of cowork usage falls into connective cross-role tasks rather than job core functions, suggesting the tool is being adopted for coordination and information synthesis rather than specialized technology work. Anthropic gathered the data using a privacy-preserving aggregate-only analysis method, which capped an hourly sampling rate rather than a fixed traffic percentage. I mean, still kind of freaks me out that you were capturing all this data, you know, the tools I was using and the fact that I was accessing spreadsheets, even if it was privacy-preserving and aggregate-only.

[08:08] Justin: 'Cause no one's ever been able to get around privacy preserving and identify what people were doing.

[08:12] Matt: Never. There's a whole world of marketing people who can tell you exactly how to do that. You know, 3 signals and they can basically pin you down to any device in the world. So.

[08:23] Justin Brodley: Yeah. Yeah. That's scary. I like the back and forth between mobile. I think one of my pain points has been starting on desktop, not necessarily with CodeWorld, just in general, it could be just a regular chat, as long as it's got access to local tools, and then you sort of walk away from the computer, you go somewhere, and then it says, hey, I finished this thing, and you send it a new message, and all of a sudden, it switches the tool context to the mobile device instead of a desktop where it has access to all the files that it was working on. And so now, you're kind of in this weird limbo until you get back to your PC and sort of like reestablish the session there, and okay, now you can rewrite the files and access. So if it maintains the access to the the local tools, even if you're remote, then that's kind of neat. I like that.

[09:08] Justin: I know Cloud Code could do it before. I've only ever, like, I never got it to fully work. It was like with the /remote command. You could like then do it that way.

[09:17] Justin Brodley: Remote control works fine for Cloud Code, but if you're actually just having just a regular chat and you're having it—

[09:21] Justin: Oh yeah, that was useless.

[09:22] Justin Brodley: You know, use the file system connectors and things, it loses those connections, which is, it's probably a good feature. I mean, somebody can't remote control a computer if they get access to your Cloud session at the same time. You know, I, I'll trade the security for some convenience.

[09:36] Matt: I mean, I have a, I've used the Telegram feature a few times, which is kind of like their answer to OpenClaw. And like, that is terrible. So I haven't tried the remote because of how bad the Telegram interface was. Um, so I, if you're telling me remote's not bad, I'll check it out.

[09:50] Justin Brodley: Remote is just like, just like having a Claude Code web session, except the worker is your, your own PC at home.

[09:57] Matt: Got it. I have used the Ultra plan feature, which is kind of interesting because it, pushes all the planning up to the Anthropic website. And I don't really know what the benefit of having it run there is versus in the Claude desktop client, but it's been a few times where I've done Ultra Plan and it dumps you out into a web interface and does a heavy-duty planning session and you can edit it easily there. And then it pushes it back down to Claude Code.

[10:19] Justin Brodley: I assumed it was just so they could protect the intellectual property of their prompts and their workflows.

[10:25] Matt: That could be.

[10:25] Justin Brodley: And then now they've, now they're sort of hiding the, the thinking and summarizing the thinking instead of passing that back to help stop distillation.

[10:31] Matt: But yeah, I have noticed that they don't really show you a lot of the thinking, uh, trails anymore, which is a problem for some use cases that need high audibility, uh, because they need to see the thinking logic because then that's how they prove the data is valid. So some of that's— it's a hard problem to solve. All right, let's move on to AWS. Upgrading your EKS cluster with confidence now using Kubernetes version rollbacks. This is for EKS, allowing customer administrators to reverse a minor version upgrade within a 7-day window if issues arise post-upgrade, addressing a longstanding limitation of open-source Kubernetes where control plane rollbacks were not supported. The feature returns clusters to the previously validated production version rather than an emulated state, and it includes automated rollback readiness checks through Cluster Insights that flag node version compatibility and add-on dependency issues before proceeding. For EKS Auto Mode clusters, both the control plane and the managed nodes roll back together, with the process respecting existing pod disruption budgets to maintain workload stability. A cancel API is also available if you need to stop a node rollback mid-process and adjust your approach. Something that you constantly find that Claude does not build into its APIs. I always have to remind it, hey, I need a cancel method. Uh, the control plane rollback process takes roughly 20 minutes, similar to a standard upgrade, and the feature is available at no additional cost across all AWS commercial regions.

[11:42] Justin Brodley: That's not a feature for me.

[11:43] Matt: I'm still ECS all the way.

[11:45] Justin Brodley: Yeah, ECS fan. Like, I mean, why do you gotta pay like 10 cents an hour just for the control plane? So I would just of the value system.

[11:53] Matt: I mean, for people who really want, you know, Kubernetes, maybe they're multi-cloud and they need Kubernetes and Azure and AWS and GCP and Oracle and they move things around.

[12:01] Justin Brodley: Yeah.

[12:01] Matt: But yeah, if you just have an Amazon-only workload for containers, like just do yourself a favor, just use ECS.

[12:07] Justin: But I need Kubernetes. Everyone says you need Kubernetes. I need to run my RabbitMQ session in Kubernetes. Come on.

[12:13] Matt: I've got, I don't think I've ever had Claude recommend Kubernetes to me on, or EKS over ECS. So just, I mean, it always recommends Fargate. It's always pushing me to Fargate, but it's definitely not recommending EKS. EKS to me, which is appreciated.

[12:24] Justin Brodley: Fargate Spot though, right? Not just regular Fargate. Yes, yes. Yeah, I got some Fargate Spot stuff going on right now. Thanks, Claude.

[12:30] Matt: Yep. It's always Fargate Spot, like cheap.

[12:34] Justin Brodley: I really wonder, not to drag on about AI the whole show, Kubernetes is just such a massive piece of software. I think we're talking about AIs and how we have thinking and reasoning abilities and things now. Reasoning doesn't just mean logically working through a problem you give it in text. Reasoning can be any kind of workflow that needs to analyze inputs make a decision. And so at home, I've got containers running on a couple of machines now. I started off with K3s, which was okay, it wasn't bad. I did Longhorn and some other stuff. And I'm like, no, actually, why not just use AI for this? And so I've got Claude in a loop checking every 5 minutes, are my containers up? Where are the free resources? If they're not up, move them, put them, you know, redeploy them. I've literally got a container orchestration in in a paragraph of a prompt and a cron job?

[13:25] Justin: I mean, I have so many more questions.

[13:28] Matt: Is that token efficient? Like that doesn't seem very token efficient. Are you running it via a local model?

[13:34] Justin: So you don't care about the Docker?

[13:36] Justin Brodley: Yeah, no, I'm running local models. I'll start off with Claude, but it works just fine with Qwen.

[13:43] Matt: I've also just like, I mean, I just, maybe I'm lazy.

[13:46] Justin: I just, I guess I don't care as much about my containers at home. So if they die, I'll notice when next time I try to go pull up my recipe servers.

[13:55] Justin Brodley: You're just doing it for the sake of doing it though, just to prove that you don't need gigabytes of packages to orchestrate complicated things when you've got a massive model.

[14:05] Justin: A loop.

[14:06] Justin Brodley: Yeah.

[14:07] Justin: I have an infinite loop that will solve the problem.

[14:12] Matt: Yep. Amazon CloudWatch Pipelines now support processing and enriching OpenTelemetry metrics during ingestion, allowing teams to add business context tags, strip high-cardinality labels, and rename metrics centrally without modifying application instrumentation. This addresses a common pain point where customers previously had to build custom processing layers or change source instrumentation just to transform OTel metrics before storage, which added operational overhead and costs. Practical use cases include tagging metrics with team ownership or cost center data from sources you cannot modify, and reducing storage costs by removing unnecessary high-cardinality labels from custom workloads. Features available in all AWS regions where CloudWatch Pipelines and CloudWatch native OTel metrics are supported. I mean, they could have just fixed the issue where you can't add additional metadata to the tagging. That would have been a solution for this too. But okay, OTEL, let's do it that way.

[14:57] Justin Brodley: Yeah, I like that. I like injecting immutable things in there. That's really nice because right now with Durelic and things, it's all defined as tags in the config file. And people don't have to provide it or they can provide mistakes. So I guess that's, if anything, the best feature is injecting immutable data, as long as it doesn't get stripped out later, 'cause they're intelligent.

[15:24] Justin: Yep.

[15:24] Matt: I mean, my first OTEL experience has really been heavily tied into Grafana and using those tools, and I see the rough edges real quickly, but it's nice. It's great. I run it on my Synology. I, you know, get all my data from Claude Code. I feed it into the system. It's great. I love it.

[15:42] Justin Brodley: Is it just a common schema spec or like what, what, what makes it special?

[15:46] Matt: I don't know. Claude wrote it. I just made Claude write it. Nothing special. I said, I want Grafana on my Synology that gets all my Claude outputs. And so it shows me tool usage. It shows me token burn. It shows me cost estimates, how much cache I hit, et cetera.

[16:00] Justin: That's very cool.

[16:01] Justin Brodley: That's cool.

[16:02] Matt: It was a fun, fun side project to learn Grafana one day. And then I was like looking at the templating engine for the dashboards and I was like, this is terrible. I wouldn't wanna write these by hand. And thank God AI exists. That's how it basically went down for me.

[16:14] Justin Brodley: Yeah, I haven't used AI for it, so I haven't used it yet, 'cause I did the same thing and I installed it on the Synology and then opened it and thought, how, and how do I get my visualizations exactly?

[16:25] Justin: You can, I built a dashboard for, you know, my last day job and, you know, I built Grafana 10 years ago and it was crap, I'm sure at the time. You know, and now I was like, oh, I want this. Okay, now tweak this. Oh, and make this be a dropdown list to support multiple different regions and dynamically pull all this data. And it was all doable, but I think that what I iterated through in the span of 8 hours probably would've taken me multiple months of engineering to like figure out the dynamic syntax of all this, 'cause it wasn't possible, you know? And that's where I think people had full-time Grafana admins back in the day doing this. And now you can iterate over these things. Like, I was working on Home Assistant, the same thing. It's so complicated, but I can just tell AI, no, make this dashboard, change this, or how, how do I connect this IoT device to it, and go.

[17:19] Matt: This is probably the reason why we'll never, um, you know, we'll never truly have a, you know, replacement for Kubernetes, because the issue is all these complicated things just got really easy with AI. And it's like, you know, there are competitors to Home Assistant, but Home Assistant's been kind of a stalwart of home automation forever. And so in this case, you know, the customer is basically, you know, some of them who don't like it have gone to those other tools, but those tools don't have the popularity. So now it's like, well, Home Assistant doesn't have to improve because everyone can just use AI to make it better for them. So, yeah. And if you're, and I assume if you're a heavy duty home automation person anyways, you're, you're gonna definitely be using Claude or some other coding tool like Codex or whatever, cuz Home Assistant is not for the weak.

[17:58] Justin: So, well, that's why I've had the container in my house running for probably multiple years. I would say 3 years now. And I think I connected like 2 or 3 things to it and then the container would die. I, you know, tried to update it, like never actually used it. It was just like collecting metrics. And then this is the first time I'm now actually using it cuz I was like, oh, notify me when the lights are on downstairs and automatically turn them off. You know, and things like that. Like just doing basic things that like I could never figure out before, or I never had the patience for, which is probably what the real issue was.

[18:30] Justin Brodley: That's cool. I plugged one of those kilowatt things into my dishwasher and now Home Assistant watches that and it sees when the power ramps up and the dishwashing is going on and it waits for the heat cycle and then it sends a message to say, okay, ready to put the dishes away now. And so I send that to my son and he's responsible for the emptying of the dishwasher when Home Assistant tells him to. You better get used to the robot overlords giving you instructions, I think.

[18:54] Justin: Yeah, well, I get more around the power of my house too, but I cross those bridges where I have money to buy all the extra tools for it.

[19:02] Matt: Well, since Jonathan and I are both active ECS users, we will appreciate this next feature. Amazon ECS now offers a live deployment timeline in the console that shows each deployment phase, service events, and task launch and termination progress with automatic refresh, eliminating the need to piece together deployment status from multiple tools. Feature includes real-time circuit breaker status with live task failure proximity and threshold tracking, plus health checks at both container and Load Balancer levels, giving teams earlier visibility into deployments heading towards failure. Failed tasks surface directly in the deployment timeline with diagnostic context and deep links to AWS CloudTrail, reducing the manual investigation work typically required to trace the root cause of a deployment failure. It's available at no additional charge for all ECS services using the rolling update deployment type across all AWS commercial regions and gov Cloud.

[19:45] Justin Brodley: Yeah, that is nice. The thing that bugs me about ECS is he, like, he never should— did it fail? And it hasn't told me yet. Or is it just being randomly slow for this particular deployment? Because sometimes it's like 4 minutes for a container to come up and you're like, well, are you broken? Refresh, refresh, refresh.

[20:00] Matt: Is that a Fargate problem for you or is that an ECS host-based problem?

[20:04] Justin Brodley: It's a Fargate problem.

[20:05] Matt: Yeah, that's a Fargate problem. Because I use ECS hosts.

[20:08] Justin: Yeah.

[20:08] Matt: And the Fargate premium tax doesn't make sense to me for it, because I ran it like— The Cloud Pod website is a static container that runs all the time. I don't have that problem on ECS hosts. So maybe, maybe look at some spot instances for ECS hosts.

[20:21] Justin Brodley: Yeah, I thought about doing that. It's just, I spin the whole thing up and down when I'm working on it, when I'm not working on it, then I kind of looked at the cost model, some of the stuff, it doesn't make sense to keep a larger instance because I need to burst into like 2 or 4 gig to do a workload and then down again. So it's all kind of mapped out nicely, but yeah, it is a lot slower than I would like, really. I kind of want something as fast as Lambda, but as long-running as Fargate.

[20:49] Justin: Okay. Well, is that that Lambda VM that we talked about last week?

[20:53] Matt: Yeah, that's true. That's available.

[20:55] Justin Brodley: Well, the capacity providers for Lambda. Yeah, I looked into that because I really wanted to run serverless functions on a GPU instance. Because, you know, building a podcasting site, I wanted to do Whisper transcription. And Whisper on Lambda without a GPU, it takes about the same time as the LLM. You know, I only got one call. Thanks, Amazon, for not increasing my limit there. But it takes about 34 seconds to transcribe 30 seconds of audio. Whereas if you run it on a GPU instance, it takes like fraction of a second to transcribe 30 seconds of audio. So it was gonna be faster to spin up a Spot instance with the GPU, run the workloads there, and then tear it down than it was to use Lambda. But they charge a fee on top of the compute though. I'm really disappointed about that. I thought it was just you pay for the EC2 and then they point Lambda at it or your pool. But they charge like a 15% management fee on top of the instance cost, which is a bit disappointing.

[22:00] Matt: Indeed. Enforce zero data retention now on Amazon Bedrock with Bedrock projects and service control policies. So Fargate tackles a real compliance headache where some new Bedrock models like CloudTable5 require sharing data with third-party providers. So AWS added tools that organizations centrally block that behavior across every account, no exceptions. The key concept is that your configured retention mode acts as a ceiling, not a floor. Setting an account to allow provided data share doesn't force sharing on every request, Models support zero retention and still operate the way, or regardless of the account level setting. Bedrock projects available only on the Bedrock mantle endpoint. Let teams isolate workloads with different retention needs in the same account, so a research team can experiment with data sharing models without production project stays locked to zero retention. For airtight enforcement, service control policies override even account administrators and root users, blocking any attempt to enable data sharing organization-wide. Yeah, I mean, Fabric created a problem and they solved it. Thanks.

[22:52] Justin: Mm-hmm.

[22:53] Justin Brodley: I feel like, you know, the past couple of years have been AI stories just dominating. I think data stories are going to be dominating the next couple of years. It's always going to be about access to data, sharing data, and tracing where it's gone or what it's been doing. That's my prediction for—

[23:09] Matt: Well, I mean, I think we're trying to use human concepts of least privilege and role-based access and ABAC access to give it to agents. And so you basically put a lot of context overhead on top of agents to deliver this data. And I think what you're going to find is that we're going to have to rethink the paradigm for AI access so that the data itself is where the data retention, you know, the data access policies live. And then AI just honors that. I think that's what we have to get to to solve this for reals. But how we tackle that, I don't have an answer to. If I did, I'd go get a startup right now because I think that's the next step. AWS Security Hub now extends monitoring to Microsoft Azure resources, covering VMs, Azure Container Registry images, Function Apps, and Azure identities alongside existing AWS coverage, giving multi-cloud customers a single console for security posture management rather than separate tools per cloud. The service checks Azure resources against CIS benchmarks for Microsoft Azure Foundations and service misconfigurations, internet exposure, and vulnerabilities using the same finding format automation workflows as AWS findings. Including existing EventBridge integrations for automated responses. A 30-day free trial is available for Azure Monitoring, starting with customers to create the Azure integration. After the trial, pricing matches the cost of monitoring equivalent AWS resources. So this is an S2 announcement because both GCP and Azure have done this forever where they'll access all of your AWS security data because they believed in multi-cloud before Amazon did. So congrats Amazon for joining the rest of us.

[24:35] Justin Brodley: I wonder if they'll take a peek at the security posture of other people's clouds and say, hey, you know, our cloud, on average our cloud is more secure than their cloud.

[24:42] Matt: I mean, I don't think they needed Security Hub to do that.

[24:44] Justin Brodley: No.

[24:45] Matt: I think they've been able to do that way before.

[24:47] Justin: I was gonna say the CISO of Amazon said Azure's not good. So, you know, they already have the data to back them up there. I mean, it is nice to kind of have this consolidated point of view. I know that when I looked at it on Azure for AWS, it wasn't as comprehensive as like Amazon's own one, was what one was. So, you know, there's a little bit of nuances there that it's not going to be as good for the one that isn't the primary cloud for.

[25:18] Matt: All right, let's move on to GCP. AlloyDB now includes 3 new general available AI functions: ai_summarize, ai_aggregate_summarize, and ai_analyze_sentiment, which allows developers to run sentiment analysis, text summarization, and grouped summarization directly within SQL queries without building separate data pipelines. Smart batching for AI functions reduces latency and cost by deduplicating prompt overhead across rows, with Google reporting up to 2,400 times performance improvement, processing around 10,000 rows per second of AIIF and AI-RANK, currently available in preview. The optimized AI functions feature trains a lightweight model proxy model on your existing embeddings and data, allowing decisions to be processed natively inside the database rather than calling an external LLM. And core AI functions AI Generate and Rank, AIIF and AI Forecast are now generally available, making AlloyDB a more complete option for teams. And all I can say is this seems like the most expensive, dumb place to put AI that I could ever think of. Like, why put it— why put it in such a slow place?

[26:15] Justin Brodley: No, no, it's so cheap though. That's— this is the— I'm gonna— yes, yes, but no. Because when you— when you— when you're doing it in a database, instead of making single calls from your— from your whatever service, you make a single call, you pay that whatever latency costs, and then the the time to run the inference. But if you can do it, if you can run in batches, and GPUs, big GPUs love running things in batches, load all the data in and you process everything in parallel, because you know, they've got 26,000 cores to do the work with. Like the cost to Google to run inference across 10,000 rows is more or less the same as it is to run one. And in perspective, like my GPU that I have running QuenCoder right now, if I run a single instance at a time, I get about 180 tokens a second, which is really impressive. But if I use Claude Code and Light LLM and it batches those requests and then runs inference on them all in one go, I get like 5,000 tokens a second out of the GPU. It's an amazing uplift in performance by batching.

[27:24] Matt: I mean, so I agree that it's not a huge, you know, huge cost. I mean, I still feel like if you had to do a SQL query first and then feed it into this function inside the database, like it's still like, it's a slow process, is it not? I mean, like, I get what you're saying. I just, I guess maybe I need to think about it more, but yeah.

[27:43] Justin: I just don't like the idea of my database, and maybe I'm misunderstanding this, but like my database is running AI queries for me?

[27:52] Matt: Do I have GPUs now on my database servers?

[27:55] Justin: Am I putting business logic into my SQL database, which is something I avoided like the plague to not have stored procedures? I guess I'm missing something here.

[28:05] Justin Brodley: I mean, it could be a database of customer service calls or messages from support chats and things like that. And so, once a day you run a batch analysis across the entire set. Or whenever the customer calls, you want an analysis across there and their previous 1,000 conversation interactions with you or something. And it very, very quickly can score everything and give the customer service person, who's probably going to lose their job shortly anyway to AI, give them some kind of indication about you as a person. I don't know. I mean, it's a good place to put it.

[28:39] Matt: But why does it need to be in the data fix?

[28:42] Justin Brodley: I mean, that's where the data is. Like, where else are you gonna, I mean, you either take it out and move it someplace else and then do the inference and then put the result back or just make it native. That makes sense.

[28:53] Justin: I just, I feel like the database here is an all-in-one printer. It does nothing well, you know, at that point. Like, you're adding things onto it that shouldn't be there.

[29:03] Matt: I mean, I was at a company where they put session store onto SQL Server and then crashed SQL Server trying to look up every web session call. So like, I'm a little, I'm a little gun-shy on putting things in the database that don't belong in the database. So like I sort of agree with Matt on this, but I, I see Jonathan's point too about like, well, you know, if you have all the data there already doing a quick LLM type lookup against the GPU on the, on the server, that might be beneficial to you. So I could, I see both arguments, but I also like, this has to be a very limited use case. I think in most architectures, like it shouldn't be your business logic as Matt's concerned about.

[29:37] Justin: Because that's what a developer's gonna do. The developer's gonna just throw everything in the database in a stored procedure or LLM call, and now like your database, which already is an expensive resource that isn't horizontally scalable, is now gonna be doing more things and there's just gonna be production outages because of it.

[29:57] Justin Brodley: Now, I think when they say you don't need to call an external LLM, I mean AlloyDB is certainly calling out to inference somewhere else in Google. It's not running on your instance. It's just taking a batch, doing something, doing an insert back again.

[30:12] Justin: Okay. I just don't like business logic in the database. Like to me, the database is storing the data.

[30:18] Justin Brodley: I mean, I think it's gonna have to blur though. It really is. It's gonna get very blurry. Do you want, I mean, they don't say where you store the prompts for summarization or things like that.

[30:28] Justin: In the database, of course. It's gonna be a different SQL call. Probably.

[30:31] Matt: Select star from LLM.

[30:33] Justin Brodley: Probably a special table.

[30:34] Justin: Yeah.

[30:35] Matt: Right.

[30:35] Justin Brodley: I don't know. But it is, it does kind of stink a little bit of stored procedures for AI.

[30:40] Justin: Right.

[30:41] Justin Brodley: But it's needed. It's needed. Everyone's going to be doing this. Guaranteed.

[30:46] Justin: I'll call you on my first outage because of this.

[30:49] Matt: I hope not to.

[30:50] Justin: Well, I'm just going to yell at you.

[30:51] Justin Brodley: Okay. I'd love to see what prompt injection protection they have though. That's going to be hilarious.

[30:57] Matt: I'm sure they haven't thought about it. That's how it works.

[30:59] Justin: And then tracking the logic of it, of, you know, how it, you know, what's the reasoning of it and storing all that, which I'm sure would store right back in the SQL database.

[31:09] Matt: Yeah. Well, Google Cloud has published a reference architecture for offloading LLM KV caches to manage Lustre as a shared external file system tier, targeting enterprises running long-context inference workloads that exceed local CPU, RAM, and SSD capacity on multi-node GPU clusters. Jonathan, you're going to have to explain that to me in English because I don't know what I just said.

[31:26] Justin Brodley: I mean, if you're Anthropic and you're running Claude and somebody is working in the session, then they go away for an hour and then come back again, they're not going to keep that KB cache on that particular node where it's kept. So they need a place to put it. And so I guess they're talking about a shared file system so that all these nodes can save the KB caches for when the person comes back the next time. And they may not hit the same host because it may be busy when you arrive, but you can still get the cache. I don't think there's anything special in the, in the story itself. It's just saving the cache to a network-attached storage device. But yeah, AI.

[32:02] Justin: Okay, so I went from 10% understanding to about 70% understanding with that. I'm moving up in the world.

[32:10] Matt: Yes. Let's move on to Matt's favorite cloud, Azure.

[32:12] Justin: Yo, always fun.

[32:15] Matt: Microsoft is launching the Frontier Company, a new operating business backed by a $2.5 billion investment embedding 6,000 industry and engineering experts directly at sites to co-design and deploy AI systems tied to measurable business outcomes. The offering goes beyond traditional consulting by combining deep industry knowledge, change management, enterprise AI engineering into a continuous improvement loop. The early deployments at organizations like LSEG, Lambda Lakes, and Novo Nordisk. A core principle of the model is that customer data and IP are never used to train models in a way that benefits other organizations, addressing a significant concern enterprises have raised about AI vendor relationships. The platform supports model flexibility across OpenAI, Anthropic, Microsoft AI, open source, and specialized industry models, meaning customers are not locked into a single provider. Microsoft extending this through their global SI partners, including Accenture, Capgemini, Ernst Young, KPMG, and PwC, which suggests the primary delivery channel will be partner-led rather than direct Microsoft engagement for most customers.

[33:12] Justin Brodley: I mean, the, the cynic in me thinks that they're doing this as a separate company so that in 18 months' time when AI can replace those forward-facing engineers, that they can lay them all off without it damaging Microsoft's own reputation as an employer. This has got to be a, this has got to be a really short-term thing. This has got to be information gathering. I mean, they say it's customer-focused, but I don't believe for a second that they're not going to take lessons learned back and use it to improve products, which will consequently help everybody. But at some point, we don't need people for that anymore.

[33:45] Justin: They built a funding program for their partners. It's all they did. Here's a bucket of money for partners to go out and help our customers. I feel like that's all they really did here.

[33:54] Justin Brodley: Yeah. And Frontier Company, Gold Bull Room, whatever.

[33:57] Justin: Yeah. Just, and it's interesting what you said is like, you know, all these massive companies have had layoffs in the last couple years and they're starting to take hits from it. So is this a way to kind of, it's a shell company essentially for to hide the layoffs and negative press of it.

[34:14] Matt: Oh yeah, I think it's that. I think it's, you know, you don't have to give these people stock 'cause it's not a publicly traded entity. It's just an investment vehicle. So you don't have to get them triple stock. And, uh, he said, you know, there's also the things around the fact that you are competing with all the cloud hypervisors who are all doing FDE or forward deployment engineers. And so Microsoft doesn't ever do anything directly because that's not their model. So they're gonna want to adopt partners to do that. And I'm sure the partners looked at and said, well, we don't wanna take the risk to undermine our own business. So if you create this third-party thing that we all invest into and have access to, then therefore, yes, you have the ability to get rid of them later when you don't need 'em anymore. 'Cause AI caught up or you're not necessarily, you know, hurting your rates when you're at something like Accenture or Capgemini.

[35:00] Justin: Oh, I read this originally that Microsoft put all the money in, but you think all those companies, you know, put some money in? Now I'm kind of curious.

[35:11] Matt: Hmm. I'm not sure if I'll put any money in or not. I'll let you look it up while I move on to the next story. So in case you don't have enough chaos from your Azure deployment, you can now deploy Azure Chaos Studio Workspaces in public preview, offering a scenario-based approach to chaos engineering that tests real-world outage patterns like zone down, DNS outage, and SQL failover rather than isolated faults. General availability is targeted for late 2026. The service addresses a common gap in cloud resilience. Many outages stem from misconfigurations, rather than platform failures, and Chaos Studio helps reverse these before they hit production. This aligns with Azure's shared responsibility model for reliability. Workspaces reduce setup friction by using a managed identity to auto-discover resources in a subscription or resource group and recommended applicable test scenarios with a library of curated scenarios covering compute, database, DNS, identity, cache, and messaging failures. It's integrated with GitHub Copilot and a new GCP server lets engineers or AI agents provision Workspaces, run drills, and pull correlated Azure monitor signals directly from tools like Copilot, Claude, or Cursor. Without needing to script against the Chaos Studio REST API directly. Each test runs, it generates a structured drill report detailing injected faults, affected resources, and recovery timelines, useful for audit evidence, change tickets, or service health reviews.

[36:22] Justin Brodley: By the way, I missed the lightning round. This would just have been gold. Plot twist, this service has been in limited availability and accidentally turned on for every customer for the past 3 years.

[36:35] Justin: With all the SODA outages and like blips they've had, it would not surprise me.

[36:40] Justin Brodley: Yeah, hopefully their first customer's GitHub, right?

[36:44] Justin: Well, it integrates with GitHub Copilot, so maybe that's how they're testing it.

[36:48] Matt: Maybe.

[36:49] Justin Brodley: Nah, GitHub's been so bad lately, I've set up GitLab, which is not as good as GitHub.

[36:55] Justin: But does Copilot help you configure it so it's good enough?

[36:59] Justin Brodley: Copilot, Microsoft's dead to me.

[37:03] Matt: I kept waiting for, yeah, they announced like a month or two back they were changing their licensing terms to consumption-based, and I wanted to try out their Copilot PR reviewer, but you couldn't sign up for it for months. So then I just built my own using open source. I found open source solution for this and I was like, I'm just gonna create that. And so Matt and I have been testing it with Bolt and I'm pretty darn happy with it. It works pretty well. So far.

[37:24] Justin Brodley: Would you have a, is it GitHub Actions that fires when the PR opens?

[37:27] Matt: Yeah, it's a, it's, yeah, it's a GitHub Action. It's a, it's called PR Agent. And basically I set it up using Claude and said, make this work. And it created a thing and literally during GitHub Actions pulls down the container that it runs and runs a bunch of checks. Then outputs a bunch of data back, uh, to the PR comment, and then you can paste that into your Claude session if you want to, or have Claude get it itself with GH. Oh, nice.

[37:48] Justin: Yeah. I just have the AI bots fight each other at this point. See who wins.

[37:53] Matt: Yeah, exactly.

[37:54] Justin: If you looked at that PR from today, Justin, there's like, it's going back and forth and it's like, you are wrong. This is why it caught. It's like, okay. Though it did catch a real one today.

[38:03] Matt: I, I did initially just use Claude cuz that's but it is a little bit Claude watching Claude, which is funny to me how many times Claude catches his own mistakes, which is nice. But I do, I do intend to add, uh, ChatGPT to it as well. So I get their input as well.

[38:15] Justin: I like the fact that you broke it up into like test casing, like, you know, like you broke it up into a couple segments too. So it's more likely to catch it. It was like, you know, unit tests that you potentially missed or, you know, actual code gaps. Right. You know, and then I think there was a third one and the diagram's pretty nice cuz then I actually learned how Claude programmed my stuff.

[38:35] Matt: Yeah, the diagrams are super nice. Um, the one thing I, I, you know, again, we're on a side trail now, but, uh, yeah, the one thing, it, it gets a little bit anal about the test. Like, well, you have a test for the thing, but you don't have a negative test, and you don't have a negative negative test. I'm like, okay, we're getting a little bit picky at this point. Like, okay, yes, I had the test, so then I had a negative version of the test. I don't need to now negate that test with the— this, like, I need to get a little bit less stringent about that part. I gotta do some tuning on that one. 'Cause I'm like this, I get what you're trying to do, but it's too much. It's too much.

[39:06] Justin: It's just Bolt. Like it's not a real thing. Like it, this is just a tool for us.

[39:10] Matt: Like, exactly. Like you're, you're a little too, a little too anal about this. Uh, you did find the article about Amazon launching their own FDE program as well. I see.

[39:18] Justin: Yeah, I dropped it in our show notes where it was 2 days after. I couldn't find who was invested in it though. I might have gotten sidetracked by shiny red object syndrome of the known as the podcast. But I didn't realize that Amazon launched one 2 days before. I dropped that in the show notes too, weirdly under the Azure section.

[39:36] Matt: Yeah, and I think we talked about not talking about it because I realized it's kind of boring, but now everyone's doing it, so I guess we should. This is just them though, uh, it's just Amazon pouring billions of dollars into this.

[39:45] Justin: They're just moving money and moving engineers from one side to the other.

[39:48] Matt: Yeah, it's all just a shell game, it feels like.

[39:52] Justin Brodley: So customers paying for the FTEs, or, or the vendors paying to place FTEs in customer orgs?

[39:58] Matt: My understanding, it's the customer paying for it, but then the vendor makes the money, you know, makes money because they're using AI tokens. So I think there's, there's typically some amount that Anthropic puts towards it when they do theirs, but like it's, it's shy. It's nothing compared to what the customer's spending on tokens.

[40:15] Justin: Yeah, I'm sure if you at your day job, you know, went and got one of these, you'd be, by the time you're done, you're spending 10x tokens of what this person told you you do it a month.

[40:24] Matt: Mm-hmm. So because you've proven the reliability of your Azure environment with Chaos Studio and you have your workloads running there, you now want to back those systems up using something reliable and safe. And so Microsoft Entra Backup and Recovery has reached general availability, shifting the recovery model from point-in-time restore to a comprehensive tenant recoverability strategy. This feature is included with Entra P1 and P2 licenses, meaning organizations already on those tiers get this capability without additional costs. Lowering the barrier for adopting a more resilient identity recovery approach. Identity systems like Entra ID are foundational to enterprise cloud environments. Yeah, you're really screwed if this goes down. And so recovery capabilities for tenant-level configuration objects reduce risk from accidental deletions, misconfigurations, or malicious changes potentially caused by your chaos engineering studio. So you now have the ability to back it up before you mess with it.

[41:11] Justin Brodley: How did they not have this before?

[41:13] Matt: Right?

[41:14] Justin: Did they have third parties is where they played you to before? You know, but yeah, it was, we talked about it, I think when it was in private preview or one of those a long time ago. Took me, it took me, it feels like it took 'em a really long time, but my sense of time isn't real anymore. So what do I know?

[41:32] Matt: Yeah, I think we talked about it when it went into early access and we said it was ridiculous. Now I think it took like 6 months for it to get GA. So very classic, um, Microsoft.

[41:39] Justin: I mean, good news, if you're paying the premium license, which isn't cheap, They give it to you for free. If you're not paying the premium license, don't worry, your environment's screwed and that's okay.

[41:53] Matt: Uh, what a mess. Microsoft released Azure Linux 4.0 as a downloadable ISO that can be installed on bare metal servers and VMs outside of Azure, marking a shift from internal cloud plumbing to a standalone server distribution. It's based on Fedora using RPMs and ships with a hardened Linux kernel 6.18. Support differs significantly by deployment. Running Azure Linux on Azure comes with formal SLAs, CVE patching, and integration with Defender for Cloud, while on-premises or bare metal installs are community supported only with no official Microsoft backing. Linux has been the dominant OS on Azure for nearly a decade, and so this move formalizes and extends the reality by giving Microsoft its own distribution to complete with AlmaLinux, Rocky Linux, and other enterprise Linux options. I mean, at the end of the day, is this thing going to replace Windows Server? Probably not. ZDNet seems to think it will, but, uh, I don't think so. Uh, and the only thing I'm gonna use this for is probably to go into my WLM on my Windows box when I rarely need window, you know, an Ubuntu prompt on my Windows box.

[42:49] Justin: Oh, you think they're gonna swap it on the Windows for, with, for WSL to Azure?

[42:54] Matt: Yeah, you can use this as one of the WSL images.

[42:56] Justin: Oh, okay. Yep. Everyone I know uses Ubuntu. I just assumed that was the only one you could use.

[43:01] Matt: Oh no, there's a bunch. Ubuntu, you can do Debian, you can do Arch. Yeah, there's several. I was looking at it the other day because I had just upgraded it and I was like, what are the new ones nowadays? And I thought we had Rocky Linux actually, because I like to do more with Rocky, but I saw Azure Linux was available in the latest update from WSL.

[43:16] Justin Brodley: So why, why though? Like why, why not just use Fedora? Like what are we putting?

[43:23] Matt: What's the point? Is Fedora even really still a thing? I mean, it's based on Fedora, but I thought Fedora died.

[43:27] Justin: No, CentOS that died.

[43:28] Justin Brodley: No, Fedora's still around.

[43:31] Matt: Sentos.

[43:31] Justin: I had the same thought process when you were reading the article. I might have thought about looking up and I finally remembered.

[43:39] Matt: Yeah, Sentos died. That's right. Which I'm still bitter about.

[43:41] Justin: I mean, I know that more and more of the internal services are based on Azure Linux. So, you know, the Load Balancers, the APIMs, a lot of those services are starting to be more and more based on those services. Now I have other opinions about that, but we'll bypass those because of NDAs.

[44:04] Matt: Fair enough.

[44:04] Justin Brodley: I mean, I think the cloud vendors do a great job of making you feel like, oh, it's some magical service. It's an Elastic Load Balancer or it's something else. Like, you know, under the hood, it's just some clever scripting with Ansible or something else and HAProxy.

[44:19] Justin: On Azure, I believe, yeah.

[44:20] Justin Brodley: Or pick a— or nginx. No.

[44:22] Justin: Well, Azure, you're on the Load Balancer, you still specify the number of servers you want. That's not elastic if I have to tell you how many servers I want. I got my app—

[44:33] Justin Brodley: There's no magic.

[44:34] Justin: —upgraded with Kubernetes for the day, we're good. Yeah.

[44:39] Justin Brodley: I mean, everything's just a service running on a machine at some point.

[44:43] Justin: It's someone else's servers, that's all it is.

[44:45] Justin Brodley: They've done a great job of making you feel like it's some special thing that they do. Orchestration. It is a special thing that they do. I, I appreciate it a lot.

[44:53] Justin: But we should find that, um, I think Jonathan wasn't here the time we talked about the guy that went on like a 10-page rant on Substack about Azure. Oh yeah, I would be curious to see here, uh, Jonathan's opinion about it. I'll have to find it from our show and send it to you.

[45:09] Matt: We'll have to dig that back up and have an after show on that someday.

[45:11] Justin: That's pretty much what he goes through, which is like Azure is just people logging in and manually doing stuff on all these systems, and it's just, you know, Ansible or whatever it is, you know, under the hood of them running it to make these systems work.

[45:27] Justin Brodley: Yeah, I interviewed somebody who used to work at Oracle, and they told me the same thing about their cloud. You know, they got, you know, a team of people logging into Load Balancers, debugging things, and, you know, it looks like it's all automated in the console, but actually there's a whole bunch of people Actually making it tick.

[45:45] Matt: Yeah, and Turk sounds like fun, but so Oracle's gonna go down now. Yeah, Oracle and Turk. All right, well, moving on to emerging clouds, Cloudflare has a bunch of updates this week because it's their Content Independence Day. And if you remember last year, they had a big huge thing about how they blocked all AI training crawlers by default for new domains. They built the jail and they do all the craziness and they've made my life much harder with Vault to get in the show notes. Because you run into sites all the time that are being blocked by Cloudflare bots controls, which is silly for your blog where you're trying to announce new features, vendors. But, you know, because it's Content Independence Day week, they have several new features, including expanding its AI bot controls beyond a simple block all toggle, introducing 3 distinct categories for customers, including free tier search bots that index content for later retrieval, agent bots that act on behalf of users in real time, and training bots that absorb content into model weights, which is kind of nice, I guess. They also have a new research program they started to help AI search engines identify fresh, high-quality content without redundant crawling. Uh, their data shows over 50% of crawl traffic from good bots refetches pages that have not changed, creating unnecessary costs for site owners. And which makes sense, because if you're on Cloudflare Workers, you're paying for all that access. And then finally, they're launching a monetization gateway, very similar to Amazon, that lets customers charge for any asset behind their network, including APIs, datasets, web pages, and MCD tools. Using usage-based pricing enforced at the edge rather than at the origin. System is built on X.402, an open protocol that uses the long-dormant HTTP/402 SAS code to handle payment negotiation in line with standard HTTP requests with no redirects, no account creation, and settlement in stablecoins like USDC in under a second. So yeah, if you're interested in what they're doing, we link to all the different articles from their Independence Day. That's cool.

[47:27] Justin Brodley: I know Amazon tried a long time ago with the, you know, request to pay for S3 access.

[47:32] Matt: Oh, they just did it. They just did it. 2 weeks ago here on the show, we talked about it. They have a new WAF capability to make you pay for bot access to sites now as well. So they just did it again.

[47:42] Justin: I thought Cloudflare had that last year. They announced that too as part of the Maze Boxing.

[47:47] Matt: Not the monetization. I think they said they wanted to do something around monetization, but this is— they've actually now released it.

[47:52] Justin: Oh, I thought they said they were doing it.

[47:53] Matt: So that's my fault. Eh, it happens.

[47:55] Justin: It's just going to be bots paying each other bots at one point.

[47:57] Matt: Exactly. All right, gentlemen, we made it to another week of cloud in the books. So thank you very much. We'll see you guys next week.

[48:06] Justin: Yep. I'll see you later. Bye everyone. Another week of cloud news wrapped up. Bolt will collect the news. Justin will get the notes. Jonathan will write some code. Ryan will watch the perimeter and Matt will reluctantly watch Azure till next week for AI, Amazon, Google Cloud and Azure. And hey, maybe even Oracle, who knows? Check out thecloudpod.net for our newsletter. Join our Slack, message us on socials, or leave a review.

[48:39] Matt: All right, I have an after show for you. Sony, the maker of the PlayStation, is announcing it will stop producing physical PlayStation game discs in January of 2028, citing that digital downloads now account for 78% of full game unit purchases in their most recent fiscal year. The shift completes a move to licensing-only model, which is a distinction worth noting. Customers purchasing digital games are buying a personal, non-transferable license rather than owning the product outright per Sony's own terms of service. I mean, the reality is they never owned them outright, even on a disc, but, uh, at least the disc was transferable. Xbox tried to do this a few years ago as well when they released the new Xbox One, and there was much pushback on that. But, but even then, I think, uh, the majority of the Xboxes sold today are discless. Versus the disk version.

[49:23] Justin Brodley: Yeah, it's a bit disappointing. It makes sense, but it's a bit disappointing, especially as there's such risk in owning digital assets. It's tied to an identity, which is easy to steal, easy to lose. And games aren't cheap, $65, $70 for some games. You have an account that's 10 years old, you may well have $4,000 or $5,000 invested in there with in-game purchases and other things. And all it takes is for somebody to steal your credentials and take your account, and it's all gone. Mm-hmm.

[49:58] Matt: Yeah. And Sony's not known for their, uh, overly secure networking.

[50:04] Justin Brodley: No, PlayStation Network is, is, is not, has not had a great track record, but then nor has Xbox, so.

[50:11] Matt: No, neither one have had a great track record.

[50:12] Justin: Wasn't Sony the one that installed a rootkit on everyone's Windows box back in the day too? Yes. Yeah, I think they were. Yeah, just want to make sure we're talking about the same company.

[50:22] Matt: And they had that, they had that issue, you know, after they, you know, they had the movie making fun of Kim Jong-un where they ended up, you know, getting hacked and all their movies got released to the internet, etc.

[50:32] Justin Brodley: So yeah, it's, I mean, it's also interesting that, you know, Sony being one of the largest manufacturers of CDs and DVDs over the years as well is, uh, is cutting the, cutting the ties with their own media. Sony and Panasonic.

[50:47] Justin: I assume they just are slowly trying to kill that product line. So this was probably their biggest purchaser. Yeah.

[50:55] Justin Brodley: I mean, it means you can sell the same thing to, um, you know, a family may need to buy 2 copies or 3 copies and now you can't go trade it in at GameStop.

[51:05] Matt: So it's, um, so I mean, sometimes you get, you know, if you get like, there's things like Xbox Family, right? That if Xbox Family allowed you to share a game across multiple people, I'd even be willing to pay slightly more than that than the retail price, but they typically don't even allow you to do that today. And then there's a risk again of the family being hacked as I know Jonathan has some bitter experiences with. Yeah. But I mean, I mean, there's also like this kills GameStop completely, right? I mean, this is one of the big reasons why Xbox walked back their thing was that GameStop had such a big fight and people really loved GameStop at the time. I think that love has sort of died since the pandemic. I don't think GameStop has the same cloud that it used to, but I mean, like, there is, it'd be nice, like, hey, I bought a game and I don't like it. There's no recourse typically for refunds on Xbox or Sony networks that I'm aware of for a game that you don't like or, you know, it doesn't work where at least you gotta take the CD back to the store. Maybe you wouldn't get full credit cuz it was open box, but you at least got some money back to put towards another game. And now, now you're completely at the mercy of these marketplaces. So yeah, I, it does. It's definitely, it's what the whole world is doing. Everything's moving to subscription. No one owns anything anymore. Everything's a lease, everything's rented. I mean, it's a risk and it's definitely a challenge that I think future historians will look back at this time and be like, this is a really bad time in some ways.

[52:23] Justin: I mean, I really dislike the fact that everything is subscription-based. Like it really is a big pet peeve of mine is you don't own anything. It's, I pay $10 a month or on a good day, or $20 a month for most things at this point. But you don't get anything. You get it for that time period, you know? So you don't actually retain anything anymore. You mean even, what was it, the car, was it BMW or Mercedes? I think it was BMW that did the subscription for the heated steering wheel. I think it was.

[52:55] Matt: Yeah, I remember they, they tried that. It didn't roll very well, but it's kind of come back in different ways.

[52:59] Justin: I think they did it again and it stuck this time.

[53:01] Matt: Yeah, I think it came back for a different thing. Yeah. Yeah.

[53:04] Justin Brodley: Remote start, I think Chevy still gates behind a pay thing. Um, but that's 'cause like you gotta pay for the cell, the cell phone modem or whatever fee.

[53:13] Justin: Yeah. I think OnStar. Yeah.

[53:14] Matt: I mean like OnStar was one of those things. Yeah. It was, and then like those, those things have like upsell hotspots they can turn on that use cell phone data to turn a hotspot in your car. And there's all kinds of different uses that they do on those things. But that makes sense to me. You're paying for a consumable product, cell phone service, right? Just like you pay for internet anywhere else. But yeah, trying on my heated seats or my heated steering wheel.

[53:34] Justin: That's— You're right, they canceled it in February of this year. I looked it up 'cause I was curious.

[53:40] Matt: But they have, they do keep trying to do things like some companies are trying to think, you know, CarPlay and Android Auto behind a paywall. And some are, you know, trying to do different other things where, you know, like there was a, someone had an idea of like, well, you can sell a 6-cylinder car and you only activate 4 cylinders or something like that. And I was like, I don't even know how that works from like a mechanical perspective, but like how ridiculous, like it's already there. I mean, but like, you know, they do it on chip fabs, so why, why not? You know, you sell a Tesla, did it, a chip fab that had a, you know, bad, uh, core. Tesla did.

[54:10] Justin Brodley: I don't know if they still do it. They're doing the batteries, their standard interior, and they only activated like 3 of the speakers in the cabin. If you upgraded to the premium subscription, they turn on the other speakers.

[54:18] Justin: I'm like, it's just what they did. They did it for the batteries because there was a big announcement when there was a hurricane hitting Florida that they unlocked the extended range on all the batteries and if you were located in the state of Florida, or I guess surrounding states, they, they gave you extended range for your power of your car to help you get out of the state.

[54:39] Justin Brodley: I was like, I'm, I'm more okay with that though. I'm more okay with that because that's a business decision, says it's cheaper to, to manufacture one type of car and then charge less for it.

[54:48] Justin: But the same thing works with the speakers. It's cheaper just to shove them all in and have every car be the same than be where you have half the cars with and half the cars without. I have to put the right inventory through.

[54:57] Justin Brodley: It's not a subscription though. Oh, it was a one-time purchase. You buy the bigger battery or you don't, but yeah. Yeah. I don't know, it's still— But even then, like, even then, like, the extra battery capacity, but by them maintaining some buffer of the battery, it extends the lifetime of the battery. And so, by extending it during the hurricane to let people drive further, it sort of shows their cards a little bit and shows how they're managing to get batteries that last. 10 years or more. Agree. Yeah, I don't know. It's, uh, it's kind of annoying. I mean, even if, even if you do own disks and things though, you often have to have license keys for things and that all relies on the infrastructure for, you know, key management still existing years later.

[55:41] Justin: Do you think I can activate my Windows XP license that I'm sure I have somewhere? My key?

[55:46] Justin Brodley: You still got the volume license from, uh, from years ago?

[55:48] Justin: Sure. Somewhere. Some Windows ISO image I have saved.

[55:53] Matt: Yeah, I don't know.

[55:54] Justin Brodley: I mean, as, as a, as content creators, it makes sense that you want, you don't want your work to be pirated. So you want to have some kind of rights management built into it. And as soon as you let people have copies that could be copied, you're going to lose revenue. But at the same time, I think people should still be able to have an immutable license, which isn't something you can lose.

[56:15] Justin: So there are articles of how to activate your Windows XP created in 2026. It's crazy.

[56:23] Matt: It must be AI generated. I'm sure, but all right. Well, I think we wrap it up here, guys. Sad about lack of physical ownership. We'll see. See you guys next week.

[56:34] Justin Brodley: Yeah. See you later.

[56:36] Matt: Bye.