[00:00:07] Speaker A: Welcome to the Cloud Pod where the forecast is always cloudy. We talk weekly about all things aws, GCP and Azure.
[00:00:14] Speaker B: We are your hosts, Justin, Jonathan, Ryan and Matthew. Episode 280 recorded for the week of October 22, 2024. Evidently the Cloud Pod was always right Good evening. Full house tonight. Jonathan, Matt and Ryan. How you doing?
[00:00:29] Speaker A: Hello.
[00:00:30] Speaker B: Hello.
[00:00:30] Speaker C: I'm well back or almost back on my travels.
[00:00:34] Speaker B: I apologize since both Ryan and I are at a meeting in LA and on AirPods and terrible audio in our hotel rooms. So sorry about that.
[00:00:43] Speaker D: Looking forward to you guys dropping out in the middle as we're trying to talk. It'll be fun.
[00:00:46] Speaker B: Yeah, as long as wifi just totally fails us. Yeah, that can happen. It could. Yeah, we'll see. You guys can wake up from there. Well, we got a bunch of stories to get to this evening and so let's jump into it. So first up, AI is going great with the new announcement of the Claude 3.5 Sonnet and Claude 3.5 Haiku models announced this morning. Hot off the presses, the Claude 3.5 Sonnet delivers across the board improvements over its predecessor with particularly significant gains in coding, an area where it already led the field for anthropic Cloud 3.5 haiku interestingly matches the performance of Cloud 3 opus, the prior largest model, on many evaluations at the same cost and similar speed to the previous generation of haiku. Cloud 3.5 Sonnet also includes a groundbreaking new capability in beta called Computer Use, which is a terrible name available today as an API. Developers can direct CLAUDE to use computers the way people do by looking at a screen, moving a cursor, clicking buttons and typing text. Sonnet 305 is the first frontier AI model to offer the computer use functionality, although we do warn that the feature is still highly experimental, at times cumbersome and error prone, as well as things that are effortless for human like moving a mouse or scrolling, dragging or zooming are very difficult for the AI to complete.
The idea is to make Claude complete individual tasks without always needing to leverage an API. Think like click on a GUI or upload a file from a computer into an Excel document and these type of solutions that exist quite often in built in test environments for things like Sauce Labs or Browser Stack. So the ability for it to perceive interactive computer interfaces makes it available to start using AI for some of those testing use cases with your ui, which opens up a lot of interesting ideas. It just sounds like it's still a little buggy at this point, but something to look forward to.
[00:02:26] Speaker A: Yeah, it's been great for like QA testing or some automated QA testing.
[00:02:30] Speaker D: It would be almost like an interesting, like, user experience test too here. Claude, go figure out how to use this tool and see how it does.
[00:02:38] Speaker B: I mean, it's not that smart.
If they could do that, then I think we've probably reached an AI in Urbana that they can actually figure out your UI interface.
[00:02:48] Speaker D: Well, in theory, right, it's based on other UI models and UI systems and humans correlate in the same way. So in theory, if you go completely, really far out of the way. Now, I understand I'm stretching here. Then it's not going to be usable and the average person that just is going to go in quickly and go play, it's going to be confused.
[00:03:08] Speaker A: If you can take pictures of the screen, then it can identify where buttons and things are without having to know the name of the objects and the, in the, in the DOM and stuff like that. So you could say, you know, give the instructions, find, you know, click on this, click on this, click on this, do this stuff. And it'd be really easy to automate tests that way instead of having to know the, like, the names and the divs and things in the. On a page, especially for web testing, because if a developer changes those, then you got updates tests where if you say, click on the button that says do this, then it. Then it can. And something I really appreciate about Claude, although it won't generate images, it's really good and at analyzing images and describing exactly what's on the screen or exactly what things are doing in the image that you give it. So I think it's kind of cool. Looking forward to playing with that.
[00:03:52] Speaker B: Yeah, I quite often take. I make Claude describe and create the image description and then I go put that into Gemini or ChatGPT to actually create my image because I find that its descriptions are better, as well as a bunch of things in Cloud that I really like particularly. I was working on a powerful presentation and it actually created the whole PowerPoint presentation in React, which was kind of cool, and you could download it and play with it and stuff like that, which, yeah, if you try to do the same thing in ChatGPT, it just looks at you like, I don't understand what a PowerPoint is. Here's an outline of one, but it can't create you one.
[00:04:25] Speaker D: To be fair, Copilot says the same thing.
[00:04:27] Speaker B: Copilot also doesn't know what a PowerPoint is, even though it lives inside of PowerPoint.
[00:04:31] Speaker D: So, yeah, thank you for explaining the.
[00:04:34] Speaker B: Joke instead of help our listeners out, you know, if they haven't experienced Copilot and the joys or lack of joy that copilot for Office365 brings to you, they are not missing out.
[00:04:45] Speaker A: I think I'm going to have to step up and pay more to Anthropic now because it's not available on the Pro plan, which is $20 a month only through the API, and that's a separate plan entirely. So I may have to.
[00:04:55] Speaker B: Well, I think I might terminate my Gemini subscription and just add more money to Claude because I'm just so impressed with Claude for the most part that I don't think I'm getting as much value out of my Gemini subscription. And honestly, I'd rather have AI that works.
[00:05:13] Speaker A: Yeah, I will say that the new version, I found it to be a little less reliable with long contexts. So if you have a long conversation about especially debugging something, it seems to lose track of what it's doing or it kind of regresses things. So if it's writing a function, it'll drop some features out that you added earlier in the chat. So I've kind of gotten the habit of terminating the chats pretty early on and then starting again in a new chat. And so I'll ask you to summarize. Like summarize the conversation today, summarize the decisions we made, things like that, to hand off to the next developer, and then tell the next chat, this is what the last guy left for you. Let's start from here. And it seems to work pretty well like that. But I also think my system prompt doesn't work as well as it used to.
So maybe they've tweaked some things in tuning. That's making it sort of less responsive to the way I was giving it instructions before.
[00:06:14] Speaker B: That's too bad. Maybe that'll get tweaked as you continue to get better at your prompts and stuff.
[00:06:19] Speaker A: I mean, it's noticeable, but I'm not complaining. It's still absolutely amazing. I'm like doing the work of three people right now with the help of Claude.
[00:06:31] Speaker D: Don't tell your boss.
[00:06:33] Speaker A: Hey, Ryan.
[00:06:37] Speaker D: Don't worry, I wasn't paying attention.
[00:06:39] Speaker C: No, not at all.
[00:06:40] Speaker D: Squirrel.
[00:06:41] Speaker A: Squirrel.
[00:06:41] Speaker B: Yeah.
All right, let's move on to aws away from Claude. Microsoft, then Google, now aws. I'm positively glowing with all this nuclear energy, personally. They are announcing three deals, including an investment in Startup X Energy and two development agreements with Energy Northwest and Domini Energy to add around 300 megawatts of capacity in both the Pacific Northwest and Virginia. The agreements include constructions of several new small modular reactors or summers, which are an advanced kind of nuclear reactor with a small physical footprint to be built closer to the grid. So on top of their agreement to co locate a data center facility next to Tallinn Energy's nuclear facility in Pennsylvania we talked about a few months ago, but everyone on the cloud side is now officially on the nuclear bandwagon.
[00:07:23] Speaker C: It's really the only option right, for energy.
[00:07:26] Speaker B: Like I really.
[00:07:27] Speaker C: Yeah, I mean it's so energy intensive to run AI workloads and you can't really depend on you know, like a cloudy day ruining solar or non windy day. Like it's can augment with that. But it is kind of interesting. I'm really curious to see what they've done in terms of like nuclear waste and hopefully these smaller footprint reactors make that at least easier to manipulate versus like, you know, the giant amounts of nuclear waste that you have to track or train through town.
[00:07:59] Speaker A: Yeah, I think they use these small pellets and so they're easy to handle, they're easy to store. But in like, in the grand scheme of things, I feel like the nuclear waste scare of the 80s and 90s was, was far.
It's just a huge overreaction. Like well what are we going to do with this stuff? Like well what do you do? All the ash that you, that you come out of, that you get out of a coal plant or something you just bury in the ground. At least there's far, far less. And it really isn't terrible to handle.
[00:08:24] Speaker B: I mean in fairness, if you spill ash, it makes a mess, but you can clean it up pretty easily. If you spill nuclear waste, you know, can't go in that area again for a few hundred years. So I mean it's, it's not quite the same. Like I get, I get what I'm trying to say here, but I'm not going to fully buy in.
[00:08:40] Speaker A: I won't go into the chemistry ash is horrible. It's full of all kinds of shit.
[00:08:43] Speaker B: Well, I mean it's terrible too but I mean like I can get a bulldozer and I can remove it within a month or two of cleanup versus nuclear, which I cannot.
[00:08:52] Speaker A: Well, these pellets at least are easy to handle. So you know.
[00:08:55] Speaker B: Yeah, I mean these are easier to deal with.
[00:08:57] Speaker A: Yeah.
[00:08:58] Speaker D: Jonathan, just go grab all the pellets, it'll be fine.
[00:09:02] Speaker B: Well, our long national nightmare here in the cloud pod is officially over as Dr. Matt Wood has announced that he would be leaving AWS after 15 years. Matt had been long involved in AI initiatives and was behind many of the warning childs like DeepRacer and the recently terminated Music Solution. He was appointed VP of AI in September of 2022, and over the last two years there's been several missteps in AI, with Amazon missing out on investments in Cohere and Anthropic and is settling for a joint investment with Google in Anthropic. Amazon CEO Matt Garman is aggressively moving to Rights a ship Aqua, hiring AI startups such as Adept and investing in training systems like Olympus. And so we don't really know if Matt Wood resigned or was asked to politely leave the premises, but he apparently has a new job that he'll be starting next week. I'll update you when he posts that update, but I'm just glad not have to sit through one of his boring keynotes again, you guys.
[00:09:57] Speaker C: That's what all I could think of too.
[00:09:59] Speaker D: Yep, that's where by my way, I was like, ooh, I wonder who they're gonna replace him with and how is that keynote gonna be? Is it gonna be a little bit more lively?
[00:10:07] Speaker B: I hope so.
[00:10:08] Speaker C: I mean if they don't replace them with some sort of AI avatar, that's just an opportunity lost as far as I'm concerned.
[00:10:14] Speaker A: Yeah, plot twist. They'll replace him with an avatar of Matt Wood.
[00:10:18] Speaker B: Well, yeah, he sold his like this to them before resigning.
[00:10:21] Speaker A: Like should have read the contract but.
[00:10:26] Speaker B: Well, we wish him the best of luck. Hopefully he lands on his feet at something interesting that doesn't have keynotes that I've sit through regularly.
So Amazon is announcing Support for Amazon. CloudWatch evidently will be ending soon and I remember evidently and thinking it was weird when it was announced and so I went back to our show notes from December 2021. It was announced at Re Invent in 2021 and our comment at the time in the show notes was AWS releases cloudwatch evidently a capability that helps developers introduce experiments and feature management in their application code. And the team that's us remains confused as to why this is a CloudWatch feature which now still two years later of being terminated for end of life. I still don't remember why it's in Cloudwatch and not just in other places as well.
And so because no one uses it, they'll just do a service on 10, 17, 2025. So you have a year to migrate to alternative solution and their recommendation for this one is to move to app config which is a feature of AWS Systems Manager which existed before this feature existed. So again, it was one of the questions we asked at the time was why not just exact config for this? Why to have a competing service.
[00:11:31] Speaker C: I do love that there's no way you can find evidently, you know, because it's part of CloudWatch. But you also won't be able to find an app config because it's buried in nine layers of.
[00:11:46] Speaker A: It was only a CloudWatch feature because they like this. They like the same toppings on their pizza.
[00:11:53] Speaker B: That was the group that likes pineapple.
[00:11:54] Speaker A: That's what it was. That's right.
[00:11:57] Speaker B: Let's see. Okay, well, I don't think anybody was using it, which is why no one's really talking about evidently, since I think we talked about it in 2021. I don't think there's been any features announced for this. I don't think there's been any press coverage on evidently. So I don't think we're missing out on this one. Going away, guys.
[00:12:14] Speaker C: No.
[00:12:15] Speaker D: Nope.
[00:12:15] Speaker A: That is a terrible name.
[00:12:18] Speaker B: Yes.
For those of you who are in this AI space and are interested in serverless agentic workflows with Amazon Bedrock, Amazon is launching a new short course developed in collaboration with Dr. Andrew Ng and DeepLearning AI. This is a hands on course taught by Mike Chambers, teaching you how to build serverless agents that can handle complex tasks without the hassle of managing infrastructure. You will learn everything you need to know about integrating tools, automating workflows and deploying responsible agents with built in guardrails, AWS and Bedrock. And while I'm very excited about the concept of serverless agentic or even agentic AI in general, I'm not sure that I would do it on Bedrock.
[00:12:56] Speaker C: I mean this is a great way to find that out, right? Like I have done this with vertex AI and I've played around there. I kind of appreciate this type of marketing, right, because it's really the only way I'm going to get to play with it is if you give me like sort of a hands on environment which I can do because I, you know, I try to use my personal AWS account for things and it always just depends on how invested I get.
[00:13:19] Speaker B: So this is.
[00:13:20] Speaker C: We'll see. I'll probably try it out and see.
[00:13:22] Speaker B: If I like that route.
All right. We get two new features in the Lambda console which hasn't had much love recently. The first up is you get Amazon CloudWatch metric insights. Now in the CloudWatch console or Lambda Console? Sorry. The Lambda console also now supports real time log analytics via Amazon CloudWatch logs live tailing capability. Both of these are great quality of life solutions for those using Lambda heavily and allow you to write in your console in CloudWatch. Sorry, in Lambda, see your CloudWatch metrics that matter to you and help you debug and troubleshoot everything you need. Now they've been adding these type of features to things like rds and to Elasticsearch for a long time where they embed native CloudWatch metrics inside of the dashboard. So I'm a little surprised it came so long after for the Lambda console, but I appreciate that it's now there.
[00:14:10] Speaker D: I feel like the live tail is fairly recent and I used to a couple of weeks ago in Elastic Beanstalk. Don't ask questions, but helping out somebody with Elastic Beanstalk, we'll just move on.
And it was a really nice feature of being able to go in there and see real time, hit the API, see the logs on the server and kind of do it all in there. So I'm looking forward to actually having to be able to grab my Lambdas and immediately be able to see the output versus wait a couple seconds, refresh the dashboard, kind of go through that whole cycle having it all in one place. It's like you said, a great quality of life improvement.
[00:14:48] Speaker B: Yeah.
[00:14:49] Speaker C: When they've made this improvement in other services, that having the contextual awareness of the service really makes it a lot, you know, a very enriched interaction. And so like I remember working with UCS, right, you could do your container logs and CloudWatch logs, but if you actually view the embedded one through the UCS console, then you actually, it would group all the similar tasks together so you could actually see, you know, those logs from those individual things versus just randomly, you know, from the task ID in the console. So I imagine that in a Lambda console that'll be even more, you know, powerful.
[00:15:24] Speaker B: So that's cool.
[00:15:24] Speaker C: I like it.
[00:15:26] Speaker A: Presumably this is like a developer experience kind of feature rather than an SRE kind of kind of feature. I would think tailing logs, I mean, SRE will be retrospective anyway and looking. So that happened minutes ago, not in real time. So it's good that they're focusing a little bit more again on developer experience because I think it's where they severely lack, which is kind of bizarre because the documentation has always been like the industry leading quality, but then they're documenting terrible workflows.
[00:16:00] Speaker C: Industry leading doesn't mean leading you right off a cliff.
[00:16:03] Speaker B: Yeah, oh yeah. They definitely have always, you know, unless you're using the serverless framework or some of the other, you know, libraries to do serverless development in a hardcore way, you have to kind of do it in the console or through infrastructure as code where you're uploading files to that process, which is a little clunky, but you're trying to do this like real time developer. I can see why you want to do it in the console and have those tools available to a developer. Makes a lot of sense, especially if you're killing now the cloud ide, you know, which they announced the deprecation of a few months back, which would be one of the ways you potentially would have done, you know, these things. So I appreciate, you know, again, the quality of life. But yeah, I think you're right. It is a developer experience enhancement, which is nice for them.
[00:16:42] Speaker A: Yeah, it's super nice because even like the SAM local stuff, you could do a lot of testing locally, but as soon as you need to test things that are integrated with the rest of the ecosystem, it's just insufficient. So this is great.
[00:16:55] Speaker B: All right, well, if you have a web application that is used by Google and Mozilla and. Or a jre, actually Oracle JRE that is leveraging a website that supports entrust public TLS certificates, you're gonna have a bad time. Starting in November, as Google, Mozilla and Oracle have all announced they will no longer be trusting any certificate provided by entrust for after November 11, 2024. If you imported an entrust certificate by ACM For ELB or ThoughtFront, you will need to reissue those certs before November 12th, otherwise those users will have errors on those two browsers. And before the show, Jonathan was asking, well, why is Entrust in bad hot water with Chrome? And so I went and found this blog post where they said over the past several years, publicly disclosed incident reports highlight a pattern of concerning behaviors by entrusts that fall short of the above expectations and is eroded confidence in their competence, reliability and integrity as a publicly trusted CA owner. Ouch.
[00:17:49] Speaker A: Yeah, What I can understand about this is why are they still permitted to operate as a ca? Why is it down to Google and Mozilla and Oracle to cut them off from the client at the client instead of whoever signs entrust certificates saying, you're not good enough.
[00:18:08] Speaker B: Well, I mean, if they're, if they're a root certificate, you know. Yeah, if they're a root ca, it doesn't matter, right?
[00:18:14] Speaker A: Yeah, but you still have, you still have to apply to be a root CA provider, there's still somebody who gives you, gives you that permission.
[00:18:21] Speaker D: You have to get people to put your root CA into Chrome and all the other browsers. So if they've put them in, then they are trusted at that point. So they have to build that trust, get them in, get the distribution out. That's how people knew ACM was going to start to become a thing before it was rumored because Amazon was built their own root certificate that they were slowly propagating through all the web browsers and everything.
[00:18:47] Speaker A: I didn't think they were actually root. I thought, I thought they were still signed by somebody else. But I'll. I'll. I'll concede they were for a while.
[00:18:54] Speaker B: Yeah.
[00:18:55] Speaker A: Okay.
[00:18:56] Speaker D: The one piece I forgot here was that you actually could upload certs to acm. I've. The last time I really uploaded a cert was back when you had to do into IAM and then attach it to your elp.
They remember those days.
[00:19:10] Speaker B: Yeah, I mean it was, I mean, basically once ACM came, no one upload certs anymore.
[00:19:15] Speaker D: I know it was.
[00:19:17] Speaker B: Yeah.
[00:19:18] Speaker D: And now I'm on Azure and I hate every time I have to deal with certificates, I'm like, can't this just be managed?
[00:19:24] Speaker B: I mean, even Google has a certificate solution, so that's pretty sad. On Azure side, they have it for.
[00:19:30] Speaker D: Some things, but not for everything.
[00:19:33] Speaker B: That's a problem.
[00:19:35] Speaker D: Yeah.
And start pinning also becomes a thing, which, you know, can affect stuff also.
[00:19:41] Speaker A: Ah, yeah. The thing that security people love, but it's just a real pie in the.
[00:19:45] Speaker C: Ass no one else does.
[00:19:46] Speaker A: Yeah, yeah, yeah.
[00:19:49] Speaker B: They're all about managing risk, Ryan. It's not about managing your ease of use.
[00:19:53] Speaker A: That's managing checkboxes, not managing risk.
[00:19:57] Speaker B: Maybe.
All right, Amazon announces a seamless link experience for the AWS console mobile app. And why I wanted this, I'm also super afraid of this, uh, because the mobile apps, you know, from Google, which I was trying to use just recently for like some admin stuff on Google Workspaces.
Yeah, they're all pretty limited. They don't really cover a lot of the things. And so deep linking to a service that doesn't maybe exist in the console mobile app seems like a bad experience. And they do try to fix that, which is nice. But basically this allows you to send a link, you know, so if Ryan's on his computer saying, hey, you know, ECS is borked, I need to restart containers.
Can you, you know, can you look at this and you want to send me a link from the Console he sent to me. I'm on my mobile device and get that link. It'll open it up in the ECS portion of the mobile app so I can see the same data he was looking at.
You used to go to the native browser. You had to re authenticate and against the web version versus the mobile app and that's kind of a pain. So this is a nice quality of life improvement if you're a heavy user of the mobile app, which as much as I would like to be, I am not because they're terrible customers benefit from using the mobile app to the supports biometric authentication as well as mobile optimized customer experience. And in the few cases where they don't have a service that's supported, they will apparently now open that experience in a native browser inside of the Amazon console mobile app which if that works, okay, I'll accept it. But I'm worried it's not going to work well. Yeah, but we'll see.
[00:21:28] Speaker C: It doesn't work well on any other app that I've used that like that. But all right. It's better than I guess just hunting me to the OS browser.
[00:21:36] Speaker B: Yeah, it's a good reason. It's a big reason why I don't get bankrupted on Instagram because like the ads pop up and I'm like oh I'm interested in buying this and I hit shop and it opens up at the browser inside Instagram. But then I can't use Apple Pay or any of the ways I normally pay. I think it actual credit card and that stops you from buying 99% of the time.
[00:21:52] Speaker D: Sounds like Target just Instagram roll by because you're.
[00:21:57] Speaker B: Unless you're going to send me to a native web experience that supports Apple Pay, I'm not. I don't give a crap about you.
[00:22:01] Speaker A: Oh cool.
[00:22:02] Speaker B: Is it.
[00:22:02] Speaker A: Is the marketplace there in the mobile app? Can I start sending people?
That's real shame. I was going to send a bunch of links out to people that subscribe to my products in the marketplace and see how much money I can make from missed clicks.
[00:22:15] Speaker B: I mean maybe you could try. I mean honestly I try to use the mobile apps for the cloud providers as little as possible, but I would highly doubt Marketplace has mobile enabled themselves.
You could try. You could try.
[00:22:29] Speaker C: Like the emergency that would have to happen for me to be using the mobile app. Like it'd be pretty bad. Like I'd have to be the only person left on earth.
[00:22:37] Speaker D: Well at that point why that can.
[00:22:39] Speaker C: Reboot the CC2 server and I'm on an island in Tahiti, but I have my phone.
[00:22:45] Speaker B: I mean, I have done this before. I have rebooted an EC2 host from my phone in the mobile app. It has happened. I can count on one finger how many times has happened, but I can tell you I've done it once.
[00:22:57] Speaker C: Yeah. Yeah.
[00:23:01] Speaker A: 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? Our chair will click the link in the show Notes to check them out on the AWS marketplace.
[00:23:40] Speaker B: All right, so stop me if you've had this scenario before.
Your marketing team perhaps needs to put objects into an S3 bucket. And so you give them an Amazon account and you go give them access to the console and they go to the S3 bucket and then they can't find their bucket because they don't understand the next button and how to get to the right bucket. And there's no way for you to filter that down easily for them without making it easy. And so you maybe then you created a custom web application that you're trying to do the same thing and again you're showing them all of the buckets because the darn list bucket API did not support any type of filtering. Which is also fun when you're in the AWS CLI and you type list buckets and it just scrolls for days. You're like, great, that's not helpful for me in any way possible shape. And so Amazon's fixed that finally. It's only taken, you know, 10 years since Amazon S3 became a thing or longer. And now it's fixed it because you can now restrict access to the List Bucket API with a filter that supports not only pagination, so you can list buckets across multiple pages, which also fixes my other problem and allows you to now query by the region as well as a name prefix for your bucket. So you can find all the buckets labeled marketing underscore, whatever.
[00:24:48] Speaker D: It's amazing how many times they've had to. Somebody's been like, okay, they just need access to this bucket. And like someone put gave them just access to the Bucket. And then they were like, if they can't, doesn't work. And I'd be like, did you do list? And then literally your scenario would come up. And it's amazing. It's taken 15 years for this to get fixed. Like, I understand S3 is in its own world, and I am, because it pre exists. I am. But like, this feels like it should have been something a little bit easier, but I hate to see the tech debt they uncover to try to actually build this out for people.
[00:25:24] Speaker C: Yeah, I think it, like, normally I.
[00:25:26] Speaker B: Would agree with you.
[00:25:27] Speaker C: Like, it's all the tech debt that they had to overcome must have been. But I just don't see it. But, you know, like, I think it's just they forgot and someone complained loud.
[00:25:39] Speaker B: Enough, someone with a big enough paycheck, a big enough check, said, if you don't fix this, I'm not renewing my edp. And they're like, okay, sir, we'll do it.
[00:25:46] Speaker C: Okay. Yeah.
[00:25:48] Speaker A: It's strange because the names are global names, and so the buckets existing in a region, they must know. They must have known on the back end which region those buckets were in. So it's not like they had to go back and add data to some kind of objects or whatever because they know where the buckets are.
[00:26:05] Speaker C: Yeah, you have to create them in a region. It's just a global namespace.
[00:26:08] Speaker A: Yeah, that's.
[00:26:10] Speaker C: I mean, that's horrible tech debt from everything being in U.S. tirefire one for at the beginning. But yeah, I don't know.
[00:26:19] Speaker B: But I mean, it would be nice because if they add this into the console, ideally you'd be able to have the ability to see all buckets in the S3 console or filter by the region that you're in, which would be also a nice quality of life enhancement.
[00:26:30] Speaker C: It's going to the individual region page.
[00:26:33] Speaker B: Yeah. So I mean, I'm hoping maybe this will be a nice quality of life improvement, but we'll see. I'm just glad it exists now because. Yeah, that was a painful, painful experience.
[00:26:42] Speaker C: Yeah.
[00:26:43] Speaker B: So I will tell you now that Amazon supports the new Claude 3.5 sonnet, but they were five hours after the announcement from Claude. But because we talked about AWS first. They're here first. And Google was within like 20 minutes of Claude announcing 3.5. They were supporting it. So Amazon, you got to pick up the paceman.
[00:27:01] Speaker C: They probably have AI writing the press releases versus Amazon.
[00:27:07] Speaker B: All right. GCP this week. So I have a thing that I did that I want to talk to you guys about and it has to do with this feature. So I foolishly, before I went on vacation, agreed to do a talk on Synops for AI workloads, which I have experience with, but I'm not an expert in by any stretch. And I agreed to that on the Friday before I left for vacation and I was coming back. And I have basically I had to have a deck done by Wednesday for a presentation I'm giving on Friday. So I did not do myself any favors because basically I had no time. And so I read a bunch of articles why I was bored on my vacation about finaster AI to get some more data to kind of think about how I wanted to do this presentation. And I just happened to listen to a podcast where someone was talking about how amazing this feature from Google called Notebook LM was. And it solved the exact problem I was having, which is that I was trying to use Claude and Gemini or ChatGPT to upload a FinOps document and then ask questions about the FinOps document, which worked relatively okay except for I kind of ran to the Jonathan's problem where like after like four or five questions it forgets what the grounding you gave it was.
And so this Notebook LM feature basically allows you to upload links or documents or all kinds of different content. It basically creates a small grounding environment for you around Gemini to 1.5 and then you can basically use it to ask questions in a chat format. You can create study guides, you can do all kinds of things. So I was able to basically ask a bunch of questions about the documents that I uploaded to and was able to kind of create my presentation which actually went really well. It was really well received. So I appreciate the help that opagellum gave to me. But I want to talk about a feature that I think you guys will find interesting. So this announcement is for basically they're providing you ability to customize audio overviews. So one of the features that the Novik LM can do when you upload all these documents to you is it can actually create an audio overview of the content that you've gift as a study guide. Like almost like a podcast. It can ask you quizzes, they can do all these things. And so now you can, with this new capability you can tell it, hey, out of the 25 documents I uploaded to the thing, I'd like you to focus the audio guide on a specific document or like you to make sure you cover these four or five specific topics in basically the audio overview. And so it produces basically a two person podcast that you can listen to and I will now play for both all of you, so you can check this out. It is pretty darn impressive.
As soon as it loads here in our.
[00:29:37] Speaker E: All right, everyone, welcome back. Today we're going to do a deep dive into something that's probably keeping a lot of you up at night.
The cost of AI. Especially when we're talking about cloud based AI services.
[00:29:51] Speaker B: Yeah, it's a bit of a wild west out there right now.
[00:29:53] Speaker E: Totally. And I think a lot of you listening are probably trying to figure out a few things. Like first off, how are these things even priced? It's not your typical software subscription.
[00:30:03] Speaker C: Right.
[00:30:03] Speaker E: Then there's the whole forecast and thing. How do you predict your AI spending accurately, especially with everything changing so fast.
[00:30:11] Speaker B: Absolutely.
[00:30:11] Speaker E: And finally, are there any practical tips out there to avoid those nasty surprise bills and actually optimize your costs as you scale?
[00:30:20] Speaker B: Right. Because no one wants to get caught off guard.
[00:30:22] Speaker E: Exactly. So in this deep dive, our mission is to give you a clear framework for thinking about AI costs. Army with the knowledge to make smart decisions and hopefully avoid some of those common pitfalls.
[00:30:36] Speaker B: Sounds good.
[00:30:37] Speaker E: So to kick things off, let's set the scene. There are basically three main ways that organizations are deploying AI right now and each has its own cost implications. You've got your third party vendor services like Think OpenAI. They're the big names offering prepackaged AI solutions. Super easy to get started with, generally high quality, but you often have less control and they can be pricier.
[00:30:59] Speaker B: Yeah, the convenience comes at a cost, literally.
[00:31:02] Speaker E: Then there's the option of using third party hosted open source models. So platforms like antiscale or hugging GPT, these give you more flexibility and often better cost efficiency, but they require a bit more technical know how.
[00:31:17] Speaker B: Definitely you need to be a bit more hands on with those.
[00:31:20] Speaker E: And for those who really want complete control, there's the DIY approach. Building your own AI solutions on cloud providers like aws, Azure and the like.
[00:31:28] Speaker B: Right. The full control freaks. This one's for you.
[00:31:30] Speaker E: Exactly. Ultimate customization. But it demands some serious expertise.
[00:31:34] Speaker B: No kid.
[00:31:35] Speaker E: So right off the bat you can see that.
[00:31:38] Speaker B: All right, so that's, that's pretty darn impressive in my opinion. Like it's. Yeah, you can definitely tell at different, like levels of, you know, how technical you want it to be. I chose kind of a medium technical ability for it, you know, and that's what I gave in the guidance for this near future, but gives you kind of an idea and it's sort of funny to me because it has some of the inflections that you would have in a podcast when you're thinking. And so that, you know, we're not on the job yet, but, you know, maybe someday.
[00:32:04] Speaker C: Yeah, it's completely replaced my role in this podcast of just going. Yeah, absolutely. That's true.
[00:32:11] Speaker A: Yeah, that's. I think I got, like, three awful things to say. One, I felt like I was listening to NPR a little bit.
[00:32:17] Speaker B: It has definitely an NPR vibe to it.
[00:32:19] Speaker A: Episode of Planet Money or something like that. The man's voice is amazing.
The woman's voice I find too fast, and it kind of a little bit choppy. You can tell that it was alternating between two. Two tracks. And because she was interjecting, it was like a little pause in the conversation. I think they record two separate tracks and then merge them so they can, like, overlap slightly a little bit. But the thing that really bugs me about that is it's. It's like the useless person on the podcast or the useless person who's just there to agree and just repeat things that have been said already or just ask questions to kind of lead the conversation and. And the like. In that. In that example, the woman's part added absolutely no value whatsoever. And I. You're welcome.
[00:33:05] Speaker C: My servicing last.
[00:33:06] Speaker B: Yeah.
So I will say that if you listen to the whole 11 minutes, which I did.
[00:33:13] Speaker A: Yeah.
[00:33:14] Speaker B: She does get more involved later on in the podcast and a little less of a responsive. But, yeah, I didn't want to more than listeners with 11 minutes of podcast.
[00:33:21] Speaker A: No, no, it's just. It's just a pattern. I notice a lot in especially tech podcasts or tech YouTube channels and things where you've got, you know, the bunch of white guys and then the token female who is just there by eye. Candy, honestly. Yeah, it's kind of sad.
[00:33:41] Speaker B: It's word choices are also very AI word choices. It's one of the ways that I always spot AI content is a view of words that people don't naturally use. And you see some of that in the podcast and you go through it more. And again, it's not meant to be a podcast. It's meant to be an audio guide to help you learn.
But still some stuff.
[00:33:58] Speaker A: Yeah. Just needs to clone our own voices now.
[00:34:01] Speaker B: Exactly. I was like. I was trying to see, like, I do four. I was trying to ask if I could do, like, four different voices. It'd also be nice to be able to change those voices. It doesn't have a lot of customization features yet of how you Put that together. But you know, again, power of AI, it's coming for all of us, even us podcast hosts.
[00:34:17] Speaker A: I mean, I can be critical, but technically it's really impressive.
[00:34:21] Speaker B: It is really impressive.
[00:34:22] Speaker C: Yeah.
[00:34:24] Speaker B: All right, next up, Compare mode is now coming to Google AI Studio. This is your companion for choosing the right Gemini model, either being, you know, Gemini Ultra or even the Gemma models. This to pair mode basically allows you to run the same prompt across multiple models at Google, assessing the cost, latency, token limits and response quality, allowing you to evaluate the response across various models available in AI Studio side by side. This will help you pick the right model for your use case and allow you to save, hopefully money as well.
[00:34:54] Speaker A: Yeah, save money after you've spent three times the money comparing the outputs of three separate models. Like it was conveniently placed after the FinOp section.
[00:35:04] Speaker B: Yes, conveniently placed.
[00:35:06] Speaker C: I also wonder how much this is going to like, you know, the more expensive models are going to perform better in most cases. And so like it's going to be, it's going to lean you in that direction, or at least it seems like that's going to be the case.
[00:35:19] Speaker B: But actually part of my talk that I gave was talking about the exact issue that, you know, you can choose Gemini Ultra, but it's going to cost you a lot more money and you should pick the right model for your use case, you know, so if you're doing a simple text summarization use case, you know, a text bison Gemma model makes way more sense. It's more cost affordable. But it's interesting you mentioned the cost of testing these models.
I was looking at the research. Gardner says that basically 70% of the cost of a model will actually come from the inference of the model, not from the training side of the model. So if you're just choosing and evaluating models based on a certain set of use cases, that's probably actually a pretty low cost comparatively to running the model long term in your solution, which is pretty interesting. But we can do a whole after a sad journey talk on what I learned about AI finops, because it's quite interesting.
[00:36:07] Speaker A: I think the problem is when presented with models of different qualities, different numbers of parameters, different whatever, people will always want the best one.
[00:36:17] Speaker B: Until you see the price tag, then maybe you don't.
[00:36:20] Speaker D: I was going to say from a business point of view, you're going to. You might want certain models in certain places and others. And others. So while this area of your application, you might want to be very exact in a very good model over here, you might Be able to say, get away with a lower end model. You'll always need to run the, you know, X32, you know, whatever on AWS. You might be able to run the T3 here and you get enough performance to solve your problem.
[00:36:48] Speaker A: I mean, speed of response is going to be important for chat applications for sure.
But even just thinking about how you quantify the difference between different models, especially if it was just text generation, is really hard. Like how, how do you objectively say, well this one's better than this one for these reasons or not? I think that's gonna be a really hard analytical task.
[00:37:14] Speaker D: But that's up to your business. You know, it depends what the, what the objective of that, of that API call is. Maybe, you know, you're talking about speed. Well, maybe it's take this data, run some, you know, let's just say, hey, simply just get our, get a show notes right? Because I definitely never have done that for our show notes. You know, summarize this article for me. Well, I don't need that instantaneously. I might be able to get away with hey, give me five show notes and email me the results in three minutes from now. And that's fine versus I need to write this second. Well, that's a different use case.
[00:37:49] Speaker A: Oh yeah, for sure. The quality of the text that comes out though is harder to quantify.
[00:37:53] Speaker D: Yes.
[00:37:54] Speaker A: What does acceptance criteria look like when you write a story that says I'd like to go away and compare the output of like Claude 3.5 sonnet versus opus versus Gemini 1.5. Like that's a difficult task.
[00:38:06] Speaker D: Yeah, yeah, for sure.
[00:38:11] Speaker C: No, you just ask the, you ask AI to analyze the responses and give you which one's the best.
[00:38:16] Speaker B: That's actually probably true. You ask the AI to double check itself. Right? That's one of the cool use cases I learned about agentic AI agents and how you, you know, if you use multiple agents you can give them different tasks and one of the tasks can be the peer review task and that, you know, ask you basically with the expense of like be a scientific peer review and like really scrutinize the outputs of the other models. So you can do all kinds of things like that. When you get into some of the Gentec use cases.
[00:38:41] Speaker A: That's cool. So first we'll do like this testing of which is the best model for the peer reviewing. And then once we pick the best peer review model, we'll hey, hey, turtles.
[00:38:50] Speaker B: All the way down, man. Turtles all the way down. I can't help you. Here.
[00:38:53] Speaker C: Yep.
I'm hoping I can use this tool to kind of back into like figuring out how tokens are actually used and the limitations there. So like, because I so far I've failed to understand it and so now maybe if I compare across several models and I can see the utilization, maybe maybe I'll get some understanding.
[00:39:10] Speaker A: I have a YouTube link for you if you want to get into some Python.
[00:39:17] Speaker B: It's really just put in the show notes for all listeners now.
[00:39:19] Speaker D: Yeah, I kind of want this now.
[00:39:22] Speaker B: Yeah, I want to see it too.
[00:39:24] Speaker D: We know what our after show is next week.
[00:39:28] Speaker B: I'll watch the video on YouTube and talk about it. All right. Google supports Cloud 3.5 and the Vertex Model Garden as well as the computer use case capability if you're excited about that use case. So you're welcome. Also be able to the 2024 accelerated state of DevOps reporting a report has now been published have several findings as well as a full report that I have not fully read. So I this is an article based on the blog post as well as the executive summary that I did read and I have a heavy reading material later this afternoon or later this week to read through the rest report. But one of the highlights of the report is the widespread AI adoption is reshaping software development practices, with over 75% respondents saying they are relying on AI for at least one daily professional responsibility. They also said that more than one third of the respondents said AI experienced moderate to extreme productivity increases from AI. However, at the same time it said AI adoption may negatively impact software delivery performance and a reduction in delivery stability.
And despite all these productivity gains, respondents did report they had little to no trust in AI generated code, which feels like they were playing both sides of the equation a little bit. Like well, I like it a lot, but I don't want to be that like it because I don't want people to replace me. That's how that feels a little bit.
Platform engineering is another area of increased adoption. Apparently, per the report, they have four key findings around platform engineering, including increased developer productivity. There's a heavy prevalence for platform engineering and larger firms. There's a potential performance dip initially as you start adopting platform engineering. And there's a need for user centeredness and developer independence in your platform. And then developer experience again appeared as a pretty key cornerstone to the success and the overall efficiency of your DevOps organization. I'm a little bit curious to read more about the platform engineering because that's what I'm doing a lot of these days in my day job. And so I'm very curious to learn a bit more about that, but maybe I'll share any additional anecdotes I have after I have a chance to read the full report. This was published like literally two hours before record time. So yeah, sorry it wasn't more thorough today.
[00:41:36] Speaker C: Yeah, same thing. Like I'm really interested in the, specifically the performance dip right at the beginning and then what, what they're looking for when it comes down to the developer independence. Because that's always something I, I battle with, you know, in the platform engineering work that I'm doing, which is, you know, trying not to become just the next service catalog of ready to launch Compute, you know, cookbooks or what, what have you. And so it's being kind of interesting. Look forward to that.
[00:42:09] Speaker B: All right, moving to Azure, the new Secure Sandboxes at scale, or Secure Sandboxes at Scale with Azure Container Apps Dynamic Sessions. Container Apps is a serverless platform that enables you to run containerized workloads without managing the underlying infrastructure. Dynamic Sessions add the ability to execute untrusted code in a secure sandbox environment at scale. Dynamic Sessions provide secure ephemeral sandboxes called sessions for running that potentially malicious code. And the Dynamics are deal for running untrusted code in a hostile multitenant scenario, including running code generated by LLM. So don't just use it in production. Running code or commands submitted by cloud app users, or running cloud based development environments, terminals and more, all from the secure Sandboxes.
[00:42:53] Speaker A: That's really good. Imagine you have a service where you want people to be able to define something as code, like a dashboard or some kind of agent for AI or something like that. And you want to test it in a sandbox where it's not going to have any production impact if it fails or goes into some infinite loop or something. It's great. It's really nice to have an isolated place to go and test things.
[00:43:13] Speaker D: I feel like this is also going to be one of those features that's like, do you have this disabled in production?
[00:43:20] Speaker B: The sandbox didn't have access to production. So again, it's a place to experiment and to do things that you want to make sure are in a safe environment.
[00:43:29] Speaker A: So yeah, that's cool.
[00:43:33] Speaker B: So I had to ask Matt, why does anyone trust Microsoft for anything related to security? Exactly.
[00:43:40] Speaker D: Because security is their key focus this year. Because they're so good at it already that they really need to step it up. They don't need to step it up at all because they're so good at it.
[00:43:50] Speaker B: Oh, good. Well, I got some nonsense for you for this week then.
Microsoft has notified customers that it is missing more than two weeks of security logs for some of its cloud products, leaving network defenders without critical for detecting possible intrusions. Per the notes sent to customers, a bug and internal monitoring agent resulted in a malfunction in some of the agents when uploading log data to their internal login platform. The notification assures you that it was not caused by a security incident and only affected the collection of log events. And I said to myself, well, what possible services could you have lost logs for? That would be bad for my organization. Oh no. Entra. Entra id, which is Azure, Ad Sentinel, Defender for the cloud and Purview, which is their DLP solution.
And ironically, this comes a year after federal investigators complained that Microsoft was withholding security logs From a certain U.S. federal government department that had their emails stolen by Chinese hackers in the company's heart and government only cloud. So not a great look, Microsoft. Not a great look.
[00:44:51] Speaker C: Wow.
[00:44:52] Speaker D: Yeah, they, you know, they said a couple, what was it, like six months ago? Or what feels like six months ago? Security is going to be their focus and they really have to stop messing around and actually make it be a focus because there's only so many hits before people really start yelling at Microsoft being like, guys, you can't lose our security logs. That feels like 101. These systems need to be tested through and through before we promote it. Especially for like your dlp, your ad, your SIEM software. Like, you can't be missing these things.
[00:45:30] Speaker B: Absolutely not.
So I also the fact that they only notified customers who are actually impacted not realizing it was going to get leaked. It didn't allow them to control the narrative on this at all. Which makes it sound even more worse than it probably, maybe is. But still, you can't tell me Entra ID lost log for two weeks. Just. You can't. Sorry.
[00:45:51] Speaker D: And the fact that it took two weeks to even real. Like, I don't even understand that part. Like, how did it take two weeks?
[00:45:58] Speaker B: Well, it was probably by the time someone got through the myriad of Microsoft support options to report it to the proper team that they didn't have logs from two weeks ago, that's how they took two weeks to find it.
[00:46:08] Speaker D: That's true. Yeah. Because you know, their first line of the first tier of support is so amazing. All the outsource vendors that they use.
[00:46:15] Speaker B: Yeah, the first line I would have told you to. Here's the API to use to get the Logs from the, you know, from Entre. And then you say that, yeah, I tried that, it doesn't work. And then they get you, the next ad, he tell you, well, you can use the console. You know, you don't get through like at least four people escalation before you got to somebody who actually could look at the problem.
[00:46:31] Speaker D: Then it goes into a, goes to the project group, which probably is a set, which is definitely a set bls, if you're paying for a crisis manager, which also means that for the product group, that's a level three case, which means they have, I don't know, a three, a three day SLA on those, which, which has taken me two weeks to get through a three day sla, obviously. So they're really good at that. And then from there you might get somebody from product that actually looks at something. So, yeah, no, two weeks seems reasonable.
[00:47:01] Speaker B: All right, well, it's fine.
Robust cybersecurity measures are vital for organizations to address evolving cyber threats and navigate regulatory requirements and their impact on compliance strategies. And so Microsoft this week is giving you a guide to help you with your NIS 2 compliance under the European Union Union set of security measures to mitigate risk of cyber threats and overall levels of cyber securities. But I just can't. How do you explain to the EU that you're missing security logs for two weeks and I'm out on this?
[00:47:27] Speaker C: Yeah, yeah. Leverage tools to navigate compliance, but we can't maintain compliance. That doesn't really.
[00:47:36] Speaker B: Yeah, Maybe, maybe, maybe they shouldn't try to help me with my NIS 2 or any other ISO things. Just not going to trust you. Sorry.
[00:47:46] Speaker D: I wonder what type of that's going to be on their, like ISO or their SOC or whatever audit they do next year. Like, you lost two weeks of data.
[00:47:56] Speaker B: I mean, does it, does it even come up on your sock? Because, like, are you going to, are you going to admit to the auditor you lost in two weeks? You're going to hope they don't, you know, are you going to see if they actually, you know, request the two weeks data in the two weeks they lost? You know, you know, like, how do they actually find it in the audit? Like, that's the truth.
[00:48:10] Speaker A: If I was the auditor, I would specifically ask them, show me the logs from Entra ID from these specific two weeks.
[00:48:16] Speaker B: Show me, show me 52 weeks of logs, please.
[00:48:19] Speaker A: Yeah, yeah, absolutely.
[00:48:23] Speaker D: I say this as somebody doing a current SOC and ISO audit for my day job right now, and I'm like, yeah, that sounds about right. That's something they would ask.
[00:48:32] Speaker A: I mean, it's not like they, they'd be technically out of compliance. I mean, they have the process in place and it's the process that gets verified and sometimes things go wrong.
[00:48:42] Speaker D: Two weeks.
[00:48:43] Speaker B: I mean, technically they logged it. They just lost them after they logged it, so.
[00:48:47] Speaker A: Well, it says, it says it didn't upload the logs to the logging platform, which means that if a customer had been sitting there watching the console expecting to see logs, they wouldn't have been able to see them. So it did take them an extraordinarily long time to notice that they were actually completely missing any kind of logs from these systems.
[00:49:03] Speaker B: Well, I mean, if you're a security team, you're overworked already. You're probably using some type of tooling and AI that's going to help manage.
[00:49:10] Speaker D: Copilot for security.
[00:49:12] Speaker B: Yeah, Copilot for security. So there's no logs for it to alert on. I'm not going to get notified until I have to go, you know, investigate like, you know, hey, so, so access the system illegally. And now we're trying to go do an investigation and they can't have, they can't find logs. That's.
[00:49:25] Speaker D: But you wouldn't have gotten notified because your SIEM tool, which is Sentinel, doesn't have any of that stuff in it.
[00:49:30] Speaker B: Doesn't have the logs.
[00:49:31] Speaker D: No, somebody did access it.
[00:49:37] Speaker B: Okay, we got to move on from this. Yeah, that hurts my brain.
If you've been anxiously awaiting for some ARM based virtual machines on Azure, they're pleased to announce The Azure Cobalt 100 based VMs are now generally available. These VMs leverage Microsoft's first 64 bit ARM based Azure Cobalt 100 CPU which has been fully designed in house, hopefully not by the security team. The new Cobalt 100 instances are in two varieties, a general purpose DPS V6 series and a memory optimized EPS V6 series. The DPS and DPD SV6 offers you up to 96 VCPUs and 384 gigs of memory, while the DPLS V6 only offers 192 memory and 96 V CPUs and the EPS series offers you 96 V CPUs and 672 gigs of RAM for your ARM. Pleasure. Michael Kirmeyer, VP of business strategy and Operations at Databricks, has to say we are really excited about the new Cobalt 100 VMs. We are making the primary platform for our Databricks SQL Serverless offering on Azure as they offer outstanding efficiency and allow us to deliver significant price Performance improvements to our customers. Customers using our Azure Databricks classic jobs offering will also greatly benefit from cobalt VMs by selecting them from their job cluster nodes, achieving noticeable performance improvements while keeping operating costs down. And if you're using Databricks SQL Serverless, you know, you might want to do some testing just to make sure everything is still kosher. Not that I'm. Not that I'm doubting Microsoft's first foray into a true ARM based processor that they custom designed in house, but I'm just saying I maybe wait for me to.
[00:51:14] Speaker D: I mean I remember playing with the Gravitons when they first came out and they were pretty nice and so it is something that I kind of will throw into some dev and other environments to see how well they are. And what's nice is they're actually pretty well available like I'm looking at it and it's a good chunk of reasons that are available day one.
[00:51:38] Speaker B: That's good coverage. I mean ideally Microsoft wants you to use them because they're going to make better margins on them. They're not paying any intel tax anymore.
[00:51:47] Speaker A: Actually really affordable. I just looked at the prices at 96 core 672 gig of memory with a three year savings plan. Just over $1,500 a month and that is an amazing amount of compute for $1,500.
[00:52:00] Speaker C: Not bad at all.
[00:52:01] Speaker A: Half that price on spot if you can. Well you can get it I guess but yeah, that's super impressive.
[00:52:08] Speaker B: It's only been 280 episodes that we've talked about new VM types and that one of you has actually got the pricing for me. I really appreciate you Jonathan because I was supposed to do that before the show and I forgot. So now that you just did that real thank you.
[00:52:23] Speaker C: Respect.
[00:52:23] Speaker A: Yeah, yeah that's. That's pretty cool.
[00:52:28] Speaker B: Azure is announcing two agentic capabilities that will accelerate the gains and bring AI first business processes to every organization. The first is the ability to create autonomous agents with Copilot Studio which will be in public preview next month. And Second they've introduced 10 new autonomous agents in Dynamics 365 to build capacity for sales, service, finance and supply chain teams. Earlier this year they announced Copilot Studio and private beta with basically allowing customers to reimagine critical business process with an AI agents draw on the context of your work data and Microsoft graph system record dataverse and fabric and they can support everything from your IT help desk to employee onboarding and act as personal concierges for sales and services. Some of the autonomous AI things they added to Dynamics 365 Salesforce qualification agent in a professional where timely literally equals money. This agent enables sellers to focus their time on the highest priority sales opportunities while the agent researches leads, helps prioritize opportunities and guide customer outreach with personalized emails and responses. The supplier communications agent, which will talk to your suppliers about order statuses and the customer intent and customer knowledge management agents which allow you to basically improve your first impression. As these two agents are game changers for customer care teams facing high call volumes, talent shortages and heightened customer expectations. And they will very quickly direct you to the documentation, I'm sure. For the first level support, these agents work hand in hand with the customer service representative by learning how to resolve customer issues and autonomously adding knowledge based articles to scale best practices across your care team. So agenting agents, guys, it's going to be a big thing this next year.
[00:54:01] Speaker A: Oh yeah.
[00:54:01] Speaker C: Oh for sure.
[00:54:02] Speaker A: They're not just. They're not just agents, they're the AI workers for hire.
[00:54:10] Speaker C: Going back to platform engineering, I completely intend to start using some of these agentic processes for internal uses. Right. So onboarding to companies is never easy. And so if you can have a curated guide that can sure be us through that process. Sounds fantastic.
You know, it's great.
[00:54:29] Speaker A: Like a choose your own adventure AI. It's like, yeah, you've started working at this company to the north. You can see the window you can jump through.
No, no, it's going great. Really is.
[00:54:45] Speaker B: Yeah, it's going fine. Unless you do the logs then we don't have those for you, but.
[00:54:49] Speaker A: Well, you know, they could build, they could build an agent that logs in and checks to see if the expected number of logs are there on a regular basis.
[00:54:57] Speaker B: You only received 18,000 petabytes of logs this last five seconds. We should have received at least 75,000 petabytes. Makes sense. It's Windows. The logging is ridiculous.
All right, guys, it's been another fantastic week here in the Cloud. We will see you all next week.
[00:55:15] Speaker A: Excellent. See you later.
[00:55:16] Speaker D: See you everyone.
[00:55:16] Speaker B: Bye, everybody.
[00:55:20] Speaker A: 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.
