284: Amazon Q uses machine learning to get smarter, but Bond's Q can turn a wristwatch into a laser beam. Your move, AI.

[00:00:00] Speaker A: Foreign. [00:00:08] Speaker B: Where the forecast is always cloudy. We talk weekly about all things aws, GCP and Azure. [00:00:14] Speaker A: We are your hosts, Justin, Jonathan, Ryan and Matthew. [00:00:18] Speaker C: Episode 284 recorded for the week of December 10, 2024. Amazon Q uses machine learning to get smarter. The bonds Q can turn a wristwatch into a laser beam. Your move. [00:00:29] Speaker A: AI. [00:00:31] Speaker C: Good evening. Jonathan, Ryan and Matt. It's full house. [00:00:34] Speaker A: Yeah, yeah. [00:00:35] Speaker B: Hey Justin. [00:00:35] Speaker C: Hello. [00:00:35] Speaker A: Hello. [00:00:36] Speaker D: Hello. [00:00:37] Speaker C: I mean it's glad you're here because there's a lot of things that were covered at Re Invent since we last recorded and unfortunately you guys weren't able to join us for the prediction. So by default you guys can't win. So between Jonathan and I and we'll see how we, how we end up doing this. But Re Invent happened last week here in beautiful Las Vegas where they had a lot of AI, a lot of Q, a lot of bedrock on the main stage. It's kind of interesting. I actually saw it today. Forest Brazil had an interesting blog post and he was talking about this is the first reinvent where they didn't announce a ton of new top level services. Which is accurate because even though the last few years of top level services they've announced are always super niche or don't really have a lot of market value. So sort of interesting perspective that Forest had on it. But what did you guys think from what you've been able to capture from the Twitters and the X's and all the different social media around Re Invent as well as looking through our long list of show notes? [00:01:35] Speaker A: Yeah, I was, it was interesting because I hadn't really put it directly to words like Forrest did, but then when he reading his post sort of gelled what I was already thinking, which was, you know, this was first reinvent in a long time where, you know, the new services, besides the bajillion AI based ones, announcements, you know, were all stuff that I could see as being like useful and, and you know that. And there are improvements on the existing services for the vast majority. So I was, I was pretty stoked about a lot of those. And so rather than focusing on, yeah, that one last corner of one of the last unserved market, it seems that they're focused on improving what they have, which is great. [00:02:19] Speaker B: Yeah, I think it's kind of in line with the fact that they've killed off a lot of services which probably didn't have very many customers. So there's definitely a focus on polishing the stuff they've already got rather than building Any new stuff right now? [00:02:30] Speaker C: I think what's interesting too, his comment about the show floor for the exhibition hall, that there's a lot more partners who provide developer tools versus trying to compete with CodeCatalysts and CodeDeploy and all these other things. And, you know, he felt like there was an overall feeling of the development productivity might come back in a big way, which would be interesting to see because I don't know that Q solves a lot of developer productivity concerns yet, but it is interesting, you know, the general feedback. And he's not the only one I've heard say, you know, this is one of the best reinvents that's been around for years. Which is. Yeah, I don't think anyone said that. And at least the last five I've been to, it's always like, oh yeah, it just keeps getting worse and worse or bigger and bigger and know, not that great. So you. The fact people said this was actually one of the best ones they felt in a while is pretty impressive. [00:03:14] Speaker D: Maybe it's because none of us were there. [00:03:16] Speaker A: Yeah, I mean, it, I mean, it. [00:03:17] Speaker C: Definitely, definitely ups the numbers for people, I'm sure. [00:03:20] Speaker A: Yeah. I mean, a lot of it is, you know, the reason why we haven't been to Reinvent is, you know, not only is it shifting sort of interest in expanding, you know, workloads into other clouds, but it, it did become a chore to go to Reinvent and it was a particularly useful, you know, we had started taking the, watching the, the keynotes from our Airbnb because it was like navigating those crowds for the announcements wasn't really worth it. [00:03:47] Speaker C: And so it was. [00:03:48] Speaker A: Plus I made bacon. Yeah, Drunken bacon. You know, my favorite post that I saw on, I think it was on Blue sky was that, you know, the Reinvent tagline should be this space closed for private event. Just because that felt like that was the experience. Right. No, no one was open if you weren't on the list or had, you know, that thing like there's no just impromptu meeting people anymore. It was all very organized. And so it just sort of, I don't know, diminished. And so what I hear, like, it sounds like if there's new life being invigorated, blown into it, that's fantastic. Maybe consider, you know, going next year. [00:04:24] Speaker D: It got too large essentially for what it was. And, you know, hopefully it wasn't there this year, but hopefully, you know, they kind of figured out how to make it work a little bit better because, you know, it was like, you know, between the shuttle buses and everything else, it just wasn't working anymore. [00:04:41] Speaker C: I mean, they definitely kind of went back to their roots after Pandemic. [00:04:44] Speaker A: Right. [00:04:44] Speaker C: Because they. They definitely didn't have the venue sprawl and the massive amount of busing that was happening, you know, pre Pandemic. And even, you know, they've grown a little bit bigger than a couple hotels now still, even for this year's reinvent, they're just not as much. I mean, I think at the peak, it was six different venues across six different hotels. It was crazy. So I do think Pandemic and post pandemic conference attendance has helped them keep it a little more tight. But I still think they're missing out by not regionalizing re invent. I think they would be better off served having an APAC version and a European version and a US version. But that's not how they're set up. From a deployment and release process for product management. It's all about hitting the reinvent deadline. [00:05:25] Speaker D: But you're seeing more announcements also, I feel like at the summit than you used to see in the past. So I think they're kind of spreading the love out a little bit more. Among some of the other. Other locations where they announced like Storage Day and some of these other ones, they're getting more of these announcements throughout the year that hopefully kind of make Reinvent just be, you know, almost like a thought process experiment, you know, we'll talk about later. But you know, obviously Vogel always has these good ideas and concepts that he's talk about. And that to me is a lot of the value is like sitting down listening to some of these things or talking with other people at this. At the events and, you know, so while the announcements are fun and seeing the semi truck drive on the stage, you know, is fun. You know, the. The. There's not a lot of those types of things anymore that they can pull out. [00:06:18] Speaker C: Yeah, yeah. RIP semi truck is that service is dead too. [00:06:22] Speaker D: Yeah. I mean, how many semi trucks did they actually move to aws? That's really the metric I want to know with and without armored cars, because that was an optional add on. [00:06:32] Speaker C: Yeah, I imagine they only ever built a couple of those trucks, so there. [00:06:37] Speaker D: Was more than 2. I know that because at one point they determined they could only have a region load two at a time without breaking a region. So at one point there was definitely more than two trucks. [00:06:52] Speaker C: Interesting. [00:06:53] Speaker A: I will say, as a digital attendee this year, like the turnaround on the videos for not only just the big keynotes but also the session library was amazing. Like it was same day in most of the things that I wanted to look for rather than sort of getting the next day on keynotes and then weeks later getting kind of a bulk push of all the sessions. And so I will say ever since I started attending remotely, I get to attend a lot more sessions than I used to. [00:07:25] Speaker C: I think the beginning of the week they were pretty quick on most of the talks. Getting them out onto, you know, the keynote came out pretty fast. They had, you know, there may be like 10 hour delay at most on talks, but then by we got to Thursday, the Thursday ones took like into the next week. So it's like they, yeah, they had some burst capacity to really hit a lot of the, the key important ones around new announcements on Tuesday and then as you got into the more standard fair, they didn't have that priority. [00:07:51] Speaker A: Ah, okay. I will say I, I was waiting for Swami's and I never actually got to it. [00:07:56] Speaker C: I mean his wasn't out until the following, you know, the morning of Warner's keynote because I was looking for it Wednesday night, I couldn't find it so. Well, let's talk about how we did in our predictions. So Matt and Ryan were not here. Matt was able to send in his predictions in advance, Ryan was not. And so Jonathan played him using AI. [00:08:16] Speaker A: Which slowly, slowly being replaced. [00:08:20] Speaker C: Yeah, but I'm saying that both of you did better on this reinvent prediction show than you've ever done before. So maybe this is the right strategy for you both. [00:08:28] Speaker A: I'm never voting again, that's for sure. [00:08:30] Speaker D: I think I got one last year. [00:08:32] Speaker C: I think he did too. [00:08:33] Speaker D: Now I want to look back at the show notes. [00:08:36] Speaker C: So Matt got something new on S3, which we'll talk about a little later, which is S3 metadata service and S3 tables, which we'll talk about a little bit later. Ryan got expansion of AI driven workflows in data lakes, which I was shocked on because when that one came out of the AI, I was like, there's no way. So yeah, no, they gave you a bunch of Q and bedrock capabilities for data lakes and security analysis of data lakes. So. So yeah, I think that one was firmly there. Jonathan scored zero. Sorry. Although you had some good ideas. A couple of them kind of got nailed in pre invent announcements. So you were on the right track. Just. [00:09:11] Speaker B: Yeah, I'm gonna, I want to talk about the, the one that I thought was going to win and then didn't, which was the automated rag or the vector thing because I. Yeah. [00:09:20] Speaker C: To S3. [00:09:20] Speaker B: Yeah. So I think the new S3 metadata service as announced is. Is pretty garbage. Really. All it is is just a shortcut to listing the contents of your bucket. The default contents of the iceberg tables is literally just the key names, the file sizes, the encryption type. It's to avoid people doing massive recursive lists on trillions of objects in S3. So it probably saves Amazon a ton. [00:09:46] Speaker C: It really solves them a problem. Not really. [00:09:47] Speaker B: Yeah, it solves their problem. But digging into the service a little bit more, which maybe we can do later, I don't know. Like, you can extend those, those links so when you query the metadata service, you can actually link to your own tables. So it's not. Not a big step as you put objects into your own bucket to actually find the embeddings for the text and then save those to a table as well. And then you can use the S3 metadata service to pull out those embeddings and you could use them directly with a large language model. So I think, like, they're halfway there. I would be really surprised if within the next few months this, what I predicted, didn't become an option. [00:10:26] Speaker C: Yeah, I, I suspect it's. If it isn't my next reinvent, it'll be at next year's reinvent is my feeling. [00:10:32] Speaker A: I'm guessing it was a missed reinvent, you know, target. [00:10:35] Speaker B: Maybe. I wouldn't be surprised. Like it's so such an obvious thing given what they've already announced. Yeah, yeah. [00:10:43] Speaker C: Well, you know, that's the way it work, you know, sometimes, Jonathan, that's the. [00:10:47] Speaker B: Way it goes every year. [00:10:50] Speaker C: You're just, you're just too, you're. You're on ability to execute. You know, your developer's execution at Amazon is not good. So that's your problem. I got new element multimodal replacement or upgrade to Titan, which not only did I get multimodal, but I also got Titan got deprecated and replaced with Nova. So, yeah, that's a double win in my book. [00:11:08] Speaker A: I miss that. They totally deprecated Titan. [00:11:10] Speaker C: I. Yeah, Titan Zoundova. They rebranded it. [00:11:13] Speaker A: They just rebranded it. [00:11:14] Speaker D: Okay. [00:11:14] Speaker A: Yeah. [00:11:15] Speaker C: And then I had mentioned that competitive VM offering to Broadcom, and I didn't get this point. I'm not arguing for this point, but they did say on stage, they did announce a new VMware service they talked about on stage, but they also announced Amazon Q developer to help you migrate your VM workloads. Now, I don't really know what that means. For a Q developer because unless your VM is completely built by code, I don't see how that's really going to help you. But that's, you know, it's interesting they directly had an attack against Broadcon in some way. Even if not, I think it's bullshit. So I was, I just, I put that in parentheses here in the thing so you guys see that. But yeah, so we basically ended up with a three way tie between Matt and Ryan, who can't win, and myself. And then I also pulled out the tiebreaker as well because they said AI artificial intelligence 45 times on stage and I had said 35 and in price is Right rules that means I win because Jonathan said 72. Now Jonathan would definitely have won and I'll tell you this, if they didn't shorthand AI and artificial intelligence now with Bedrock or Q. So places where they would have said AI or artificial intelligence are now saying Q, they're saying Bedrock models. So we might need to retire this tiebreaker for next year, I think. [00:12:29] Speaker B: So the other reason it would have, would have won is if I'd actually gone with 36 like I said I was going to without felt that. [00:12:37] Speaker C: So another, another reinvent Allen. I think I got this one but it was good. I, I, I think, yeah, even AI Ryan came out with impressive results. So that's, yeah, maybe you should just use AI to help inform your decisions next time around. [00:12:50] Speaker A: I mean, what, Yeah, I mean the, the proof is in the pudding there, so. [00:12:54] Speaker B: Well, to be fair, AI generated like 20 different suggestions and I picked the top three. So there was some human involvement in picking those. [00:13:01] Speaker A: All right, so that's, that's where I'll fail is I'll get the same 20 and then I'll pick the three. I think is very interesting. And I won't get a single question. [00:13:08] Speaker C: Yeah, I was trying to convince Drop that. To pick the three worst. Yeah. [00:13:14] Speaker D: What was your prompt for that? [00:13:15] Speaker B: What was my prompt? My prompt, I used the pre invent all the pre invent announcements and titles and I asked Claude was it or is it Chat GPT? No, I think it was Claude to look at those and then predict which services may follow on from those services that have been announced. Pre invent up at that point, which. [00:13:37] Speaker C: Was still, there's still another 100 or so pre invent answers that came out after recording. Right. Yeah. [00:13:44] Speaker D: So next year what we're going to do is we're going to choose the top 10 of each of those or go through. We'll have Chat GPT And Broad and Claude go at it and maybe we'll throw in Nova and then we'll bet. And that will be the tiebreaker. We'll bet which one has the most announcements that's accurate. [00:14:02] Speaker B: Oh, yeah, we can just do that. Just get three AI predictions and we'll bet on the one that's going to do. [00:14:07] Speaker C: Yes. [00:14:07] Speaker B: Yeah, we could do that. [00:14:08] Speaker C: I mean, that sounds great. So the guy who wants to write the show notes have to go listen to in that context of doing that. So, yeah, I'm, I'm out on this one. But you guys are going to do that homework. [00:14:19] Speaker A: Doesn't AI do that part too? Like, aren't we fully replaced at this, this point? [00:14:23] Speaker C: I mean, it might someday. I mean, it was, it was pretty nice to, you know. So the way that we've calculated out how many announcements and how many times it's had artificial intelligence in the past is that I take, take the YouTube Downloader plugin from Chrome and I go to the video and I download it and I put it through a transcription service and then I basically do, you know, account using a simple Python script. That's how I typically do that. This year now, AI is built into YouTube. And so I just asked Gemini, right, on YouTube, how many times they say AI and it gave. It was amazing because not only did they tell you, like, they've said this many times, they gave me all the timestamps. So if you wanted to challenge this, I have all the timestamps too, which is great. That's awesome. And then I also was able to ask it, like, what are the announcements that I said on stage? And it gave me a list and the timestamps of all the announcements, I was able to cross check my list. AI definitely helps a lot. I'm not upset about it. [00:15:09] Speaker B: Maybe there's something next to we won't be doing the show and we just delegate it to somebody else. Something else. [00:15:16] Speaker A: A few weeks ago, we did the demo of the, the, the Pot AI podcast. I'm like, yeah, we don't have to do anything. [00:15:23] Speaker C: Yeah, just sit back and rake in the, the monies. [00:15:27] Speaker B: I, I heard somebody actually using, using that as another podcast. It's a theories everything podcast. And at the end of one of his interviews, he, he did like an AI summary of the interview. And it was, it was, it was literally the, the podcast, two podcast hosts talking about the content of the thing. It was cringy, but it was actually really, really good. Like, if I hadn't just spent two hours listening to his actual interview Show. I would have spent more time on it, but I listened to about 10 minutes of it and it was really, really good. [00:15:57] Speaker C: Didn't still have the, the situation where the woman. [00:15:59] Speaker B: No. [00:16:00] Speaker C: Just kind of the throw to host or was it better? [00:16:02] Speaker B: No, no, no. It was still a man and woman, but the, the woman was much more involved. It wasn't, wasn't just the, the bit on the side. [00:16:15] Speaker C: All right, well, I'm sure. Go ahead, Brian. [00:16:18] Speaker A: I was just laughing because you're saying sit back and rake in the monies. And I was like, yeah, except for the amount of money we're going to spend on tokens on AI Premium. [00:16:28] Speaker C: All right, well, let's get into Pre invent. I didn't count how many there were, but it's over. Over 180 some announcements for sure. Um, so I. If you want to know all of them, you can go to our show Notes and see them all, but we're only going to talk about the ones that we thought were actually kind of interesting to us. And so the first one up was a time based snapshot copy for EBS. So now you can specify a desired completion duration, 15 minutes to 48 hours when you copy an Amazon EBS snapshot within or between Amazon regions or accounts. This will allow you to meet your time based compliance business requirements for critical workloads, mostly around doctor Capabilities, which I'm just glad to see this one finally because having it built in directly to the console to guarantee that my EBF snapshots make it to the other region is a big quality of life enhancement. [00:17:17] Speaker B: Do you pay more to have it move faster? [00:17:21] Speaker D: I was looking for that when I was reading over this. It looks like you. There's throttling throughputs so you get X number of megabits per second between regions is what I saw. And then there is a completion duration price too, per gigabyte. So like you can't just say everything goes fast because they still have a throttling limit, but there is then a price point at that point also. [00:17:47] Speaker C: Yeah, it says throughput. There's a default per account limit of 2000 Mbps between each source and destination pair. And if you need additional throughput in order to meet your rpo, you can request an increase by the support center with a maximum per snapshot throughput of 500 megabits per second. And that one can't be increased. [00:18:03] Speaker D: I mean it's 2 cents per gigabyte if you want it in 15 minutes. Like that's not cheap too. That will Add up quickly. [00:18:14] Speaker A: Okay. I miss. I misunderstood this when I read it the first time. I was. Because I thought it was like, oh, cool, you can set it to backup every 15 minutes. And I was like, I don't care. But now I guess it's actually saying you can speed up the transfer to make it within 15 minutes, which I didn't understand. Okay. [00:18:29] Speaker D: Yeah. [00:18:30] Speaker A: Yeah. [00:18:30] Speaker B: But even. Even up to 48 hours is a half a cent per gig on top of the standard thing, which is kind of worrying because I really hope that it would finish within 48 hours or even 24 hours or even 12 hours without paying extra. [00:18:43] Speaker A: Yeah, right. How big is that volume? [00:18:47] Speaker B: Yeah, it's not. It's not bad. To be fair, if you. If you need RTO and then. Then 2 cents a gig on top of the transfer cost is not terrible. [00:18:59] Speaker A: Yeah. Why? As long as it's not in addition to whatever transfer costs for inter. [00:19:04] Speaker B: It is in addition to. Yeah. [00:19:06] Speaker A: So, yeah. [00:19:07] Speaker D: Yeah. [00:19:07] Speaker A: Why. Why am I paying twice if I want it in 48 hours, 50 minutes? I guess if you have the compliance can meet. [00:19:13] Speaker D: Yeah. It's contractual concerns. You can meet your RPO and definitively state to a customer that you are going to have this. Be there within this time frame and solve that. Check that box. So you can say, I have an RPO of 24 hours and guarantee it essentially, versus before you're like, it gets there. When it gets there. [00:19:35] Speaker A: Yeah. [00:19:36] Speaker C: Best of luck to you. [00:19:37] Speaker B: Yeah. We did a backup, but it took three weeks to get there. Yeah. [00:19:41] Speaker C: Yeah. It doesn't really hit your RPO target. [00:19:43] Speaker B: No. That's good. Hopefully they do the same thing for S3 replication as well. [00:19:47] Speaker D: Actually, between regions, I think there is already actually. There's definitely a feature of S3 replication where you can set, like the time of how fast it is. [00:19:57] Speaker B: Okay. [00:19:59] Speaker D: Controllable time for predictive replication. SLI 2019. S3RTC is what it's called. I knew this because I wrote an exam question about it a long time ago. [00:20:13] Speaker A: Nice. [00:20:14] Speaker C: All right, let's move on to the next thing because we got a lot to get through. Keep us moving here. Amazon and Anthropic deepened their collaboration with another $4 billion investment from Amazon to also use their Neutrinium 2 chips, which came up later on Main Stage at Monday Night Live and as well as on Matt's presentation, basically around their new Ultra cluster and some 64 trainium 2 chips in a single box. Pretty impressive. So basically, Amazon gives the money with the right hand to Anthropic and Then Anthropic gives it back to Amazon on the left hand by spending it on aws. But good to see the expansion of that partnership continuing. All right, this one caught all of us by surprise, so we said, no way. But new physical AWS data transfer terminals let you upload to the cloud faster. So we got rid of the trucks, we got rid of the disks that we send you in the mail. But if you have your own disks that you'd like to bring to a physical location in either Los Angeles or New York, you can connect them with the cable directly to the Amazon cloud through a public endpoint that is available, I assume, in a secure building or something. You know, it wasn't really clear to me from just reading through the article, but basically you reserve a time slot to visit your nearest location and upload that data quickly to your AWS public endpoint. So another way for you to move your data by a sneaker net. [00:21:35] Speaker A: Well, so like, Jeff Barr, you know, did the remote scouting for this, and it was like, when I arrived at the building, I was kindly escorted to be able to work easily in the instructions provided at the time of reservation. They escorted me to a room in the back with a. Was it like a. A patch panel and a server. It was kind of nuts. [00:21:54] Speaker C: Like, wow. [00:21:56] Speaker A: It's kind of crazy. Yeah. [00:21:58] Speaker C: I mean, the pictures, though, you showed of the building, like, I don't think that's an Amazon building, but it's kind. [00:22:04] Speaker A: Of like some generic office building. Yeah, yeah, Very, very generic looking. Yeah. [00:22:11] Speaker C: Yeah. This is the pilot location in Seattle, so maybe it was just because it's a pilot. Hopefully the real ones are a little nicer. [00:22:19] Speaker D: I kind of want to go and just bring like, you know, five gigabytes and just see what it is. [00:22:24] Speaker C: I want to bring my. My one terabyte Western digital external drive. Like, hey, I'm gonna upload this to my S3 bucket. [00:22:30] Speaker D: Yeah, upload my backups go into New York once a month and upload directly. It's 300 per port for Dan per data transfer, but no ingress fees. [00:22:43] Speaker B: Yeah, no ingress fees anyway. [00:22:45] Speaker D: Yeah, you're not supposed to know that. Yeah, they call it out. No per gigabyte charge for data transfer. [00:22:55] Speaker B: Okay. [00:22:56] Speaker A: All right. [00:22:56] Speaker C: I mean, is that down? I can download to this thing as well. Can I download it? [00:23:03] Speaker D: No, no, no. [00:23:05] Speaker A: Yeah, it's. It's very specific to upload in all the documentation. So I assume that, although it'd be. [00:23:11] Speaker C: Funny, it would be fun to see that. Well, in an announcement that would have Been perfect at reinforce or on mainstage. They've announced that the new AWS Security Incident response service designed to help organizations manage security events quickly and effectively. Service is purpose built to help customers prepare for, respond to and recover from various security events including account takeovers, data breaches and ransomware is now available to you. So all for your security incident response team to be able to address issues in your organization quickly and efficiently. So that's, that's pretty cool. I like this one. [00:23:44] Speaker A: Was it a new service? I didn't really understand this one like. Or was it. Or is it like the Security Incident Response sort of team or. [00:23:51] Speaker B: No, it's a new service curated experience. [00:23:53] Speaker C: I mean it automates the triage and investigation of the security funding from GuardDuty and then integrates that into Security Hub. And then there is 24,7 access to security experts from the AWS customer cert team to help you address issues of account takeover. Those things, I mean it's partially, it's humans, it's human spackle. [00:24:12] Speaker A: Yeah, I just could have sworn that all the integration between GuardDuty and Security Hub already existed. So when I was playing with it before this, I'm like, I thought that that was a thing. [00:24:22] Speaker B: That's what they wanted you to think. [00:24:26] Speaker C: That's what marketing told you. But it never really worked the way you thought it did. [00:24:30] Speaker A: No. [00:24:33] Speaker D: I mean this is their competitor to Copilot Security. You know that Microsoft is pushing very heavily. [00:24:41] Speaker C: I'd say it's more of their competitor having Mandiant in the back door at Google. Like oh hey, you need access to a security expert. Here's Mandiant. Like, hey, we have, we have a security cert team. Their best, they're Amazon specialists. [00:24:56] Speaker B: There are a lot of cloud cost management tools out there, but only Archera provides cloud commitment insurance. It sounds fancy, but it's really simple. Archera gives you the cost savings of a one or three year AWS savings plan with a commitment as short as 30 days. If you don't use all the cloud resources you've committed to, they will literally put the money back in your bank account to cover the difference. Other cost management tools may say they offer commitment insurance but remember to ask, will you actually give me my money back? Our chair A will click the link in the show notes to check them out on the AWS marketplace. [00:25:35] Speaker C: All right, keeping moving here. AWS private link now supports cross region connectivity. This until now interface VPC endpoints only supported contribute to VPC endpoint services in the same region. This launch enables Customers to connect to VPC Endpoint services hosted in other AWS regions in the same AWS partition over interface s endpoints which having designed some of these network things, this is handy because some of the limitations of being restricted to specific regional targets was a bit difficult. [00:26:02] Speaker D: Yeah. Especially if you have a customer in a different region or you then had to launch your infrastructure in that region where your customer was just to have that private link to them. So this will solve a lot of, you know, resellers and you know, of services problems. [00:26:18] Speaker C: Another networking enhancement, Amazon Cloudfront now announces VPC Origins, which is a feature I've wanted forever. Basically this allows a customer to use cloudfront to deliver content from applications hosted in VPC private subnets. And with the VPC Origins customers can have their ALB, NLB or EC2 instance and that private subnet is accessible only through their Cloudfront distribution. So you don't have to do the dance where you go from Cloudfront to a public endpoint to go to your private endpoint anymore. [00:26:43] Speaker B: Yeah, that's nice. I kind of assume that this depends on the private link connectivity. So they also announced. [00:26:49] Speaker C: Yeah, I would assume so. [00:26:50] Speaker D: And this solves a lambda that micro that they posted I think probably like five, seven years ago, which was just a lambda that watches the public endpoints IP addresses for Cloudfront and just would update your security group rules so that you could only have that accessing it. I think I've deployed like 30 times and every time you have to do a security group expansion because it's over 50 IP ranges. It's always fun. [00:27:17] Speaker C: Yeah, sounds terrible. So yeah, so now being that little loophole would be. Is a Plus. [00:27:25] Speaker D: Yeah. [00:27:26] Speaker C: Amazon EC2 has taken the great container lineage container lineage capabilities you have there where you can see where the container got created and then how many times people added or modified to it, brought that to amis. So if you want AMI lineage, you can now get that you can easily trace and copy or find the derived AMI back to the original AMI source through the records, which is important for some organizations who have heavy duty spawn requirements and or you know, they have image factory type solutions that basically create golden images amis and they need to be able to see if it's the right one. [00:27:59] Speaker D: Yeah. And also lets you track all the updates and security. So if you start to say hey, in this ami you built off that one, it's easy to kind of track that whole lineage, which is great and will solve lots of problems for sure. [00:28:14] Speaker C: Amazon Database Migration service now supports data masking, allowing you to automatically remove sensitive data at the column level during migrations to help comply with gdpr, et cetera. Which makes DMS now even more interesting if you're trying to keep a dev environment replicated with somewhat accurate production data without having the actual customer data there. So that's, you know, DMS is more than just the migrations. It can also keep things in sync, which is this nice capability that you don't have to build in glue or some other terrible ETL process. All right, next up, Amazon S3 now supports enforcement of conditional write operations for S3 general purpose buckets using bucket policies. This enforcement of conditional rights, you can mandate that S3 check the existence of an object before creating it in your bucket. Similar, you can also mandate the S3 check the state of the object's content before updating your bucket. This will help you to simplify distributed apps by preventing unintentional data overwrites, especially in high concurrency and multiwriter scenarios. And if only someone had created S3 fuse type product that might have this problem that needs this. Oh, Amazon did. That's right. Because after they told us for years not to do S3 fuse, they finally built it and now they're realizing oh yeah, if we just gave them this ability, you could solve a big problem. So I appreciate it only took you actually living the pain that we all live for years to fix it. [00:29:29] Speaker A: Well, and the fuse interact integration that they, they released was very limited. So it's like it was. I, I still wouldn't use it for any kind of asset based like check ins or anything of sensitive data or important data. [00:29:46] Speaker C: I was really hoping for systems manager to get some updates. This year they did not. They got a new ui that makes it worse. They now streamline your node management, now provide you access to your, you know, you quickly be able to see if it's an instance, if it's an on prem instance or if it's a hybrid instance on top of Outpost or something else. So wasn't quite what I was looking for in the systems manager improvement camp, but that's what they gave us. [00:30:14] Speaker A: Yeah, I'm still just mystified by the story of systems manager and like, like the problem they were trying to solve in the ecosystem they were trying to create. I feel like they went somewhat down a path and then would realize that no one wanted this and then pivoted into this weird now abomination of a service. There are parts of it I love. [00:30:40] Speaker C: Yeah, I mean there's only parts that are good, but it's always about like, okay, you have all these core Amazon teams and they've been running forever and they have their way of doing things and the way of expectations and all that. And then someone said, well, we really need assistance manager functionality to manage all these different things. And instead of taking people from those teams who knew how things were built, they took a brand new team and taught them all brand new things and said, build a service, manage Amazon. And then they built this. That's why it feels so disconnected from the other paradigms you're used to inside of Amazon. So I was hoping for more. They did integrate Q developer into it, if you want to use that as well as you have a couple, you know, some new dashboards help you manage unmanaged nodes as well as identify reports you can download. So, you know, they had a couple other little improvements, but nothing, nothing too exciting. All right, moving right along here, the next one up in the pre invent side is new enhanced root cause insights to help explain cost anomalies. They'll tell you why your cost just ballooned three or $4,000 without you having to go figure it out yourself, which is handy. They also gave you a new savings plan purchase analyzer which allows you to quickly estimate the cost, coverage and utilization impact of your planned savings plans purchase. Which is kind of the opposite of giving you the prediction or like giving you the recommender. It's now saying, okay, if you bought the recommendation, here's what it actually would do. So now you get both directions of modeling, which is good. And then AWS Compute Optimizer now supports Idle Resource recommendations for you as well. So three nice finops improvements. [00:32:07] Speaker A: Yeah, the Idle resources is interesting because I thought that that was the whole point of underutilized, you know, instances so you could optimized down to a different instance type. But I'm not sure what the level separation is that they're announcing here. [00:32:24] Speaker B: Well, I think before it was just based on CPU time, which is just a terrible metric for something that you need to have running. But I'm assuming that they're bringing more data here, like network traffic and various other things as well to decide whether or not something really is idle. [00:32:38] Speaker C: Yeah. So the new Idle Resource recommendation, you'll be able to identify idle EC2 instances, EC2 auto scaling groups, EBS volumes, ECS services, ring and forgate and RDS instances. And you can view the total savings potentially of stopping or deleting those idle resources. So it's using still CPU but just more services are getting it. So maybe not what you want Jonathan. [00:32:55] Speaker A: I think it's more. Yeah. More incomprehensive of the just. Instead of just the instance level the stuff around the instance. So. Okay. [00:33:03] Speaker B: Yep. [00:33:03] Speaker D: I actually really like the root cause insights for anomalies. I feel like at my day job we have the anomaly detection step on Azure and we just get these things and it's like, hey, this resource group is up 3% and then you have to go dig in and spend time to say like hey, why did this increase? What changed in the last 24 hours? It's like, oh, and by 3% of this resource group it really. There's nothing in there. So it was a 14 cent increase. Thank you for wasting my time. [00:33:31] Speaker A: Yeah. [00:33:31] Speaker D: So you at least could in theory. I haven't looked at this service but at least from what I understand from it should give you a little bit more information off the bat. [00:33:40] Speaker C: Yeah. [00:33:40] Speaker B: It's still annoying though because even if you bought savings plans and you know, turning it off, you're still going to spend the money. It still tells you to turn it off. Like this instance is idle. Yeah, but I'm paying for it anyway. [00:33:50] Speaker A: Yeah, yeah. [00:33:52] Speaker C: It would be nice if they connected those things together. Like oh well, but if I lowered your recommendation, you know, your utilization rate of your savings plan, we're not actually helping you. [00:33:59] Speaker A: Right. [00:34:01] Speaker C: And then the final feature that was a pre invent that I can't believe didn't make mainstage. They now Amazon Lambda now supports application performance monitoring OR APM via CloudWatch application signals. This gives you the ability to see the health and performance of the service application built using Lambda and makes it easy for you to identify and troubleshoot performance issues to minimize the MTTR and operational costs of running your serverless app, which you only wanted for a thousand years to have better telemetry inside of Lambda. So thank you finally delivering that. [00:34:30] Speaker B: Yeah, it didn't have Q or AI in the title though, so it's not going to make main stage. [00:34:34] Speaker C: Not going to move main stage. Yeah, yeah, that's the problem. So that was, that's. That's just the a tip of the iceberg of things that we covered from Pre invent. I mean there was things in here that we didn't talk about. Like Workspaces now supports Rocky Linux for people who want to punch their users. There's a ton of Q developer, business and app stuff that they announced as well as things like SES Mail Manager, as delivery of email to Q Business apps, which why would you want that? I don't know. But if you need it. It's there a bunch of bedrock improvements, CloudWatch improvements around application transactions, additional monitoring for local zones. There's new cloudformation recycle bins for those of you who like to delete things accidentally. You know, RDs, blue green improvements. Just ton of stuff here. All good. We just. This show would be 12 hours long if we went through everything. So check it out in the show notes to see what else got announced. Pre invent, as we call it. [00:35:29] Speaker A: To be fair, anyone requesting a Linux based workstation is already used to pain and punishment, so they're not even going to notice. [00:35:36] Speaker C: Yeah. All right, let's get into the main event with Monday Night Live, the lovely talk. And this is a good one this year if you're into how do they build some of their compute infrastructure, especially these new Ultra clusters. They talked about the Trainium chips, the nitro process for those. They only had one announcement during this keynote which was that they are giving you new latency optimized inference for foundational models by Bedrock. But just for those of you who are hardware nerds, this is the talk for you to watch. [00:36:07] Speaker A: I learned a ton. And I put learned in air quotes because most of it went over my head when they were talking about the structure and the design of the silicon at a very low level. But mondaylight Knife continues to be my favorite keynote of the conference for a long time now. [00:36:23] Speaker C: Yeah, it's always one of the highlights of the week for me, for sure. I think you and I watched it even a little bit live, didn't we? [00:36:30] Speaker B: Yeah, it's hard to kind of connect to as a, as a consumer or a user because it's not off the shelf stuff like you can't, you don't read about it in PC Magazine and then think, oh wow, Amazon's deployed 10,000 of these things. It's like, no, they built this thing, they designed this thing for this very specific purpose and, and it's absolutely amazing and you're never going to get your hands on it. [00:36:50] Speaker A: Yeah. [00:36:52] Speaker D: But this is for the nerd in us. [00:36:54] Speaker A: Yeah. [00:36:54] Speaker D: Like we want to technically know how it's done. [00:36:57] Speaker B: Yeah. Like it's, it's looking at fancy sports cars or something for, for cloud guys. So. [00:37:01] Speaker A: Yeah, I mean, I'm not smart enough to get into building my own silicone chips. I know that I can just call that right now it's never going to happen, but I can still like nerd out and see how they're doing it. [00:37:11] Speaker C: I mean, there's a CEO job open at intel right Now, Ryan, so you could build your own silicone, you'd have. [00:37:19] Speaker A: To explain to me what a CEO does. [00:37:21] Speaker C: Yeah. [00:37:22] Speaker D: Sits in meetings. Meetings. [00:37:25] Speaker A: Oh, that I'm qualified. [00:37:26] Speaker D: Yeah. [00:37:29] Speaker C: All right, well, Matt Garmin, first introduction to Matt. Who he's a pretty good public speaker. He, I'd say he did the worst version of what is AWS I've ever seen from a keynote, a CEO. But once he got past that five minute section, he, he was right into the weeds. He was smooth. Did a great job after that, but, so I won't mark him too bad, you know, for getting into the intro. Intros are hard, especially in a keynote of that size and that number of people. Any thoughts on Matt and his presentation from your guys take on it? [00:38:00] Speaker B: He does what I do when I'm reading the, either the show notes out or doing a presentation. I like, comment on what I'm saying. [00:38:07] Speaker A: Yeah. [00:38:08] Speaker B: I'm like, oh, isn't it? And that, and that's just great. And like the number of times he was like, isn't that great? Like, that's just great. He's just so excited about what he's doing and what they're building. It's, it's really good to see. So I'm not marking him down for it. It was just an observation that he kind of couldn't hold his own enthusiasm for these things in, which is why the keynote went over by 15 minutes, I think. [00:38:30] Speaker C: Yeah, yeah, he was there. Well, you also, you have, you know, several customers come on stage, but you can't really control them the same way. And you know, it's always difficult. [00:38:40] Speaker B: The one, the one very, very slow Australian customer, I think, who is talking abnormally slowly even for a keynote presentation. [00:38:51] Speaker A: Yeah, no, I thought it was, you know, for, you know, since we've had a little bit of a rotating AWS CEO in the last few years, like, definitely a great first appearance. You know, Adams took a little while to get, you know, I think by the time he was done, like, you know, it was pretty well polished, but the first few were a little rough. Yeah, I think, you know, I, I, I think the passion was there, like Jonathan said, like you could see it and I, you know, I, I don't know, like there was something a little bit more invested in what were bought in or just excited and you know, maybe that was just, it seemed like that was sort of the tone of Reinvent altogether pretty exciting. [00:39:28] Speaker C: Yeah. Well, he also brought on Andy. So we had a surprise guest with Andy. I wasn't very surprised. They talked about it I think for two days prior to it happening. But Andy came back to talk about some AI ML things of course, but no, of course I can remember what he even announced, but it wasn't anything exciting. It was beyond the other things. But let's get into some of the announcements from Matt. So first up is the interesting Amazon Nova, which is a new frontier intelligence and industry leading price performance LLM solution. This is replacing Titan. So this is basically the next generation of Titan with a new brand. It comes in four model sizes. An Amazon Nova Micro, an Amazon Nova Lite, Amazon Nova Pro and Amazon Nova Premiere. The Pro Lite and Micro are available now with Premiere coming sometime later next year, which will be a complex reasoning model. So those are still pretty new. That won't be out until they said Q1 or maybe even Q2. Amazon Nova also understands RAG functions. It has a comprehensive set of benchmarks already that you can check out and see if it compares to other things. It also has a couple additional components including the Amazon Nova Canvas which gives you image generation model so you can basically do image creation Dall E and then Amazon Nova Real, which is a state of the art video generation model which will explain why all of your Amazon prime content has gotten really bad in about two years. [00:40:53] Speaker B: Well, like it's gonna like QVC basically. Amazon.com is gonna be like a live custom QVC feed autogenerated by AI. Notice you were just looking at this thing and now you're looking at this thing and I know I've got just the thing for guys like you. It's. It's going to be absolutely sickening. [00:41:08] Speaker A: Yeah, I mean it, it was already getting that with, with live, live actors and live, you know, stream kind of doing the same thing. It was. So yeah, it's just going to proliferate to a larger amount of them. [00:41:21] Speaker B: I kind of wonder why they built their own thing though. Right? But like it. Why, why build your own model at this point? I mean I know the market is exactly saturated, but there's an awful lot of good models out there. And if you're partnering with somebody like Anthropic, why not just use. [00:41:39] Speaker C: I mean why is. Why is Microsoft creating their own models, you know, and they're partnering with ChatGPT I think, you know, and actually Swami talked about it is that it's not going to be one model to rule them all. It's going to be a bunch of specialized models for special use cases or special needs that make sense to them. [00:41:55] Speaker A: But these aren't specialized. [00:41:57] Speaker D: That's the Problem. Exactly right. [00:42:00] Speaker C: I don't know. They're foundational first and then they get specialized from there. [00:42:03] Speaker A: No, I agree. I, I do think that this is going to be a whole bunch. But it is funny. I do, I think it's really more of, you know, how many digits of PI can we calculate? I really do feel like that's, it's exciting. It's the new hotness. And that's why is developing these giant big models. [00:42:19] Speaker B: Yeah. And they are quite different, like architecturally. I was digging into slight aside now I was digging into the Transformers Python module. I was like, well, how exactly does this work? Because all these models have different architectures, different numbers of layers, different numbers of whatever. I was like, actually the Transformers Python module is literally just a community assembled thing which supports every single type of architecture that you could possibly want. So, you know, when Quinn comes along from Alibaba Cloud, they make a PR to the Transformers model so that it supports the Quinn models. I was like, oh, that makes sense then. Because otherwise how could you possibly support all these differences? But yeah, anyway, carry on. [00:43:05] Speaker C: He also brought on Amazon S3 Tables, which is their new native Apache iceberg format support port inside of S3. It comes as a competitor to parquet files. And this allows you to have basically table buckets that can act as iceberg tables, which can be handy for your AI ML use cases and training models. It makes sense they built that for that. And then also the inquirable object metadata for Amazon S3 buckets, which the guys kind of mocked earlier. This is basically providing a rich metadata service that tell, you know, so you can store 20 elements, including the bucket name, object key creation, modification, time, storage, class encryption, status tags, and user metadata that you can define. They showed on stage an example that's using a Hike image and basically showed several of the parameters of an image, including the image size, et cetera. So you have that available to you in the metadata without having to download the object or some portion of the object to determine that data. [00:43:58] Speaker A: Yeah, I can't remember if we were actually making fun of this during the show or when we were just preparing for the show, but it's definitely a feature for Amazon themselves, because it was. I've abused Amazon S3 queries for this exact purpose. I'm sure I wasn't alone. [00:44:15] Speaker B: Yeah, I mean, how do you. How do you build a service that does billing for S3 that has to troll through tens of trillions of. Of objects anymore without having a something like this to. To help you? [00:44:25] Speaker C: Yeah, so I mean, it's definitely nice though. I, I do appreciate being able to like put the object size into the metadata. There's some just dumb things like oh, like what files am I trying to clear out of a bucket? [00:44:37] Speaker A: Things like that. Oh, the amount of dumb things I've done. Yeah. With like trying to do labels or like actually, you know, like it wasn't that long ago where you know, we're running a service to sort of put that data into a data storage for other for querying outside. So yeah, no, this is great. [00:44:53] Speaker B: Yeah. [00:44:54] Speaker C: Then you know, he went a little bit off the deep end talking about Q and all the different Q things they have now. Agents action integrations, Q business capabilities, transformer capabilities for.net applications and mainframes and VMware workloads. Investigator immediate operational issues with Amazon Q developer in preview, which is sort of a nice way to be able to like query your infrastructure's code while you're troubleshooting an issue. Like why can't I talk to this node and it'll give you information from Q And then. Interesting. Also a GitLab Duo integration with Amazon Q. So if you're using GitLab, you can natively query into your code from Q as well. So a bunch of things in the Q continuum, as I called it, and then several things in Bedrock including multi agent collaboration capabilities, a new factual error prevention for LLM hallucinations with mathematically sound automated reasoning checks as well as ability to build cost efficient, highly accurate models with Amazon Bedrock model distillation. And since none of us care about either of those two areas, Q or Bedrock, I'll share for you guys, see if you have any comments, but otherwise we'll move on to SageMaker 2.0. [00:45:59] Speaker A: I mean, I like hallucin. I like hallucinations because I think it's funny. I still think that's funny. I don't like it when my AI gets me wrong. Wrong answers. But hallucinations is hilarious. [00:46:10] Speaker B: I think we're misunderstanding it and this isn't really a topic for reinvent but since we're never going to talk about it any other time, I think we're really misunderstanding what hallucinations really are. If you think about where ideas come from for people, it's like, you know, we have this knowledge graph in our brains and ideas form where we start making connections between things we already know that weren't really there. Like we don't have confidence that there's a relationship, but you kind of have this idea that forms and then we kind of investigate it. Like actually, yeah, that is a thing. And then it becomes, you know, like a solid concept. I think the problem with AIs is they, they have no, like, they have no way to check if the sanity check, these ideas that they come up with. So I think a lot of hallucinations are things that actually make sense in the right context. Very often just like when it writes, you know, the names of scientific papers, those are the names of papers that the people may well have written or should write or something else. So like, I think, I think hallucination is a little misunderstood. [00:47:07] Speaker A: Oh, I think it's perfect. In fact, I'm reversing it, which is every time I'm wrong. I'm not wrong, I'm just hallucinating. [00:47:13] Speaker B: Oh yeah, that's good. [00:47:16] Speaker D: I don't like when Jonathan explains things at times because I just think I went crazy thinking about what he just said. [00:47:23] Speaker B: So I'll share with you. I had a long conversation with Claude because I was doing some coding. There was a hallucination and in some code it had called this method, which I don't have in my code. And I kind of called it out. I was like, well, this doesn't exist, but it should exist. So thank you. It solved this really awkward problem that I was having, like, oh, we could do it like that. It was a hallucination which was actually incredibly valuable. And if it had these checks in place which would reject things like this. Well, he didn't tell me that I could write the thing like this. So therefore I won't, I wouldn't have found it. I wouldn't have had a novel solution to a thing. And so I went on and had this interesting conversation with Claude about hallucinations and about kind of like self reflection and how ideas are really very, very similar concept in, in human terms. And. [00:48:20] Speaker D: Did you solve hallucinations in Claude by this conversation? [00:48:23] Speaker B: I haven't solved hallucinations. But, but, but where I got to was that giving, giving them more time to think is a really good way of reducing hallucinations or turning hallucinations into something which are more useful and solid. And so I kind of came to this agreement with Claude and I have a system prompt that I'll share with you where if Claude doesn't feel like it's had enough time to think about the answer. Because if you think about it, it's always very user driven. Like user asks a question, Claude generates a response and the amount of time it can spend doing inference is limited by the architecture itself. And so I have this agreement in the system prompt with Claude. Now that on complex topics which require more thought, it will ask me to give it a moment of time and I just type the words time passes and then it can reflect again on what it's just said and on the context. And it's been super good at generating some really good answers to tough questions just by giving it a little extra time to think. [00:49:25] Speaker A: I never thought of that. That's amazing. Yeah, yeah. [00:49:28] Speaker B: It's just so freaky, like halfway through a conversation and it's like, can you give me a moment to think about that? [00:49:33] Speaker C: It's sort of what the idea of the reasoning models is, right? [00:49:36] Speaker D: Yeah. [00:49:36] Speaker C: Because the reasoning model, the whole idea is that you're giving it more time to do that thinking introspection, which is why it's kind of, everyone says it's kind of the next generation of AI in that reasoning model, that exact reason. So it makes sense. [00:49:50] Speaker B: Yeah, but good thing is you can do it with models that aren't reading models by just providing that kind of structure externally. [00:50:01] Speaker C: All right, now, SageMaker is now the kitchen sink. We're going to throw everything into it, so it's going to be very confusing from now on out, but basically they're announcing the next generation of Amazon SageMaker as a unified platform for data analytics and AI. The all new SageMaker includes virtually all the components you need for data exploration, preparation and integration, big data processing, fast SQL analytics, machine learning, machine learning, model development and training, and generative AI application development. The SageMaker you know and love today is now known as Amazon SageMaker AI and it is also integrated with the next generation of SageMaker while also being available as a standalone service for those of you who still want that capability. But the new SageMaker is SageMaker Unified Studio, which is a single data and AI development environment, brings together functionality and tools in a range of standalone studios, query editors and visual tools that they have made that they have today, including Amazon Athena EMR Glue Redshift, Apache Airflow and the existing SageMaker studio. They've also integrated the Amazon Bedrock IDE, which is an updated version of Amazon Bedrock Studio to build and customize generative AI applications. In addition, Amazon Q provides you AI assistance throughout your workflows in SageMaker. So your key capabilities are the Unified Studio SageMaker Lakehouse for unifying your data across S3 data lakes, Redshift and third party and Federated data sources, data AI and Governance, which includes access to SageMaker catalog and built on top of Datazone data processing with all of your Athena EMR and glue model development using the foundational models and SageMaker AI and then your generative AI development built and scaled with Amazon Bedrock as well as SQL analytics for Redshift, all built directly into SageMaker Unified Studio and Preview. [00:51:43] Speaker A: I mean, SageMaker was already a kitchen sink for ML solutions, right? Like all the different things and it made it really difficult to sort of summarize what it was useful for. And now it's so much worse and. [00:51:59] Speaker C: It'S going to get confusing because you know you're doing something in SageMaker Studio. Say you're querying S3 and it's happening to be using Athena under the hood to do that. As a developer of SageMaker Studio, do you know that it's using Athena? So then when the FinOps team comes to you and says, hey, our Athena bill just blew up and they're like, we don't use Athena, we use SageMaker. I can see these conversations already. [00:52:20] Speaker A: Oh, absolutely. That was the first thought I had was it's this abstraction layer where you're going to be able, you're going to be using services that you don't know and I'm sure like you said on the back end, it's going to be directly exposed from a finance perspective, but that, that primitive service, it's gonna be chaos. [00:52:38] Speaker B: I'd love to play this stuff. I just don't have the money. [00:52:43] Speaker A: I don't have the money. I don't really have the problem statement, like I would love to do more data research like that. Like I've really, you know, like there's a couple things where I thought, you know, it'd be really fun to do, but it's so time consuming to set it all up and getting all that. Like it's, it's a, it's a job to get that just to set. Just getting it all in a place where you can run a query is a full time job. [00:53:07] Speaker D: Well, because then you're also like me who would say, well, I'm not going to set this up unless if I set it all up in code. So you really have to understand all the very detailed underlying IAM permissions and everything to make it all work too. Which is even more fun. [00:53:20] Speaker A: Yeah. [00:53:21] Speaker D: Then you hate everyone. [00:53:23] Speaker A: Agree. It's true. [00:53:25] Speaker C: All right. And then the last couple things from EC2, part of Matt Garmin's, which he admitted was his favorite service. He was the head of E2 for a while. So basically the Trainium 2 instances and Trainium 2 ultra servers for AI, ML training and inference are now available to you as well as the new EC2 P5EN, Nvidia H200, tensor core GPUs and EFAV3 networking is all available. [00:53:50] Speaker B: I thought that whole session was kind of amusing in the end because it's with this great partner with Nvidia. They've partnered with us, they've helped us build this amazing thing out. And by the way, and new chips kicked their ass. [00:54:04] Speaker C: Yeah, you shouldn't use them, you should use ours. Yeah, yeah, it was a pretty good, you know, knee jerk there through that. [00:54:11] Speaker B: Yeah, that was kind of funny. Those new Ultra servers are fantastic. And when training three comes out, if they're as good as they want them to be, I mean, the size of cluster looking at building for anthropic is like comparable with the top five or ten supercomputers in the world. Yeah. [00:54:28] Speaker D: What I found interesting is they were like, here's the Trainium 2. Oh, and by the way, Trainium 3 is coming out in like the next like six months. Like at least that's the way I kind of gathered it was like, why? I guess they needed the announcement to like get the hype out, but also like start with this, like move over to that. Like felt interesting. [00:54:48] Speaker C: And they didn't say anything about training M3, what it actually would be like. They just said, yeah, we're working on the next generation of Trainium 2 chips. And it's like, okay, but no details. So I guess they, yes, they pre announced it, but they didn't pre announce anything with substance. [00:55:00] Speaker D: Oh, I thought they gave a timeline, but maybe I just assumed, oh, they. [00:55:05] Speaker A: Give timelines all the time, they just don't meet them. [00:55:08] Speaker C: Yeah, nor is there any details of like, is it going to be 100% better than Trainium 2 or is it 0? I mean, training 1 versus training 2, it's a whole different architecture, you know, it's a pretty significant improvement. And the fact that you can do inference on a trainium 2 but you couldn't do on the trainium one is a big deal. And so that's part of the reason why they're going after Nvidia is that they think they can provide you a better optimization for inference in addition to training using training M2 over what you can do with H2 hundreds. [00:55:36] Speaker A: I love that we just discussed AI in that much detail and I still feel like such an AI idiot just every day. [00:55:44] Speaker C: Well then we should definitely move on to the Swami keynote then for you. [00:55:47] Speaker A: Yeah, yeah, here we go. Let's make me feel real dumb. [00:55:50] Speaker D: Yeah, Buckle up. [00:55:53] Speaker C: So I luckily watched this one at 2x later because I didn't watch it live, which I was very pleased about because it was a lot of AI, a lot of things here. Nothing we're going to really dive into too much. There are new SageMaker hyperpod recipes which was funny to me because they like. So it's a day later you're already announcing new features for SageMaker. You just announced as well as Amazon SageMaker partner AI apps are available for you there New Bedrock Marketplace for you to access over 100 foundational models in one place, which I didn't even know there was 100 of them now. So I just. It shows you how much I'm out of touch on that. They have a new reduced cost and latency with Amazon Bedrock intelligent prompt routing and prompt caching, which actually makes it easy for you to try across those 100 foundational which one is the right one as well as the right cost Use case for what you're asking it. Kendra got a call out here with new gen AI index for Kendra. I was like, I thought Kendra was dead, but apparently not. Oh yeah, Because I got a new feature. [00:56:43] Speaker A: I haven't heard that in a while. [00:56:44] Speaker C: Yeah. Amazon Bedrock capabilities enhanced data processing retrieval. Bedrock guardrails and support multimodal toxicity detection with image support so you can detect if people are using your AI to create images of celebrities and compromising positions. Amazon Q developer to build ML models on Amazon Sage or Canvas. And you can now solve complex problems with new scenario analysis capabilities in Amazon Q. Quicksight. All from Swami's keynote. [00:57:12] Speaker A: It'll take AI to make Quicksight usable. So that makes sense. [00:57:16] Speaker C: It makes all the sense in the world. Right. [00:57:19] Speaker D: I was like, quicksight got an update. Somebody uses Quicksight. That's where my head. [00:57:24] Speaker C: Amazon does. [00:57:26] Speaker A: Yeah. [00:57:27] Speaker C: Because Amazon doesn't want to pay Tableau. I'm pretty sure that's why they had Quicksight. [00:57:30] Speaker A: Yep. You know, with all the time that Slipsky was the. They didn't make any Quicksight announcements. Really. Not really. [00:57:37] Speaker C: Yeah. [00:57:37] Speaker A: Yeah. [00:57:38] Speaker C: Which was. I thought he was sure going to be like, oh, we're going to make major improvements. This thing. [00:57:42] Speaker A: Yeah. [00:57:42] Speaker C: This didn't happen. [00:57:43] Speaker D: Even he knows it's bad. [00:57:45] Speaker C: Yeah, he did. I think after Swami there was a partner keynote. We don't have anything really there yet other than there's a new education equity initiative which they'll be applying generative AI to educate the next wave of innovators. And that's. That's it for Wednesday. Leaving Warner's keynote for Thursday, which we had no announcements. So I think we'll talk more about Warner's keynote. More from the perspective of his thing was basically, complexity isn't always bad kind of this continuation of keep it simple, always be refactoring those things. But as you get bigger and you get more complicated things, times complexity is warranted. And I think that makes sense when you think about something the size of S3, it's basically something that you inherit over time. As long as the components are simple, the building blocks are simple, then that's good. But the overall combined set of building blocks may be very complex. [00:58:37] Speaker B: He says a lot of things that. That make a lot of sense that people often probably don't think about until somebody says them out loud. [00:58:46] Speaker A: Like, it really does feel like he's putting words to thoughts I have but couldn't express myself. Like, for the last. Especially the last few years, where his keynotes have been a little bit more introspective from a technology sort of philosophy and less, you know, focused on announcements and the cool new capabilities. And so, like, it has been a lot of that for me. And so it's like, it's really great because then I just steal a whole bunch of content from it. Like, oh, yeah, this is how I should communicate that idea. That's way better than the jumbled sentence that just came out of my head. [00:59:20] Speaker B: Yeah. Complexity is weird, though, because complexity kind of emerges from what he builds. Like, you never. You never go out to build a complex system. It's just something that naturally happens. And so I. I appreciated him calling it out and saying that it's not inherently bad, as. Unless it's something that becomes unreliable or unmanageable. [00:59:38] Speaker A: Yeah. And it's, you know, like, it's a subjective measure. It's, you know, complexity is as much as you don't understand. So it's like, this is really complex because you don't get it. You know, like, yeah, there's one guy that sees it all and sees, you know, the matrix coming together. Maybe it doesn't seem as complex, but. [00:59:54] Speaker D: I also liked how he talked about like. Or I think it was the S third guy that talked about like. Okay. As we grew, we had to, like, you know, one, change languages to one, we could hire people in, but then two, also, like, how we actually grew as a team, this complex thing. Like, how do we add people to this ridiculously complex infrastructure that is S3 in a way, without further breaking it? Because people might not know all the edge cases and all the, you know, sharp edges that exist. So how do you then even take something that is complex and use it and leverage it, which you know that part. And grow a team that can continue to develop because otherwise you're stuck with the same seven people that are the only people in the world that know how this thing works. And that's not good for your business either. [01:00:44] Speaker C: Which is basically how we're dealing with all the people quitting from Amazon because we're forcing back to work five days a week. [01:00:51] Speaker D: Who's leaving that the side? But you could bring it front and center. [01:00:55] Speaker C: That's hilarious. All right, well that's it for another fantastic reinvent in the books again feedback. What I've seen from social media is pretty positive other than people who are like us, not so big on the Q and Bedrock and they're over heavily focused on AI, which was definitely here, but there was a little bit of other stuff. I appreciated that there wasn't all AI wall to wall. So that was. Maybe they have some feedback that that would be important. So I do appreciate that. Any other final thoughts before we wrap? [01:01:27] Speaker A: Well, I hope other other cloud providers take note of that because it's everyone. It's all just been AI, everything AI and like it's cool. It's the new hotness we're putting in everything. I got you, but oh yeah, stop. [01:01:39] Speaker D: Talking about AI, the Microsoft rig, whatever they call theirs. I'm stuck on Ignite. Yeah, rig nite. I was talking, my head was like, ignite, Ignite. Well, I was thinking like reinforced. I was like, no, that's the security one. I couldn't get the right letters together. They kind of did the same thing that AWS did here, which is like heavy focus on AI. But there was a lot of like nice quality of life improvements scattered around the edges that if you kind of look through, looked around them like they were good things that got updated. Yeah, see, I said something nice about Azure. [01:02:19] Speaker C: All right, well we'll be back next week catching up on everything we missed from Google and Azure, which they have had quite a few announcements to try to counter program to the noise from Reinvent. So we got lots to catch up on next week and then Amazon's already been announcing new features for a bunch of these products that we'll talk about next week and back to the normal recording schedule. Have a great week in the cloud, guys. [01:02:40] Speaker B: See you later. [01:02:41] Speaker A: Bye everybody. [01:02:42] Speaker D: Bye everyone. [01:02:46] Speaker B: And that's all for this week in Cloud. We'd like to thank our sponsor Archera. Be sure to click the link in our show notes to learn more about their services. While you're at it, head over to our [email protected] where you can subscribe to our newsletter, join our Slack community, send us your feedback, and ask any questions you might have. Thanks for listening and we'll catch you on the next episode.