episode 332 for the week of thanksgiving 2025 reinvent predictions i i just don't know how it's possibly reinvent already like can we can we be serious about this for a minute is it really reinvent come on i don't a long year yeah it's been a long year and i also don't understand how it's already over i don't know how those two things are possible at once but here we are it does seem crazy rename it rehash instead of reinvent because I think they're running out of good ideas we'll see I don't know Vogel's keynote is always really good that is true well we're recording two episodes tonight and so this one actually is from the future in the past I don't know how that works we're technically recording 332 before 331 which is just weird to my brain breaking the fourth wall I know I know breaking the fourth wall because we all decided we want to take Thanksgiving off because I'll be gone and just a pain. And then the rest of the month gets kind of weird too with the rest of the holidays. So we're going to get this done tonight. So we're going to jump right into reinvent predictions. You'll be listening to this hopefully on black Friday, if I do everything properly, which questionable because I'm going to edit this and hopefully get show notes done by myself so I can get it all scheduled to go. So while I'm on my trip, it'll just magically post. So yeah, that'd be interesting. But before the show, for reInvent predictions, just like all of our prediction shows, we roll the dice, see who gets first. And this time, Jonathan won the roll, followed by Ryan and myself and then Matt. I am the reigning champion of reInvent. I have won most years. I don't know why it is the case, but it just seems to be how it works. Oh, oh, do you now? Yes. Because I don't know how, so. Because you read notes. Oh, oh, that does help. Yes, you're right. that's good yeah he does his homework more than three minutes before the show starts exactly that's true yeah that's true yeah hey i started 30 minutes so i feel like the third place is a good place now when i drafted for my nfl fantasy team it was a bad place to be like a number eight or number nine which is where i drafted and which is what my team was losing uh unlike ryan's but uh yeah so i mean in general i'm just glad i'm not going to reinvent i think is really the part I'm most excited about, reInvent, is not being there. Someone tried to invite me to come. Like, just come hang out. Don't go to the conference. Just come. And I was like, yeah, I could do that. And then I thought about it for more than five seconds. And I was like, yeah, no, I'm not doing that. Why would I do anything near there? And a lot of the last few reinvents I had was kind of that setup anyway, because I could barely make it to any of the sessions with all the different meetings and talking to people. I don't have enough Amazon work at the day job to get all the cool meetings that we used to do at reInvent so then I actually had to go to the sessions and then I'll just go to the session and I'll be annoyed or just be drinking that's how that'll work so my liver is still recovering from other things I think the last session I went to was I was just saying the last session I went to was with Orion when they released Image Builder like 8 years ago I think that was the last time I went to the actual session. Wow. Wow, that is a long time ago. I think I went the first year after Pandemic. I think that was officially my last re-invent because I wanted to see if it got back to the old days, and it didn't really. Well, the Prediction Show works. Actually, we used to be really good at AWS because we used to be really in the AWS ecosystem, and now we're not as much. So it's actually much harder today. We're never really good at this, but I definitely feel like it was easier. So this time I had to do a lot of thinking and research on this. So we'll see where we end up. But the way the prediction show works is we roll the dice toward the thing. I mentioned the order earlier. We will then try to make our best guesses of what we think they're going to announce. And then we'll see who wins points. You get a point for every item you get correct. And if we get none correct or we have a tie, we have a tiebreaker, which is how many times did they say the word AI on stage, which will be probably in the thousands this year. I remember it being pretty large last year as well. So we'll see. So that being said, it is reverse snake draft. So first up, Jonathan, is you, and you are now on the draft clock. All right. More of a wish list item still, but they do build things for customers. So I think my first guess is serverless GPU support, or whether it's an extension to Lambda that provides inference, serverlessly or whether it's a different service. I think it's about time we had an on-demand serverless offering that provides GPUs or some kind of AI inference. I mean, technically you could do this in EKS if you're running your own nodes, but I think that you're talking specifically far-dating. I mean, yeah, you could run Knative on top of EKS, but I think you're looking for a managed service. Yeah, I would imagine it will be an extension to Lambda, and I'm not going to commit to any of the things I'm going to say now because I don't want to limit my scope exactly. I've been burned by this before, as have we all. Yes, yes. But I would imagine there will be some requirements around it, like it has to be provisioned to be a certain size instance, and it probably has to be something that you pay to keep pre-warm so they can keep that stuff in memory. But, yeah, I think there's a huge use case for something like this. that's good I think you're coming in hot on that first one I like that one that one's even on my radar so nice job thank you alright Ryan you're up next I'm going to steal the one that I think is on your radar because I got better draft position this year and that is new Tratium 3 chip nice going for the kill I thought about that I'm totally staying away from that this year because I know how constrained manufacturing is and whether they'd have the money to spend on it. I don't know. Previous years, I really hoped for Amazon Fab Plant in the US and things to help support things like that. I think you stand a very good chance of winning without that. Yeah, that's a good one. In the future, in the past, whatever order is going to be in this conversation, there was a one for Azure. Just know. I don't know how you describe that. Can I suggest that instead of just saying training, you extend your guess to also include the inferentia chips? And potentially a Graviton chip. Well, yeah, obviously I meant that. Yeah. I mean, training is expensive, but it's a relatively fixed cost once you build a model. Well, I think inference is the ongoing cost, and so reducing that cost may be more important to them. Well, historically, we've seen inference costs. In the last four years, inference cost is probably the cheapest, has the best return on investment because it's dropping in price the most. So, yeah, I would expect more inferential trips or things in general. So, yeah, I think that's a good bet as well. And I had Graviton on my list, so I assumed someone was going to grab it with my draft placement, so I'm okay with that. Alright, that puts me on the clock I'm going to go with a new Nova Premiere or Nova Sonic launch that potentially includes multi support Yep Really being last is not going well for me It is a bad place to be, for sure. Yeah, it just makes sense that Nova... They haven't technically general availability Nova yet, so they could potentially just release it, but I assume at this point it's a year old, they're going to have to do a new version of it. and then multimodal, of course, all the new hotness and models, and so if they're serious about being in this space and they want to have their own foundational model, then it's time for you to deliver on a new Nova model, so it just makes sense to me and that leaves you, Matt, in the fourth spot I'm going to go with a model router yeah, with everyone else kind of releasing them, but also the fact that setting up your code and everything else to have a specific one or ways to test it easily. They have some in SageMaker today, don't they? SageMaker. I was going to go within Bedrock, actually, was what my note was, but I was trying to be more generic. I mean, I wouldn't tell you to go narrower unless SageMaker already has it. But I don't see any routing things. You're probably safe from this one. My five seconds of Google searching says you are safe with this prediction. All right. Jonathan, you're second. Oh, wait, I thought Snake Draft goes back to you, doesn't it? No, no. I guess it's not technically a snake. It's just an up and down. Okay. Sorry. I don't know how drafts work, Ryan. Come on. This is why I'm winning in the fantasy league. Go sports! Sports. Maybe next time we should each have a certain number of virtual dollars or credits that we have to bid on the position that we want to go from one to nine. We all barely do our homework. We barely win this competition as it is. You want to make it more complicated? It's a great distraction from never getting anything right, though. This is why we didn't get right, because we made it more complicated. Yeah. This is not going to end well, I can tell you. I expect a service which allows agents to run continuously performing actions in the real world. What? What? That's so broad. It's a thing that does something in the world. That's what I'm expecting. Continuously. You're going to have to narrow that down a bit. I will narrow that down a bit. So you build an AI agent with a goal or with a set of instructions to do something. You know, it could be checking the weather every day and turning your screen to the signing for who you want to. But I think commoditizing agents that can run when they need to or periodically or all the time and then can perform actions. So it's sort of like – So is this kind of like competitive with Gemini where it can call a restaurant and make you a reservation or go buy stocks when the stock price hits a certain price? Right. I would see it as the service that would support that type of application. Okay. So you're really an agentic platform is what you're talking about. No, because that's my next one. That's not what he's talking about. I'll narrow it down and say specifically a place or a service that runs agents that you build on a schedule that you define. I'm still going to say it's an agentic platform for agents sorry Brian yeah fine well I guess Brian that takes your number second so now this is where you scramble that's fine I actually did some stuff so I think to follow along with the bedrock enhancements I think that they are going to expand their expand the number of models to make it more challenging. I think they'll probably announce over a thousand new models because I think that it's one of the ways that they are really lagging behind GCP and Azure in their magenta gardens. Bedrock has a lot of models now. A thousand feels like a lot. I mean, I mean, I was reading a terrible article that was How many models are in AWS Bedrock? I think it's like all the ones Cohere's in there Anthropix in there Gemma 2 Quinn I mean just off the top of my head those ones I guess 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 There's 13 Or could you say that they will mention on stage the number of models they now have in Bedrock. Yeah, I was thinking the expansion, just because if you look at the Azure AI Foundry, it's got 1,500 models in the catalog. I mean, it's all the models with different partnerships. So you're going with, they'll expand the number of models available to you in Bedrock. Sure. Okay. It feels very vague. I feel like you need a number. I mean, I don't know if they'll double, I would say, at least. Again, I because I don't know if this is going to step on my next one, so we're going to debate here in a second. This is going to be like teamwork teamwork draft. Yeah. In this competition. Yeah. It's a collaborative competition. It's a co-competition. I don't know. Really, it's just Justin wins every year. We get it. So I specifically, so you guys tell me if you think Ryan already has this with number two. I specifically was going to say they were going to announce a partnership with OpenAI. No, that's different. That's different. That was one of mine. Yeah. I think that's because that's more than just offering, you know, GPG-5 in the marketplace, right? That's a strategic partnership. Yeah, it's a partnership. It's, you know, everything that OpenAI has is going to come to, you know, on stage, OpenAI is there, is what I'm thinking. okay well good then Ryan I love yours yours great he wishes you luck Matt your second draft pick you all suck I'm going to go with well architected framework review expansion so could be things like a well-architected framework for AI or for green. I guess they already have sustainability. They had an AI one too. Well, architected framework. I thought this was a pretty good one. I mean, expansion is still an expansion whether or not they already have it. That's true. They have a generative AI lens today. Yeah, I missed that answer. All right. You can just go with expansion and not limit yourself to AI, which is what I highly recommend you do I don't even know if I like that one this is because I'll go there is this just the first day's keynote or is this also include this is this is typically the Andy Jassy keynote and the Werner Vogels keynote Oh well it simply that ML and the Werner Vogels But Werner doesn announce anything Werner hasn't announced anything for a while, but I do think we typically do all covers on that. I thought we just did the main one. No, it's Tuesday, Wednesday, and Thursday keynote. We don't do the Sunday Night Live, and we don't do the partner keynote. All right. I guess I'll just do with that because I have to you've officially killed all my other ones and I had six of them and I don't know how I got all of them taken out this year it's just how we roll here we're all similar thinkers I think that's part of the challenge here I think that puts Jonathan on the clock for his third and final prediction I've got two and I don't well you can put one in honorable mention after we get through this but that's fine in that case I think Werner will announce that this is his last reinvent keynote and that he is retiring interesting I don't like that I don't like that one either I also don't I also think you're that... It's not possible you're not wrong, so I just... Totally. Yep. It's a bold one. That'd be an interesting conversation if that one turns out true. Alright, that puts... Who's up next? Oh, Ryan. Ryan, you're up for your third. Alright, this is more of a wish list in my... just a kind of a gasping effort to try not to predict something AI related, but super difficult. So my hope is that there will be a major refresh, mostly most likely UI based refresh to AWS organizations. So some examples. UI refresh or just a refresh to organizations? A refresh in general, but I suspect it will be UI. But to give specific examples, something that would provide visibility of the SCPs and how they map through OUs or visibility into stack sets across your organizations. Just things that you don't really have today. You have to configure things in organizations and then go account by account to go see if they're done. All right. I'm looking through my list here to see what I want to go with for my third, because you guys all have taken many things that I liked. I did that one. That one's undone. Oh, when we talked about two. Oh, yeah, you guys are just killing my list here. All right. Wait till you're last. Yeah. I'm going to go with enhanced features for security hub to add agentic AI capability to it advanced agentic AI capabilities for security hub I only have a better one because right now security hub doesn't really have anything agentic yeah I mean it's funny because I yeah there's to do work or is that security for agents I'm looking for automation for SOC teams. Okay. Please, automate the SOC teams. All right, Matt, that puts you in the final spot. So I feel like AWS is missing an authentication layer for customers. Cognito? So they have Cognito, but I feel like Cognito hasn't really been anything for a while. So you think they're going to try to take another bite at that apple? Okay, strong. I mean, I'm already going to have my Dio's. They have six ways to do everything else. Why not? Right. I would, so actually I think if you want to go down this path, and I'm only helping you because you're after me and I don't have another pick, but agentic identity and AI identity is a really unsolved problem in authorization. Oh, that's a good one. And if you want to go down this path of authorization identity and you want to stick to AI, I would probably do something around agentic identities. I was trying to get one outside of AI. Into the rest of the card. Bedrock already has agentic identities. Like an IAM type of persona for agentic things. Because that was going to be one of my suggestions. I'm like, they don't have this already. And they do. I want to go research that because it is a hard challenge. It's a big problem. Yeah, it was a big topic. In some recent conversation I had with StrongDM, they were like, yeah, this is where we're seeing a lot of interest. I'm going to still go with it. What are you going with? Sorry, I got distracted. Essentially, for lack of a better term, Cognito 2.0, but some sort of new authentication, like end-user authentication for a client portal. End-user authentication that doesn't suck. How's that? Yeah, that's right. Not current Cognito. I mean, I wish that because I hate Cognito. Like, every time I try to use it, I'm just like, bleh. Everyone hates Cognito. Like, I've never met a single person who's like, oh, yeah, Cognito works great for me. Yeah, I never heard that either. I mean, I normally at this point in the show, when we've done predictions, I would normally... Oh, actually, I need to do our tiebreakers. So that's reverse order. So, Matt, you get to pick first number for tiebreaker. What are we doing? How many times did they say AI or artificial intelligence on stage? In which keynote? All of them. In all three of them. Oh. One number. Yeah. That's why it was a big number last year. 200. 200. Nice round number. I like it. I'm going to go 160. Ryan? I feel like this really screws the person that goes first. So I feel like I get screwed on the first and screwed on the other side. I feel like the advantage is supposed to be the person that goes last. You should get some advantage. Well, that's why, yeah. So John's in the pick. 201. Sorry. Let's see. I think last year's number, if I recall, was like 170 something. So last year we... No, no, no. Last year I'm looking at it. How many times will AI or artificial intelligence movie set. And it's only 35 or 72 was our guesses. We must have just done one of the keynotes. I thought I picked a very high number for that and it got... I think we said we were hoping... Last year we had this optimism that they weren't going to be crazy on AI. They had heard the message that everyone was getting tired of AI. I remember that talk track and then that was completely incorrect. So... Alright, well then I'm going to go with, um, your 99. All right. 99 And John balloons One if I was looking at was left I was like yeah one is probably where I would go just to be contrarian All right. Well, normally I would tell you what to do if you're going to re-invent, but I haven't been in years, so I can't really tell you what you should do if you're going to re-invent. But I do know talk talks are still very highly respected and well-regarded. Anything you can do with this on a lab is definitely recommended for your time at re-invent. and then get out there on the floor, talk to your peers, and have a great time at reInvent. You won't find us there. You'll find us somewhere else on Twitter maybe or elsewhere tweeting about how silly what they're announcing is. But I look forward to finding out how wrong we are after Thanksgiving. You want to do honorable mentions really quick? Oh, yeah, we have honorable mentions, that's right. Honorable mentions. What do you got, Jonathan? um i think i some kind of proof mathematical proof that uh an ai or one of amazon's or amazon's llm specifically have generated a particular text some kind of mathematical thing that they do to the output that they can use to verify the source of um some output i think will be an interesting one uh i'd also you know when you talk i just feel dumb sometimes. Oh, that's... It's a gift. But if you listen a bit longer, if you listen longer, you won't think the same thing. I'd like to see a marketplace for AI work. So, you know, you can sell images or services on the Amazon marketplace. I think if I build an AI-based tool, I'd like to be able to monetize that. I think a facility where you can publish AI tools and Amazon handle the billing for you doing work. would be kind of cool. You know, for a while they kind of explored the customer build API access through API Gateway, and I don't know anyone that actually used that, but it's possible. But I think that'd be really cool, a place to put your apps that you've built that you can monetize them per call or per whatever. And I think the last thing was a new device to go with their new Nova models. I know they kind of killed the Alexa thing I'd like to think it's because they've replaced the Echo and the Alexa tech with some new tech that will be a lot smarter with some kind of local inference chip to save the money but a new device of some kind for consumers Interesting I had a wishlist item What was your wishlist item? Cost savings for networking. Cost savings for networking. Oh, yeah. None of us are stupid enough to say that, but we all want it. Yeah. Like I said, definitely not putting it in the main ones, but, you know. Ryan, do you have any honorable missions? I have kind of two, and they're very related. So I was looking for something like a sort of FinOps AI recommender for either model usage. like I have prompt optimization and a couple technologies there but not actually like if you use this model you would save money I doubt I mean it'd be awesome and the other one would be some sort of savings plans for different use cases I don't think that's coming either but you know it's a good idea I don't hate your idea that should have been one of your main ideas probably Oh, yeah? Yeah, because capacity planning around GPUs has got to be a real nightmare for them right now. And so savings plans or committed use plans would be... We'll give them something. I think they have... You can do committed use, I thought. That was the way. You basically can reserve a GPU, is what they have. Yeah. Yeah. I did have a few as well. So I expected something, some big S3 vectors announcement and integration to Bedrock. That was one I thought was pretty good. But I didn't use it because, yeah, why? If it ruins, I'm an idiot. I did, I had a FinOps service for Kubernetes, but then I was like, they're never going to do that. FinOps Kubernetes service, but I'll put it here just in case it happens. and then I had Amazon Q developer with autonomous coding agents I had that which would just be the beginning of the end because Q is so bad I also had oh so right now they have a very and the reason why I didn't go with this is because I didn't feel that it would be something they would do yet so right now they have a very specific differentiation between GPU for learning, GPU for inference, and then they have Graviton. And I was thinking maybe they have a new GPU architecture that would combine basically that into one more cost-effective new type of CPU core thing with integration of all those. And so I just didn't think it was realistic next generation. That brings it closer to the Apple... Well, it could be better with the TPUs. I sort of felt like maybe that would make sense, but... Or, I don't know. Yeah, something that supports shared memory with GPU processing in it, like a Graviton with some kind of AI accelerator built in, would be a really good suggestion. Yeah. So, again, that could happen. I just didn't feel confident enough in it. Oh, then I had the one I laughed about, but I thought it was a good idea, was an Amazon Bedrock model marketplace for revenue share for fine-tuned models. It's kind of like what... It's very similar to mine. I would say that's kind of the same thing as what I'm talking about because it would be an introduction of models on the bedrock that's specifically Marketplace. All right, well, if I get that one and honorable mentions, I'll go to you and you're a main show. I've got a feeling honorable mentions may perform better than that. That would be helpful. It very well might. You never know. It'd be true to form. Yeah, it would be. The other two I had was some sort of green dashboard, like sustainability dashboard. And then some sort of specific Aurora or what was their new D-SQL that they announced last year? Like specifically like an AI feature specific for that. Yeah, I mean, that would make sense because being able to write queries or do analysis or whatever would be good. I mean I don't feel terrible about these I mean we had joked about we have no idea but I think these are good No we don't feel terrible until after re-invent Yeah then we get none of them right and then we're all like wow that was terrible So again I just remind people we are not recording Thanksgiving week which means there will not be an episode the week during re-invent We will record an episode late that week and then try to get out the following week for recapping re-invent That'll be a dedicated episode to re-invent and then we will have our normal back to our normally scheduled programming. So lots of fun. Enjoy your time at reInvent if you're going and let us know how it goes and if you had a good time. Have a great one. Have Thanksgiving. Bye, everybody. Bye, everyone.
