336: We Were Right (Mostly), 2026: The New Prophecies

[00:00:00] Speaker A: Foreign. Where the forecast is always cloudy. We talk weekly about all things aws, GCP and Azure. [00:00:14] Speaker B: We are your hosts, Justin, Jonathan, Ryan and Matthew. [00:00:18] Speaker C: Episode 336 recorded for January 6, 2026. 2025 we were mostly right. 2026 the new prophecies. Good evening, Ryan, Jonathan and Matt. Full house tonight. [00:00:31] Speaker A: Hey guys. [00:00:34] Speaker C: We only decided to take an extra week of vacation, so we're not too far behind. And luckily the cloud providers decided to mostly take off the time too, which was really appreciative this year because I think it was last year they just bludgeoned us to death after Reinvent with Amazon stories because everything it wasn't AI. They didn't announce that reinvent. They just dumped on us for the next three weeks. But since they they did a better job this year, it was a little bit nice. But this is a. This is our annual look back at our predictions, make some new predictions, talk about our favorite shows and then I I've done a bunch of homework over my Christmas break, a bunch of work and we'll talk about a bunch of that stuff today. Then we'll hit up some stories to wrap up the day but lots to get into here. So Happy New Year. Let's get right into it. So predictions we did pretty darn well. I think this is maybe the best we've done actually looking at these starting with Ryan. He predicted that we'd have quick LLM models for individuals and according and I actually used LLMs to help me validate this data. So if you don't you disagree, you can blame Claude. Basically it came back and said it was accurate and it basically Pointed out Metalama 3.1 AB instruct the GLM model, the GWEN model chosen for a second balance of performance and computational efficiency make for an ideal for edge AI deployments which is pretty nice. AI at the edge natively for lambda or Lambda esque that came out from Akamai. We saw that as well as we saw it from Cloudflare and several others and then cloud native Security Mesh it said that was a bit unclear why they said there has been some evolution in service mesh technologies. It didn't find a breakthrough to help you on that one. So it. [00:02:10] Speaker B: I definitely wouldn't give me that point either. [00:02:12] Speaker C: I don't think so either. Matt, you did this is probably your best prediction ever. [00:02:18] Speaker D: I think it was all of our best prediction ever. [00:02:21] Speaker C: In general we did really well. But I think, you know, I was very pleased with both you and Ryan who came in strong on both of these. You Had Focus adopted by Snowflake or Databricks. That both happened. Alibaba Cloud Databricks and Grafana all officially adopted Fooocus as well as Snowflake did AI security and ethical standards. The ISO 42001 standard, you know, basically got pushed out. Companies started achieving it including my and then even some of the cloud providers like anthropic got their 42,000 win certification as well. And you predicted that Amazon authenticate five plus services and you know, our little bit of research said CodeEdit, Cloud9S3 Select Cloud Search, Simple DB Forecast, Data Pipeline, QLDB, Snowball Edge and more. Although WorkMail was not deprecated so you missed out on that one a little bit. But that's okay, it's pretty good overall. [00:03:09] Speaker D: It would take 19 pretty well. [00:03:11] Speaker C: Yeah, yeah. Jonathan predicted that company will claim that AGI is achieved and I did not think this happened but they did research and said actually that's not the true Integral AI. Apparently founded by ex Google veteran Jad Turfi claims to have built world first AGI model in December of 2025. And Sam Altman called GPT5 a significant step along the path to AGI when he released it. So I don't know who Integral AI is and so it really depends on Jonathan if he feels that's a valid choice of giving him a point or not. But hey, someone claimed it so he didn't say what company would claim it, he just somebody would claim it. [00:03:45] Speaker A: Well, I did say claims not actually achieves. That was very strategic on my part. [00:03:50] Speaker C: Strategic on your part? [00:03:51] Speaker D: Yeah. [00:03:54] Speaker A: I'll take the point because I think we're close. You also said AI, I think an agent. Yeah, I think, I think an agent working for days on end is not AGI necessarily, but I think we're getting closer and closer, especially with some of the work from like Yann Lecun on his world models and things like that. It's very interesting stuff. [00:04:11] Speaker C: Yeah, there's definitely some cool stuff going on in that area. You said AI agents will book reservations on real world tasks. That is fully accurate. OpenAI's operator can do it. I actually tried a service that was kind of a AI assistant that I used to book some stuff and yeah, I've seen it. Although a lot of them fall back on humans very quickly, you'll notice. And so they typically have some things, but yeah, it's there. So I agree. I think you got that one fully. And then models that can learn in real time. You are partially accurate there. And that comes from some of Claude's memory capabilities, some of the other things. But there's a lot more research happening in continuous learning, which is probably a 2026 prediction for somebody, I'm sure. [00:04:49] Speaker A: Yeah, I might repeat that one again this year. [00:04:54] Speaker C: For me I had GPT 5, plot 4 and Gemini 3 all getting launched. That all happened. I said OpenAI will not be seen as a leader by the end of the year, which basically Google is starting to run away with it. Anthropic holds 30% of enterprise market space versus OpenAI's 25% which is a sharper reversal from the 50% in 2023 versus 12 before that. And then Sam Allman of course issued the code red memo which is what really kind of felt pushed me over the ED that he sent a memo that they're losing market share tells me that they're concerned too. So there you go. And then I said 10 companies would go RTO 5 days after Q2. I was only partially accurate. I got a few but not a lot of new ones other than the people who already announced when this prediction happened. So I'll take partial credit but not a full point there. But yeah, overall guys we, we should have, you know, invested something. Yeah, the process. So overall I, I, you know, Claude's grading of this was A plus for Matt, A for Justin and Jonathan and a B plus for Ryan. That was its overall score. So nice. [00:05:56] Speaker A: It took us six years to know what the hell we're talking about. That's great. [00:05:58] Speaker C: Yeah, doing great. I think it's actually the, the industry. [00:06:03] Speaker B: Came to us with AI. It's made, it's become kind of predictable. [00:06:08] Speaker C: So you know, basically we had the break time and I was doing reviews and during review time I have to do things to make my brain not go scream into the other of writing reviews. Not that I hate reviews but just you know, they're mind numbing. Although AI was helpful for them to share and so I, I decided that I was going to do a bunch of enhancements to Bolts and the thing I wanted to do mostly was come up with stats from the show and so I shared with you guys back a couple weeks ago. Matt missed it, so you guys probably missed it too. But I posted basically a bunch of stats from 2025 and I will post it again in our slack room for you guys. So you have it. [00:06:46] Speaker D: Just saying it's been a long couple weeks. [00:06:48] Speaker C: Yeah, that's fine. But basically in 2025 we covered 1,308 stories from 15 different unique sources and Interestingly enough, cloud provider coverage wise, Amazon still won at 39%. Azure was 22.9%, thanks to Matt's insistence on adding more stories. And GCP was 38.1%, which I think is up in general. AWS official, GCP official, and Azure Official, of course, lead most of our sources for news, although we do get news from tech News and other different websites as well. We have a breakdown of all of our top domains where we got news from. But then the thing I was most interested in was I said, well, over the entire history of the show, what's the breakdown been? And so if you go back to 2019, when Peter, Jonathan and Ryan and I started the show, we were doing 73% of our stories on AWS versus now 39%. We were doing 16% on Azure and 11% on GCP back in 2019. So, I mean, in general, I think we've become more balanced, which is sort of a win. We are still covering almost the same amount of stories back 2019. We had the lightning round and other things. I think that juiced some of the link counts, but we did 1441 in 2019. We did 1694 in 2020. Almost 2000 stories in 2021. And then we've basically last in 2024. We decided we need to cut down on stories, and so we did. And that was noticeable in the numbers. So, hey, good job, guys. We actually cut stories now, which is good. [00:08:14] Speaker D: That's cool. [00:08:14] Speaker C: Which is good because the cloud providers continue to drive all of the stories after. So then we. I said, okay, well, about host participation. We did 49 episodes in 2025, analyzed 42 of them. We only had six shows last year that had all four of us there. So little bit of a struggle. [00:08:35] Speaker A: Been m. Trudent for the. For the year, I think. [00:08:37] Speaker D: Yeah, yeah. [00:08:38] Speaker C: But I. I apparently made 95% of the shows. Ryan made 85% of the shows, and Matt made 78% shows. I was surprised about Matt and Ryan's numbers because I feel like I'm always chasing both of them, and it's either I either get one or the other. So I was. It was higher than I expected there. [00:08:54] Speaker D: I had a kid. Just say it. I had a kid. [00:08:56] Speaker C: That's fair. I know. And I am a child. Jonathan was in 12 episodes, and that was due to some personal stuff going on. But we did have one guest, which we. We do need to work on guests again. We only had lease this last year on episode 329. [00:09:11] Speaker A: That was fun. [00:09:12] Speaker C: So, yeah, Overall, you know, Jonathan has some work to do 2026. But you know, overall I think it's, you know, it's good. Take time off. Don't. You don't have to be at every show. Then I wanted to look at AI mentions an AI ML takeover. So AI was mentioned to 596 times in 2025, averaging 12.2 times per episode or 66.4% year over year growth. So we are talking about more stories than AI for sure. And you can see it in the chart pretty clearly. 2019-2020-2021-2022 before AI was big express ML stories. Then 2023. Of course it started to get bad jumping from, you know, less than a mentioned episode to 4 and a half times an episode, 7 and a half, 7.3 times an episode to 12.2 times an episode. So that's a little crazy. We talked about 19 outages in 2025. We talked about fiber and network infrastructure 12 times and we had 1, 2, 3, 4, 5 major announcements for Deepsea cables because we do love those. So there you go. We did five prediction shows. I was the 2025 champion. So thank you very much. [00:10:20] Speaker B: Nice. [00:10:20] Speaker C: I won re invent 2025, Matt won Google Next 2025 build came out to Ryan winning on the tiebreaker and then I won Ignite and then our annual predictions. Matt just won. So we just talked about it. So congratulations to Matt. [00:10:36] Speaker D: So did I tie you? [00:10:37] Speaker C: I think maybe so. But overall for all time I am a leading six two to zero to one. So in general there is a bit of a bend here but we'll get Ryan going there this year hopefully. Okay. Running jokes and callbacks. Apparently we use finally a lot in our show titles. 18 times we had a Finally Oracle arriving late to the party is a common joke. We do rack Ryan or the life size 8 years rack given to Ryan. We finally mentioned that a few times over the years. Easy is exec my prediction for 2018 finally coming true was a big one and then major cloud outages breaking the Internets buzzword bingo over 2025 the hottest terms MCP which yeah totally agree reasoning agentic multimodal agent to agent grounding and vibe coding and then out of the 49 episodes we did 58.9 hours of runtime. Our average length is 72 minutes and our longest episode was 1 hour 51 minutes. [00:11:36] Speaker B: Wow. [00:11:36] Speaker C: Episode 310 and our shortest was 31 minutes which was our reinvent prediction draft show which was a special episode. I do had to have an in memoriam service. We lost many services in 2025 including AWS Cloud Search, S3 Glacier, AWS Migration Hub, AWS Application Discovery Service, S3 Object, Lambda Health, Omnics Variant Store, Kubernetes, Nginx Ingress Controller, Azure Consumption APIs, Azure AWS Connector, AOL Dial Up Internet and RC4 encryption. And now for a moment of silence. Moving right along, Cloudflare bought replicate Atlasian bot, DX cognition bot, Winsurf and Meta bot scale AI plus another one we're talking about today. And iRobot and builder AI went out of business. One over fraud and one over not innovating I guess. I don't know. And then cloud expansions. We had Azure and Google Cloud expansions in Atlanta, US Gov, Arizona, Turkey and Thailand. And then we had some aging poorly items again. Builder AI and my robot on top of that code commit deprecation because that got rolled back and then GitHub would maintain a pen center. Microsoft that also has not aged well as Microsoft has now pulled it fully into the fold. Unfortunately most mentioned non hyperscaler companies, OpenAI of course Nvidia, Anthropic, Snowflake, Databricks, Cloudflare and then we had a bunch of titles that we almost went with 633 rejected titles partially because AI helped us with titles this year. But you know there was Several that the AI thought were the best of 2025 including from Roomba to Toomba which was quite good. Not so fast. It always helps you find gateways doing absolutely nothing. Fast and furious it was lays 320 terabits of cable across Atlantic and many more. And we'll include these in the show notes so you can see all of them. But yeah, a little bit of stats. Guys, there you go. And there are a couple things I learned about doing this project. LLMs don't really help you in most of this. This is a lot of ML Python code. So getting like analysis on something, the LL really helps you out a lot. But anything else is really just counting metrics. And to do this, you know you need transcripts and show notes and all the kinds of things. And so as part of building this all out, I built a brand new ML data pipeline for the cloud pod. So as we publish our shows, we're now getting official transcripts, we're getting all uploaded to the website and that's all going to go into our new Bolt based AI search. So you can now talk to Bolt so you can chat with him just like any other LLM. He will remember your conversation and he will use our vector search off of S3 to identify where we talked about things, so you can ask him all kinds of questions. He's available to you, of course, and our Slack team. So if you have not joined our Slack and you want to come ask questions about how many times Jonathan says he loves AI or GPUs, you can go ask those questions of the Bolt and he will tell you, which is kind of fun. And so he's out there and we'll be keeping that up to date all year long. So that a. When I do this next year, I don't have to do as much work. But number two, that we continue to build our repository. So when we have these questions in the show, like, hey, what did we talk about last week or a month ago, or did we talk about this before? We can just ask Bolt and he will tell us very kindly, which is really nice. So I can talk about that. If you guys are curious about building an AI chatbot and LLMs with vectors, I learned a lot. I learned what it's good at, and I want to learn what it's not good at, as well as how some of the esyncrasies of doing that kind of stuff is quite interesting. [00:14:51] Speaker A: I'm more curious how you remain focused on a single project, because I see you doing some updates, I see you doing the bug fixes and things. Like, if it was me, I was like, I get it 90% working. And then it's like, nah, whatever, move on to the next thing. [00:15:04] Speaker C: Well, so there's a couple things in my vibe coding that I do a lot, right. I have tests, I always have tests, and I always have a very clear direction of what I want it to do. And then I just let it kind of do its thing and then I go review it when it's done and I make sure it works. I also have the superpower that Ryan's always talked about, is that I can keep a lot of things in the air at any given time. So I have five or six Claude Windows going at any given time with different projects or different things I have ideas of, and I just kind of bounce around them all the time. So that's kind of just my superpower at work. But really, I do write very clear AI instructions, very clear tests, and everything I write has tests. So basically I can not break myself as I go through the process. Cool. [00:15:46] Speaker B: I'm fascinated by the sort of like. Because I'm having a similar experience mostly in my day job, which is like, trying to use AI for Different workloads and then falling back into more traditional technologies or different ways. And at first I thought it was like, you know, just like old dog, new tricks, trying to like just falling back into comfort zone. But I find more and more I'm identifying things that, you know, the, the large language models just are not good at. Right. And I think a lot of stats and the metrics and it feels like it should be able to do that, right, because it's conversational and it's, you know, you're building a corpus of data for it to, you know, for the model to query and do all that, but that it really can't. Right. And so fortunately we do have, you know, machine learning technologies and the ability to do notebooks and stuff and Agentic can absolutely help you make the notebook, but it can't actually do the analysis for you, which I find funny. [00:16:45] Speaker C: Well, and it's fun because you, I've always said at ML problems, it's always, you know, the challenge is not the technology. It's always like, what question are you trying to answer with the data that you have? And that actually is even true inside AI. It's, it's the same problem, but worse in many ways because it's like, you know, like for example in the 2020 review at the end, it gives you analysis of why it thinks those topics titles are the best ones. And it gives you a three thing. And that makes sense for AI to do that because it can use natural language and it can use SEO and it can use things it can compare to and it can come up with good answers. But to like, figure out like how many episodes was Matt at this year is not a AI problem. It is purely a, you know, ML problem. Like, oh, I need to know a data source of where how do I know Matt was in an episode or not an episode? How do I know who was there? And so, you know, luckily Heather does a really great job always typically mentioning in the intro to our show notes who was on the show that day. So because of that, I was able to get pretty good data on that. And the reason why there's six episodes that don't have that data is because it wasn't mentioned in those, which is fine. So to fix that in the show notes, I now have added who was here today so I can track it. And then when we do, when it does the processing for the show, do the YouTube updates and do all our things I do for the show posting, it'll now update the vectors on Amazon S3 to do that which going through S3 vectors was actually a lot of interesting fun because number one, it's not very well documented in the APIs. No way terraform or anything. Yeah, I know, Shocker. And so like, you know, Claude, want to go back to like, well, no, no, you need to use Knowledge Base. And I was like, okay, well I'll try Knowledge Base. I tried to set that up for 25 minutes and I was like, this is too slow, I can't do this. And, and the worst thing about Knowledge Base is it doesn't have a terraform module, has no ability to do anything terraform. And when you're in the website setting it up, you have to stay on the web page for it, for it to actually complete. And if you move away, it stops. I'm like, that's just silly. [00:18:32] Speaker B: That's awful. [00:18:33] Speaker C: So, you know, there's, it's. But it's interesting because it's like, oh, a lot of this is, just becomes very, very clearly, you know, the, the tools that you've had for a time, which is, you know, kind of some of my thinking about predictions and the AI bubble and what we're talking about. I'm like, yeah, it's not, it's not AGI. Maybe AGI changes things even further. But right now it has a limited capability of what it can and can't do. It can definitely do a lot of cool stuff and it has a lot of value in a lot of different places. But it is not like these companies that have laid people off in favor of, you know, AI replacing their jobs. I'm not sure that's going to pan out long term. I don't think so. [00:19:08] Speaker B: I think they're going to get burned and have to, you know, I think a lot of companies will be able to pivot. So I don't, you know, I think it's the, the people that are hurt are the ones in the short term and then the market, the hiring market will get good again when people start. [00:19:20] Speaker C: To realize their mistakes. [00:19:21] Speaker B: But it's not a replacement, it's a tool. It's, it's an augment. [00:19:27] Speaker D: It's a tool. And companies are also just doing it to shed some expenses too, using as an excuse. [00:19:34] Speaker A: I think it's gonna be painful in a few years time because, you know, AI is taking away the level one jobs, you know, the tier one coders and people aren't going to go into industry because they think, well, what's the point? I'm not going to get a job. AI is already doing This, a senior engineer can now do the, the job of, you know, five or six different people. Like Justin with his five different. Claude. Windows open. I think in a, I think in a, in a few years time, maybe, maybe five years time, it's going to be a bit of a shit show because we're not going to have the people to replace the people who are now doing the work. I mean, who knows, maybe we won't need them at that point anyway. Maybe, maybe Skynet will have taken over anyway. [00:20:11] Speaker D: I think you're further out than that because you have enough people that are still coming in and doing it. And you'll see people, I think mainly skip that Tier one and try to attempt to go directly into it. But you're not going to lose people that fast. Like, you know, you're not having a 64 year old on average be a senior software developer and leveraging these tools just on average here, you know, so you're not gonna see that. [00:20:37] Speaker B: I wonder if it goes the other way. If you think about like new college grads and hiring, you know, new grants for interns, right? Like they come in, they're super green, you got to teach them everything anyway. And I wonder if the tier one just has a completely different skill set. [00:20:53] Speaker C: Right? [00:20:54] Speaker B: Like it's, it will be vibe coding. [00:20:57] Speaker D: Yeah, but, right, that's what it's going to be. Your tier one's going to just be straight Vibe coding. I have no idea how to do this. How to do the initial part. [00:21:06] Speaker C: Yeah, yeah. [00:21:07] Speaker A: I think to, to be a good vibe coder, you need to be an experienced programmer or you need to have business experience. I don't think the people who are vibe coding right now are getting very good results if they don't have that kind of background. [00:21:18] Speaker C: If you, if you don't know the basics of what you're trying to do, then it's going to give you bad code and it's not going to work. And you know, and then like there's things about architecture and making sure you understand how things are supposed to work. And like, especially if it gets into loops, like a lot of times my, my feedback to the agent is like, why are you doing that? That's a silly way to do this. Why don't you think about doing it this way or that way? And it's like, oh, I hadn't thought about know. And so I'm giving a guidance on my experience and what I've done in the past. It's interesting because even my uses of Claude over the year has Evolved. I think at the beginning of the year, I was very big into root code and using some of those tools, moved into Claude, really like Claude. Now I'm starting more into agents and letting agents just kind of go do things based off of, you know, a master, you know, context that I'm managing and saying, okay, now is agent, go do this thing. Come back when you have it done. With my tests passing, and then it comes back and then I can review that code as a pull request. So there's, there's definitely evolutions to it. And I just like, was just reading through the guy who actually wrote Claude for Anthropic. He just had a Twitter feed where he was talking about how he uses it and he uses a different way than I use it. I'm like, oh, some things he's doing actually sound kind of cool. And I want to learn more about those. So I've got some homework to do. I started. I finally started playing with Warp. So the other, other project I did, I don't know if you guys have noticed, the Cloud Pod website has been updated. [00:22:31] Speaker A: Yep. Looks great. [00:22:32] Speaker B: It does. [00:22:33] Speaker C: That was a warp. And I spending some time like, I never built a. I hate PHP, but number two, I never built a WordPress theme from scratch. And I was looking at buying a new template because I was like, it's kind of getting dated. We need something new. And I was. So I just basically went to Warp and I was like, hey, I'm looking to build a really cool podcast website Update with a WordPress theme. And here's some examples of ones I've seen on the Internet. I like and create me something. And so when we designed it together and basically we built out what the Cloud Hub website redesign looks like. So that's pretty cool. And, you know, definitely a lot of fun things to do. And, you know, one of the things I was really interested in doing was making it really good for mobile. And so one of the things for mobile is now it's. It actually responds really well on mobile devices where the old theme did not. And then actually I also revamped some of the newsletter pipeline as well. So we've always traditionally just sent out our email of the week's show notes to our email list. And, you know, we have a few hundred subscribers on. On the mail. It's not a big audience, but I was like, well, it'd be kind of nice if we added some more personal flavor to it. So now Bolt's going to ping you guys in the room and say, hey, last week's episode just got posted. Any thoughts about last week's episode? And you guys can talk to it and it'll write a little blog post and put that on top of our newsletter as well. So, you know, you don't have to do it. But it's something if you want to. [00:23:47] Speaker D: Be more interactive somehow we all got more work to do. [00:23:50] Speaker C: Yeah, that's how it works to somebody else. [00:23:53] Speaker A: AI. [00:23:53] Speaker C: Yeah. [00:23:53] Speaker B: Yeah. [00:23:54] Speaker C: Well, I, I don't have an AI write. I'm just having him bug you to say, hey, do you have any thoughts about last week's episode? And you know, you can say whatever or like, yeah, I really like, you know, I started really thinking about what Ryan talked about with container security and you know, whatever, but because they're always like one of the things I always read the show notes. I always read, listen to the episode. I always have thoughts after the fact. So I was like, well, that be kind of fun. And since there's enough time delay between when we publish versus when the newsletter goes out, there's some opportunity there. So it won't happen every time. But you can definitely check out when one of us posts something about it. So. All right. That's a lot. That's all my, my vive coding of Christmas break. [00:24:29] Speaker B: Yeah, that's very cool. That's a. I mean that's amazingly produced a lot. Yeah. I think I cleaned a room and installed. [00:24:37] Speaker C: You were doing remodeling. I was doing Homestead. Yeah, that's a whole different thing. Definitely cannot get that done. [00:24:42] Speaker B: I did not know what a computer was for a few weeks, which was kind of great. [00:24:46] Speaker C: Which is kind of nice too. Yeah. All right, well, the other homework I gave you was I gave you our 500 page archive of every story we covered in 2025. And the homework for you all was to come up with your favorite episodes. And some of you did the homework and some of you didn't. So some of you are going to be frantically looking through the archive as we talk. So we're not going to do this in any order in particular. I'm just going to jump in with them. And so the first one that I really enjoyed was Amazon saying FU to Microsoft over their security in episode 287. Basically, Amazon refused Microsoft 365 deployment because of lax cybersecurity, which was just fun for Amazon to poke Microsoft right in the face like, your security sucks. I thought that was, that was a fun story. I remember us talking about it on the show. It was a good time talking about. [00:25:32] Speaker A: It as well, yeah, that was interesting. That was great. I wonder if they ever followed up on that. Did Microsoft ever fix their problems? Did they go back on the plan to implement it? [00:25:43] Speaker C: You know I, I tried to do a little bit of research on this and I meant to do more and I failed to do so. But yeah, I thought there was a follow up article at some point that we almost talked about on the show and then I didn't. But yeah, I'll do some homework on that because I. It might have happened or might not have. [00:26:00] Speaker D: I think one of my favorite stories was around a tool that I dislike and I've always disliked but Microsoft or AWS killing Chime so chime was dead. They said we're done, we're out and I was very happy. Granted I don't have to deal with Microsoft support or sorry, AWS support that often or my AWS tab where I had to be on those fun filled tools. But I've had a lot of pain with those tools over the years. I was very happy to see it was dead. [00:26:33] Speaker A: That's cool. I think my, my favorite is actually a reinvent thing which is the frontier agents from aws. It's such a shift in the way people think about using agents. We're not just promising them individually to go do a particular thing. It's literally goal based direction. And the Q agent which is the developer and the security agent. I can't remember the third one. I think it's a DevOps type type agent SRA. I just think that's, it's a great evolution of the technology. I'm really excited to see, to see how that carries on developing. [00:27:12] Speaker B: Well, that's a nice segue into mine because what's underpinning of all of that is the agent to agent protocol that Google announced and that's my favorite for the year just because I was really happy to see a standardization come out that were across everything and wasn't, you know, specific to a single ecosystem or cloud scalar or you know, all of those things. And I really do feel like it's accelerated agent to agent workflows and I think that the, the agent agent workflows has finally made a agentic AI kind of useful for me like where it's, it's more than just sort of a novelty act and you know I there was trying to decide whether it was agent to agent or mcp but I think agent to agent stands out just a little bit more for me. [00:28:01] Speaker D: Pretty cool. [00:28:01] Speaker A: I think in general though all those, the fact that this is Such new technology. And if we think back, you know, even 10, 15, 20 years ago, technologies would, you know, competing technologies would have completely different implementations, they wouldn't talk to each other. And I think what's really interesting about the way AI has been going the past couple of years is we're getting these standards which competitive businesses are adopting jointly because they know that the, you know, to get support from the ecosystem and to get user adoption, they need to be able to support the same thing. So we have OpenAI being the standard for the API, although anthropic's got a slight variation on that. Then we got mcp, we've got the tool calling standards now we've got the agent to agent stuff. I think it's really interesting that they're converging on a common way of speaking to each other while working on their own proprietary implementations. The models themselves. [00:28:53] Speaker B: Yeah, it's more akin to the development that was going on in the 70s, early Internet days and continuing into the 90s. It's crazy. [00:29:02] Speaker A: It seems very collaborative, even though they're obviously competitors and somebody's going to win and somebody's going to lose, but they still seem to be a lot of collaboration and there's a lot of public research still going on, which surprises me. I think there's a lot of research going on which is not public, especially around AGI and around how to do online learning and things like that. Yeah, exciting. [00:29:26] Speaker C: Another one that I, I did enjoy Amazon coming out with Nova, even though it's a trash model that no one uses. I like, I like the underdog story in general and I, I appreciate that they're trying and I'm curious to see if it gets anywhere this year at all. But yeah, they did announce some updates at Re Invent. I have not had a chance to go benchmark it myself, but yeah, it's a. It's a sort of fun little piece that I, I'm curious to see how it goes. [00:29:50] Speaker B: It's so underused. I don't know if the model sucks. Like, no one complains about it. [00:29:54] Speaker C: I sure haven't used it. [00:29:55] Speaker D: Yeah, it's just not used. That's why. Yeah, they've launched it. There's like one GPU sitting in USC's one. That's it. That's all they need for the whole thing. [00:30:02] Speaker B: Yeah, it could be amazing. We never know. [00:30:05] Speaker C: Yeah, it could be the best. [00:30:07] Speaker A: Yeah, it's. It's hard to want to use it in a way, because once. Now we've got Claude and OpenAI chat, GPT for, for cloud based models and then we've got GPT OSS, even like 120 billion parameter model you can download for free or, or a 30 billion model from Quen or Deep Seek. It's really hard to, for, for somebody like Amazon to come along and say, hey, let's use this thing. It's got no kind of no selling point. It's just, okay, look, we've got a model too, but there's been no effort seemingly into what does it do differently? Like why is it better? Why should I touch that instead of something else? At this point I think they've, they've kind of failed a little bit there. They need to, if they were able to use it, they're going to have to make it worth using. But my other favorite announcement from last year really was, was deep seat coming along and things up. I mean Nvidia stock sliding tremendously. That was just after Release of chat GPT 5 maybe or something. Is it 5? [00:31:06] Speaker D: 4? [00:31:06] Speaker A: I don't know which one it was. [00:31:07] Speaker C: I think it was four. It was four. Yeah. [00:31:10] Speaker A: The thing came out when the deep seat just appeared on the scene and like, hey, we've got this thing. It's just fantastic. It's just wonderful that so many people could be working on this in the background and make these releases just to shake up a whole industry. I do like to see the world burn sometimes. [00:31:29] Speaker B: We are a bunch of cynics. [00:31:31] Speaker C: Yeah. I mean as long as you shorted the stock of Nvidia, we're fine. [00:31:36] Speaker A: Yeah. Looking back over the predictions, you know the reason me and Ryan are not always getting the point is because we kind of secretly hoping for the doom. [00:31:43] Speaker B: Scenario sometimes I think, yeah, it's definitely true. [00:31:46] Speaker D: That's part of one of my predictions for this year. I also want to go with a sad story, but it affects, I would assume all of us traumatically, which is GitHub is going to start prioritizing the migration to Azure and the upcoming outages that will all ensue because of Azure. [00:32:06] Speaker C: Outages. [00:32:07] Speaker D: Yeah. So it looks like it was about episode 317. It wasn't that long ago but I'm like waiting for all the outages to start to happen on GitHub and there I have a lot of minor ones of actions or small things but I'm waiting for the major one of our door crashing and all of GitHub is down for days and just watching the whole world freak out because nothing works anymore. Speaking of being cynical. Sorry. [00:32:35] Speaker C: I mean it's already had like Two outages in December or November that I were pretty impactful to the day job. So, yeah, yeah, I'm not looking forward to that actually. So. [00:32:43] Speaker B: No, and it's going to continue. I, I don't know if we're going to get the big one or at least, you know, I hope not. But I do think that the little ones, this, like, kind of minor interruption ones, like, I think we're going to see that over the next two years. It's going to be ridiculous. [00:32:58] Speaker D: Yeah, I think you're going to see a lot of small outages. I don't think you're going to get the big ones because hopefully they've built, especially at the scale that they're at, enough, you know, redundancy and, you know, decouple their services that you're not going to see that massive cascading outage. [00:33:15] Speaker C: Hopefully they haven't done that, but we hope they did. [00:33:18] Speaker B: Yeah, we'll see. [00:33:19] Speaker C: Luckily, GitHub is relative or Git Internal is relatively distributed, so. So maybe they have some better choices for Dr. Well, I, I actually, this story got me thinking because Jonathan decided to buy a video card that can run AI models locally. And then I've been kind of on a mission now, like, I kind of want to do it myself because I'm spending a lot of money on anthropic tokens. And so I was looking through the thing, I was like, oh, yeah, the digits computer from Nvidia, which they did actually ship and they actually rebranded it to something, something else. But I'm wondering this week at CES if they're going to have a new version of this coming out with maybe a newer processor. But yeah, this ended up becoming the DGX Sparks, which has a GB10 Grace Blackwell superchip with a pedal flap of level performance, all available to you. And you can connect two of these together on your desktop. They do cost a pretty penny. So I did not buy one for Christmas for myself, although I think about it and I have been picking Jonathan's brain about what he bought, so I might follow his suit. But, you know, the whole idea of like, well, these AI models are getting good enough that I might be able to get one at home that I don't have to pay quite as expensive for might be really cool. So definitely Jonathan's inspired me. We'll see what happens next year or later. [00:34:26] Speaker A: This year, to be fair, I still pay for my Claude Max subscription on top of what I have locally as well, because they are totally differently. [00:34:37] Speaker D: I feel like you build a router that all your stuff goes through that like what needs to go to Claude and like you just fan out that way. Like some will hit local, some will hit Claude and you know, be cost optimized. [00:34:48] Speaker C: I was so I have my Claude subscription and then I also use Bedrock Claude because I got tired of paying by the drip. So I just part of my monthly Amazon bill now which can get expensive. You're not careful. So do keep an eye on it. Budget alerts are important but I I am sort of noticing there is some deprecation happening like the Claude has a new cloud code, has a new Chrome plugin that can control Chrome which is really cool but that only works with the official anthropic APIs right now. And so I'm hoping that's because it's beta and when it comes GA they'll add it to Bedrock so I can do it without having to continue to max out my Max subscription to Claude as well. But because I also have that as well. So yeah and then of course because I talk about Claude, Cloud 4 which came out in September has been amazing. I mean Cloud Sonnet 3.5was great. Cloud 4 and then Cloud 4.5sonnet are just so good at opus. Like they're. And the opus price drop was such a huge thing. I mean I went from like oh, I don't want to use Opus unless I have a really tough problem to like I don't mind using it as much. I mean it's still expensive so I don't use it too too much. But I, I don't cringe when I have to go use it now it just stings a little instead of like ah, that really hurts. So I do really appreciate that. [00:36:06] Speaker A: I think finally adding conversation compaction really helps as well because now you can have much longer conversations in the web UI with both silent and opus. [00:36:16] Speaker D: Yeah. [00:36:18] Speaker C: And it does a pretty good job of the auto compaction too. So it keeps most of the primary important parts of contacts available to you. But I still recommend writing your contacts out to like your to do lists to file. Yeah because you will sometimes lose especially in a really long running conversation. Which is one of the things I I learned as I vibe coded more was you don't want really long conversations, you want short conversations, which is why you have some. I have so many windows and I just, I'm resetting and clearing context whenever I change thought processes or tasks to make sure I don't have that risk. [00:36:49] Speaker D: See, my problem is I get distracted because I have a really good idea and I forget to type slash clear and then I get mad, but I'm like, oh, how am I running out of space? This is the 16th topic I've thought of. Oh, I forgot to clear it. God damn it. [00:37:03] Speaker A: It's a make for code. Yeah, yeah. So if you do, if you're going to do a compacted conversation, always give it instructions. I know you don't have to give it instructions, but if you give it instructions to tell it what to remember and what's not important, it makes so much difference. [00:37:22] Speaker C: Yeah. Learning how to use cloud code or sorry, cloud MD and making sure it has those things because it uses it quite often. And then also like repetitive things that you do often you can turn into commands. So like doing a git commit and pr, now I just have a cloud command I run for that. I don't even ask it anymore. I just type slash get PR and it just creates it for me. So there's like things like that you can do. So like little snippets of code that you use often with Claude, you can do that. Or like the linting process that I do, if I don't put it into the pre commit hook, I'll just have a, you know, a Claude command I just run. And then when you get into agents, it's all about, you know, having very clear contexts for each agent, what you want it to do and giving it source. You know, like, here's reference material you might need, here's what I want you to do, here's the task, and here's the test that I want you to accomplish. [00:38:06] Speaker D: I think you're like a week or two, a couple weeks ahead of me. I didn't get that long period of time during this holiday to go work out all these things. I've like dabbled in ages, I've dabbled in these things, but I haven't put all the pieces together in my head yet. So you and I will have a long conversation. [00:38:19] Speaker B: Justin. [00:38:20] Speaker C: Yes, we should do that, have an after show about it one day actually. Probably because people are listening. [00:38:26] Speaker B: It might be a dedicated show because I think a lot of we have a lot. [00:38:30] Speaker D: Right. [00:38:30] Speaker B: Because it's, we all have different workflows and you know, like I'm, you know. [00:38:35] Speaker C: I think, and I think this is. [00:38:37] Speaker B: True about all of us. I'm a little proud of mine. Right. [00:38:39] Speaker C: Like, I've put a lot of work. [00:38:40] Speaker B: Into how I set up my environment and how I do things. [00:38:43] Speaker C: And like, I like the output a whole lot. It's kind of cool. I think I'VE realized that I've, you know, there's a couple things I really like about the whole anthrop, you know, LLM coding assisted world that we're in now. It's like it doesn't have to be one way. Like where you're in old patterns. Like you always had to be like, I had to fit into this pattern. [00:38:59] Speaker B: Or I don't work. [00:39:00] Speaker C: And at least now you can be a little more flexible. But also because there's so much innovation happening and there's so many good ideas and plugins. Like, I was just reading about Wiggum Ralph Wiggum extension for Claude, which basically is, you know, it's. I don't have. I can't really describe it yet because I'm still working on what it does exactly. But I basically, it assumes it's in danger. And so it continuously runs the same prompt against itself multiple times, basically. And then it then compares the answer across multiple of the same prompt and then, you know, combine the answer into one outcome. Which is kind of cool and. But like there's a really good use cases for that. Or, you know, like today I was. [00:39:35] Speaker D: A security agent does that's what the anthropic security one. It like spawns like, I think it's five agents run the scans and gives you like the 80 percentile of all them. [00:39:44] Speaker C: Yeah, exactly. Or they're like even, you know, hey, I did a bunch of vibe coding and now I say. Now I channel my inner Jonathan Baker and I say, you're a really anal principal engineer who hates everything I do. What don't you like about this architecture? What would you change? And it gives me a bunch of feedback and then I basically do it again. And then I make. Then I use Ryan's personality to like, okay, now I'm just an asshole. And what don't you like about it? And between the two of them, I get really good feedback on my code and it's great. And it's just like talking to both of you sometimes. [00:40:16] Speaker D: It's actually the agents called Ryan and Jonathan on his local club. [00:40:22] Speaker B: It's in one of our other chats. Like, you've got a pretty good representation of me. [00:40:27] Speaker C: And so put them in. [00:40:28] Speaker B: In AI form. [00:40:29] Speaker C: It's pretty bad of like. Oh, yeah, yeah, checks out. Yeah, you got others, Matt? I had a slew of like, just. [00:40:35] Speaker D: Quality of life things. I feel like there was a lot of little ones. [00:40:38] Speaker C: There was a lot of good quality of life. [00:40:39] Speaker D: This year, like both Azure and AWS released regional nets. Granted, like for aws, it's still one per thing but it's just less to manage because it's always been a pain in the butt to kind of manage each individual thing. ECs actually increase the exit code so you actually don't have a real exit code versus here's 255 characters but the rail parts at the end. God forbid to use Java where it's just never in the exit code. I also had things along the lines of Beanstalk because for some reason they still exist. Can directly retrieve from secret manager which was just a nice quality of life versus environment variable. Look up your environment variable, shove it in you know to figure out which secret manager secret is et cetera, et cetera. Blue green actually built into ecs. You want to set up everything all the way to CodeDeploy which I thought was a nice quality because for my home little project. So it's nice to have that versus just normally what I do is I just tell it to kill the container and replace it because no one actually hits anything I make. [00:41:44] Speaker B: And no one. No one uses ECS variants task versions. No, because they're so. They're so hard to work with. [00:41:51] Speaker D: You just override them. The other one for GCP which I think we've all hated ourselves at one point in our day jobs choose the cloud is migration of buckets from one region to another because I've definitely all have been there where somebody launches something large in one region and you have to move it versus setting up another one. And God forbid somebody did something completely illogical like hard code the bucket name into an application that you can't change without full application deployment. Definitely have never dealt with that one before. [00:42:21] Speaker B: I haven't in a while. [00:42:23] Speaker D: Yeah, that's because you have security now Ryan. [00:42:26] Speaker C: It's correct. I gotta, I gotta go back here. So you don't use task versions in ecs? [00:42:31] Speaker B: No. I mean for actual versioning. [00:42:34] Speaker C: No. [00:42:34] Speaker B: I mean you have to use a task version. But I don't have like. [00:42:37] Speaker C: I mean I, I, I use it in my. I have that's why I've ciudad to take care of it because I don't want to think about it. So it just automatically it downloads last task does whatever update to the container version I need. You know container version I'm updating to and then it pushes a new version of it and it deletes the last anything older than two prior versions of my task definition. I don't think about it beyond that. [00:42:57] Speaker B: But. [00:42:57] Speaker C: But do you roll out? [00:42:58] Speaker D: I'm more impressed you clean up yeah. [00:43:00] Speaker B: The, it's the, the ability to sort of roll out versioning like it's you. [00:43:04] Speaker C: You, you can, I mean it's clunky because I feel like the Docker container version, it's already sort of solving most of the need so the task version feels redundant to me. That's why I just build automation. Don't think about it anymore. [00:43:15] Speaker D: But it's useful if you have environment variable changes and stuff like that that you need to roll out kind of coordinated. [00:43:22] Speaker C: I mean again I'm doing most of that through parameters now and so it just doesn't matter to me that much. [00:43:27] Speaker B: Well, I mean that's basically what we're. [00:43:28] Speaker C: Saying is that we don't really use. [00:43:29] Speaker B: The functions of task versioning. [00:43:31] Speaker C: We just, I mean if I'm forced to use it. So that's how I use a task version. Okay, yeah, if I could just overwrite the existing task version I would just do that but it won't let me do that. There's hundreds of versions out there really. [00:43:42] Speaker D: Prove I did the homework. I had a fun list of announcements that were just released. The Terraform plugin for Domino's Pizza provider. [00:43:49] Speaker C: Oh yeah, that's fun. [00:43:50] Speaker D: Yeah. Terraform provider for Oracle database at Google Cloud. Because when you want to burn money that's a great way to do it. Microsoft has both announced the Solve Quantum computing chips and then everyone said no, you didn't improve it. And they said well eh, we didn't really but we're close. [00:44:07] Speaker C: And they kind of disappeared on that whole thing. There was a lot of noise then nothing, nothing. [00:44:12] Speaker B: Crickets. [00:44:12] Speaker D: I was. So I started at the bottom, I scrolled up. It was like replaying the whole year in my head of like all the stories. It was kind of fun doing it. Microsoft surprises Ms. DOS fans with remaking the ancient text editor that actually works on Linux. Kind of fun. The quantity of undersea cables which you've already identified. I think there was like six or seven that sending petabytes and petabytes and I always just like the Prime Day announcements. That's the nerd in me hearing the number of SES and SQS and SNS messages and EBs, you know, volumes that they've kit that they've used and how many, you know, dolphins they killed that year during Prime Day announcements due to global warming. You know it's always fun to kind of hear those announcements. [00:44:56] Speaker A: Yeah, the trillions of DynamoDB calls in a day, that kind of thing. Yeah that's, that's always, that's always awesome to Hear. Cool. Well, thank you for doing the homework, Matt. We late to speak extra long just to get these out. [00:45:09] Speaker D: You're welcome. I was sitting there with not able to leave my daughter's room for many hours one night and I just scrolled through it and copy and pasted. [00:45:18] Speaker C: Good for you. [00:45:18] Speaker D: That's great. Let me tell you, scrolling through 500 pages of a Google Doc on your phone, one, Google Docs isn't loaded very well and two, scrolling through it takes a long time. [00:45:29] Speaker A: Yeah, Google to Godocs does not do well with large documents like that. [00:45:33] Speaker D: No, no, it does not. [00:45:37] Speaker C: And that's why we give a new document every year for archive because every one of them is that long. Which is what is fun with doing all the analysis for the AI and the vectors was I had to parse all of that data and put it into vector format. It took a long time. All right, let's do our 2026 predictions and then we'll probably wrap it up for this episode, per Brian's suggestion. [00:46:03] Speaker A: Yeah, I didn't. I didn't check the message, but I was hoping that's what it was. [00:46:09] Speaker C: All right, we did roll before the thing to do your predictions. And so Matt, you rolled first. And as the winner of 2025, I do think that's fair that you should be first. High expectations. [00:46:19] Speaker D: I don't really want that. So I have a couple, some that are less nice than others, but I think I'll start on just kind of a sad one. I feel like there's been a lot of major outages across the vendors recently and it's time for GCP to step up and get some Hugops here. And I am predicting a outage of her GCP this year. [00:46:43] Speaker B: Thanks. [00:46:44] Speaker C: Thanks. [00:46:44] Speaker B: You're welcome. [00:46:45] Speaker C: I really appreciate it. [00:46:46] Speaker D: Now you guys work. [00:46:48] Speaker A: Is it going to be caused by AI though? That's the question. [00:46:50] Speaker C: Yeah. [00:46:53] Speaker A: Will it be the first RCA that goes out because an AI agent did something it shouldn't have done? [00:47:00] Speaker D: It just kept scaling up until there was no more scale to go up. [00:47:05] Speaker C: All right, I'm up a second. So I have two thoughts in the world for 2026. Either it's going to be a massive recession and we're all going to hate everything, or it's not going to be so bad. And so I'm going to try to go with some positivity here and say it's not going to be so bad. And because it's not going to be so bad, there's going to be a lot of AI layoff. Regret and we're going to see a return of hiring in 2026. So that's what I'm hoping for. I have, I fear it may be the other way. Which one of you may predict as well. [00:47:32] Speaker D: I had that. I'm not going to lie. That was the other one I was going to go with. [00:47:35] Speaker C: I do have other. I do have another negative one I could use as well. But we'll continue if it's still on the table when the back. Yeah. But I am hoping to see AI lay African and companies coming kind of back to some hiring in the world. So that's my, my hope. Although Venezuela makes things more complicated all of a sudden very quickly. All right, Ryan, you're up. [00:47:55] Speaker D: All right. [00:47:56] Speaker B: I think that the, the multi agent orchestration is going to blow up in a big way. And you know, just to be a little specific because that's a little obvious. Yeah, I, I feel like the, the major providers are going to announce more, more of an agent to agent integration that's more native in the sense of like the, the workflows naturally tying together between like Vertex AI Bedrock, whatever Azure. [00:48:18] Speaker C: Has. [00:48:20] Speaker B: And you know, where it's more of a. An easy button and not something that you sort of have to set up and three different ways and then plug it together and do all the scaffolding yourself. I think you'll start seeing that even beyond the hyperscalers where you'll see that agent interaction with providers as well where you can just naturally have sort of an agent workflow or a multi agent orchestration workflow that just uses agents from larger companies and cloud providers. [00:48:52] Speaker A: Makes me wonder. That's a really good prediction. It makes me wonder how long it will be before we have like a business to business AI to AI type protocol where my business wants to talk to your business but we delegate those tasks to the machines to do the negotiations. Negotiation or something. You know mentioned the first multi million dollar dollar contract negotiation being performed entirely by LLMs or, or agents of some kind. [00:49:16] Speaker D: Wow. [00:49:16] Speaker A: Kind of cool. [00:49:17] Speaker C: That would be cool. [00:49:18] Speaker A: That is not my prediction for this year. [00:49:20] Speaker B: No. [00:49:22] Speaker C: What is your prediction for this year? Your first one. [00:49:25] Speaker A: I think that the negative one is. I think there will be a major. I say major. [00:49:33] Speaker D: Ah. [00:49:33] Speaker A: Because he's such a stickler for these details. He write down every single word and hold me to it. At the end of the year. [00:49:40] Speaker C: I. [00:49:40] Speaker A: Think there will be a highly visible company bankruptcy due to the rising costs of AI inference. You know a perhaps a popular company, a well known company or well liked company will go under because of the rising costs of providing the service they're providing and will no longer be able to provide the service. [00:50:04] Speaker C: That would be pretty bad because that would. Yeah, that would also probably mean where the AI bubble is real and someone's overextended. [00:50:11] Speaker A: It's absolutely real. [00:50:13] Speaker C: I mean, based on my research of just last three weeks doing vectors, I'm like, yeah, this has some limitations. All right, Matt, your second prediction for 2026. [00:50:29] Speaker D: So given that Microsoft decided this year they solved quantum computing and nobody else has, and then magically stepped back on it, and I don't know if this is going to be positive or negative. Positive as we will get new technology, but negative as half of encryption protocols all die over. At that point. I'm going to predict a step forward in quantum computing. [00:50:54] Speaker A: A quantum leap, if you will. [00:50:56] Speaker D: There we go. There's the, there's the show title right there. [00:51:00] Speaker B: Nice. [00:51:01] Speaker D: A Quantum leap into 2026. There you go. [00:51:07] Speaker C: I, I was curious. We were talking about Marjana. I was like, you know, it's been a while since I looked that up and yeah, it's basically been radio silence since March when we first talked about it. There has been nothing in the press about it at all other than Microsoft's continuing to stand by their statement. [00:51:22] Speaker A: I'm going to hold Matt a little bit more accountable than just saying an improvement in some technology, though. Like, do you think it'll be. There'll be a real world use case for it or. [00:51:30] Speaker D: No, I think it'll still be. It'll still be in the laboratory, but it will be a good step forward that you will kind of transform the next generation of the way we think about it, you know, so if you kind of think like, how do I word this? [00:51:53] Speaker A: Like a different way of approaching it. Sort of a, sort of a revised vision of how we can get. Achieve the same thing, but maybe not with, you know, near absolute zero quantum computers in a, in a basement someplace. [00:52:06] Speaker D: Right. And it's not going to be like, you know, consumer grade or anything else like that, but it'll be, hey, we've taken a big step forward. Now granted, there might still be 15 more after that, but it'll be a good step forward to say, okay, we've solved this thing, like you said, like the near zero or whatever, the thing that they are able to solve is and be able to, you know, then you'll start to see, you know, the incremental improvements till the next big step forward. I'm waiting to see how Claude next year decides if I win a point for that or not. [00:52:41] Speaker C: Yeah, the. I do sort of feel that the key to making AGI work is quantum computing. I sort of have this feeling in my brain that that's going to be the key thing to make that happen. But I'm curious because OpenAI and everyone else seems they're close. But I feel like if you think about the human brain can do and what, you know, the ability, we can imagine infinite possibilities, which is something that quantum computing promises to do. And so combining that with a large language model type system I think is how you get there. But again, Madrana is not real and other quantum qubit stuff is not close. Then it doesn't really matter at this point. But I think that's why AGI is a next year thing. It's, it's probably within 10 year kind of problem, but it'll be interesting when it happens. All right, my next prediction for this time around, I'm going to go for the. My bad one. I start to see the first AI agent security breach. This will be an agent that basically breaches an organization and exfiltrates data automatically by itself without any interaction. [00:53:44] Speaker B: This is what keeps me up at night. Absolutely. This is trying to think through. [00:53:53] Speaker C: Emerging. [00:53:53] Speaker B: Threats on technology that I barely understand because it's coming out so fast and it's, you know, changing just the way we work. And so you're already starting to see AI in, you know, attacks where, you know, groups of people are using AI to put together some pretty sophisticated attacks on companies. [00:54:16] Speaker D: Right. [00:54:16] Speaker B: They're, they're easy. It's a lot easier for people that aren't natural language speakers to generate content for spear phishing. It's a lot easier for, you know, malicious actors to, to have an AI agent go do a bunch of research on the company real quick and, and provide sort of the tailored, you know, like this is where I think it'll be weak. And so you're already starting to see that. So I absolutely, I fear, but I think you may be right. [00:54:44] Speaker A: Yeah, it's a whole new level of robo calling. [00:54:47] Speaker C: Yeah. Yeah. I mean, I find it funny how basically within, you know, three months of ChatGPT coming out in existence, basically all phishing training was invalid. Yeah, you know, it just, it's crazy. So, yeah, it's going to get worse. All right, your third second prediction, Ryan. [00:55:06] Speaker B: I think over 20, 26, that infrastructure as code is going to move to infrastructure as intent. So I think no, you know, like we'll start instead of sort of defining configuration for you know resources and cloud providers. It will be natural language. And so you'll ask for, you know, you, you won't ask for, you know, 100 servers. You will tell it your workload and what you want to achieve and then the process will sort of con. Continuously sort of optimize and provide that. I think we'll start seeing that pattern emerge in 2026. [00:55:49] Speaker C: All right, Jonathan, your second. [00:55:54] Speaker A: How many times are you going around it? [00:55:56] Speaker C: You get, you get three total. [00:55:58] Speaker A: Yeah, three total. Okay. I think there will be an explosion this year of competition against existing SAS companies. So, you know, looking at what Claude can do and chat GPT can do now in terms of autonomous coding that can, that can last for days and days on end to, to recreate things, I'm absolutely convinced that if you don't want to pay GitHub.com for their service because you don't like them moving to visual as, as a provider, I think you will very easily be able to ask an AI to build you a very GitHub like service for your own use. I'm not saying you necessarily sell it to other people, but I think a lot of companies are going to lose a lot of business because people will recreate services that are already paying for using AI. [00:56:51] Speaker C: Yeah, I think we talked about this on an episode you weren't around for, but you know, I kind of have this belief in that, in that, you know, big companies, you know, 100 million plus revenue companies, we're already building their own things like meta builds their own load balancers, you know, these type of things. And I think what happens in the AI world is that you see the barrier to entry maybe drop from $100 million revenue companies to 50 million revenue companies. But I still think there's a large amount of medium and smaller enterprises that just don't want to have to, you know, even maintaining all of that code that agents generate and having to do it, they're not going to take the compliance risk. They're not going to want to take, you know, maintaining it and doing all this stuff. And so while they could and some might, I don't know that it's necessarily the end of SaaS as we know it, but I definitely think it changes the market dynamics dramatically. [00:57:45] Speaker A: Yeah, yeah. I don't think it's the end of SaaS for sure. There's always going to be compliance requirements, there's always going to be scaling and deployment things. But I'm just thinking just in terms of general technology, you know, we have a lot more things around the home now than we ever used to. You know, we never used to have photocopiers. We used to have to go to the library or sneak into the office at work and copy some things. But now so many people have got printers and scanners you can scan with your phone or whatever. Just technology seems to filter down and down and down, become more commoditized and I think that these tools are going to make software products, sort of put software products into that category as well. You know, make me something that works like GitHub that I don't have to pay. It doesn't have to be perfect. I just wanted to do what it does. I think that's gonna have an impact on, on the industry and I think it's a good thing because I think people will be able to build a lot more, probably get a lot more innovation, but at the same time I think it will impact the bottom line of some. [00:58:36] Speaker C: Oh yeah. [00:58:36] Speaker A: I think the key thing, overpriced companies because a lot of, a lot of SAS is very overpriced right now. [00:58:41] Speaker C: It is, yeah. I, I do think it's, it's going to be interesting of like how do I. The Data that those SaaS companies have like a salesforce and being able to know lead generation and some of the metrics and KPIs and like being able to make that data accessible to my, through agent to agent protocols, I think is really interesting. So we got a whole conversation with this because I thought it was a lot having working in the SaaS business. You know, I've had a lot of thoughts about the different ways this might go. But you know, I think there's through MCP and agent to agent, I think that's how these companies start selling the data value that they have and that experience to then empower companies to build their own tooling and agents. And I think it's part of my next prediction actually, which we'll get to when I get to it, but we'll talk about in a minute. But I think it's a good one. Matt, you're up. [00:59:29] Speaker D: So I feel like we're kind of seeing this now, but I may predict there's another kind of more micro hyperscaler that pops up. So I'll explain a little bit what I'm thinking. But you have DigitalOcean and they kind of do general cloud, you know, general, here's a small server for that, you know, can do whatever and they, you know, they, they have a very specific market segment. I predict another one like a digital ocean or Digital Ocean competitor, but, like, something that's going to, like, be a little bit more niche, whether it's going to be all we do is GPUs or all we do is something else. You know, was it Black Rack only does storage, you know, like a specific niche that comes out. [01:00:18] Speaker C: So, I mean, they. There are a couple of hyperscalers who I keep track of right now who are purpose built for AI, But I specifically think you need to quantify this to, like, what do you mean? Like, from a size perspective? Because, like, I think they exist. The question is, do you think. You think one of them becomes a competitive to, like, Oracle or solid number five above Linode or DigitalOcean or, like. [01:00:44] Speaker D: I think it comes around the same. Around the same level as Digital Ocean. [01:00:50] Speaker C: Okay. [01:00:52] Speaker A: Well, we already got like, co. I mean, you talk about like, co. Yeah. [01:00:56] Speaker C: I think Crusoe and Core Weave are the two that come to mind for AI Specific Cloud. [01:01:02] Speaker D: Yeah. [01:01:03] Speaker C: I don't know what their revenue is these days because I think they're both private. [01:01:11] Speaker D: But we'll see someone come up whether it's AI specific or not. But I feel like right now the world is still AI specific, so, you know, we'll go with AI. [01:01:22] Speaker C: Well, so according to Gemini, Core Weave is 1.20, is 5.1 billion to 5.3 billion in productive revenue in 2025, where Digital Ocean was 870 million and Crusoe was 500 million. So we'll have to see. [01:01:38] Speaker D: I don't think I'm gonna win on that one, but I think it'd be fun. I think it'll be fun to have another. [01:01:44] Speaker C: I think it's already happened. I think we're just not covering it enough here at the Cloud Bot. I think it's really the truth, what we just discovered. [01:01:50] Speaker D: That's what it is. [01:01:52] Speaker C: I'll maybe start pulling in more Core Eve stories. It's been on my radar to do. I just haven't had time to get through. All right, so kind of alluding to what we're talking about with the sass thing. So the Internet was designed by people and education to basically provide data to humans. And so then, of course, the marketing companies all abused that by now, trying to monetize eyeballs into marketing dollars and ad spend. And so I think this is probably the year that we start seeing Agent First Design start creeping into major E commerce sites and major websites. The idea that instead of fighting the AI and them crawling your website, that you want to actually take advantage of AI to help sell your Products help to do things and you know, transactions that don't require humans. And so I think we see the beginning of the AI designed web in 2026 coming out in a big way and start to see it more and more in conversation. [01:02:50] Speaker A: Yep, that's, that's one of mine. So thanks. [01:02:55] Speaker B: I'm trying, yeah, I'm trying to decide if that's one of mine because my, I have a couple different angles. [01:03:00] Speaker C: Which am I next? You are next. [01:03:03] Speaker B: I can just sidle into that then because I was very close. I was thinking we'll see, you know, like a full stack, you know, generative media platform that does, you know, the text to image, image to video, text to speech, music, the whole thing where, you know, enterprises will be able to sort of upload their, their brand guidelines and it would, you know, automatically output, you know, the full sort of content management system and then also with built in provenance tracking so you know, generating content would have, you know, watermarking that would distinguish. It is, you know, both as, you know, the ownership of it, but also AI generated. [01:03:46] Speaker C: So Soros, but owned by an entity because, you know, we heard about Disney licensing their stack to Sorrow. So you can basically go have. [01:03:57] Speaker B: It's more about the tooling. [01:03:58] Speaker C: Yeah, more tooling, you know, so, because. [01:04:00] Speaker B: Right now like you can, you can do, you know, you can, you can have, you know, Nano Banana generate an image, you can have Veo create a video, but it's, you have to stitch it all together yourself where I think that we'll start to see tooling that sort of puts it all out there together. [01:04:19] Speaker A: Yeah, yeah. Along Justin's lines, I was thinking this is, I'm not going to use this as my prediction. I was just thinking how the Clawboard Web Chrome plugin really highlights the problem with agents actually interacting with things that were designed for humans. And so find this button, click here, do this thing. And they even try to figure out how to do those things from, I mean they got access to the whole document model in the browser. But even just figuring out where is this button the user's talking about when they're saying it's, it's above the title right there. Well, where the hell's that in the code? It could be anywhere. And so we've, we've been through this whole SEO process. We have metadata in the pages and everything. I think we're going to have an agent optimization system which, where we kind of define intents in the page. This is, this is how an agent would do this. This is how an Agent should order a Starbucks through the app if the user's got a local AI assistant on their phone, for example, so that. So that those hooks are there and well presented so that the agents can perform actions on behalf of users. That is not my prediction. [01:05:24] Speaker B: That's. That should be, though. I think that's pretty cool. That's almost like, you know, like lightweight MCPS for. But like at a website page, you know, sort of. [01:05:34] Speaker A: Yeah, it could be. It could be pages, could be. It could be mobile apps. It could be anything. But I think we're going to start seeing very can't speak, declarative ways for agents to interact with systems that were previously designed for humans. I should make that one of my. One of my things. It's slightly different from Justin's. [01:05:50] Speaker C: Whatever. [01:05:51] Speaker A: Maybe I'll have that as a bonus point for when I win all four points. [01:05:56] Speaker C: All right, so then what. What do you want your last prediction to be? [01:06:00] Speaker A: My last prediction is this year we will release an entirely AI generated podcast. Essence episode. Huh. [01:06:10] Speaker C: It was. It was a wise choice. He could have gone for, like, I'll be at more than 50% of the episodes this year, and he would have. He, like, this is how he's going to do it. [01:06:17] Speaker B: I think he wants to win the point, though. [01:06:19] Speaker C: Yeah. Probably. [01:06:22] Speaker A: Just count if I. If I just, like, swipe in and then leave again. I don't know. [01:06:26] Speaker B: I mean, that's, you know, that's how their RTO policies work. Right. [01:06:30] Speaker A: Can I just give you my card? [01:06:32] Speaker B: Can you coffee badge into a podcast? [01:06:34] Speaker D: Can I just give you my voice? [01:06:36] Speaker C: I mean, that's. So the service I'm using for the transcripts now, they have that ability. I can just give. I can take your audio samples and I could just have it generate from text the podcast. So we could. But I feel like we'd lose a little something, but. [01:06:50] Speaker A: Oh, it definitely would. Yeah. Yeah. I've been playing around with Microsoft released a speech synthesizer thing. I've been playing around with voice cloning, which isn't good enough yet by any means, but I think it'd be really interesting to try and do a. A good AI generated protocol test. [01:07:07] Speaker C: I think it would be fun. I would do it as an experiment. [01:07:09] Speaker A: Yeah. Yeah. I'm not saying it would be. [01:07:11] Speaker C: I just want to see what it would be like. But until. [01:07:14] Speaker B: Until our users get back to us that it's better and we're in trouble. [01:07:17] Speaker A: Yeah. [01:07:18] Speaker C: Then we're screwed. All right, well, that's fantastic. Back at 2025, and I hope our predictions come mostly true. Not the bad ones, but the rest of them. So we'll see how we do. We will keep track as we do every year, and we'll look back in December or January next year. So all good. All right, guys, have another fantastic week here in the Cloud. [01:07:43] Speaker B: Another fantastic year. [01:07:44] Speaker D: Happy New Year. [01:07:48] Speaker A: And that's all for this week in Cloud. 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.