They discuss whether the AI agent hype is supported by actual buyers, how startups are faring as AWS focuses on large enterprises, and how many of the new technologies coming out this year will actually be transformative.
Episode notes:
This episode was recorded at AWS re:Invent 2025! Check out Ryan’s recap of events on our blog.
Duckbill provides financial planning and analysis for enterprise infrastructure to help you understand, negotiate, and optimize your cloud spend.
Connect with Corey on Linkedin and subscribe to his newsletter Last Week in AWS.
TRANSCRIPT
Ryan Donovan: Zenflow launched this week and it changes AI coding. It replaces prompt chaos with AI orchestration. Zenflow turns specs into plans, executes across code base, and verifies results of multiple agents. If you want AI that writes correct code, not just code, this is it. Download at zenflow.free.
[Intro Music]
Ryan Donovan: Hello, and welcome to the Stack Overflow Podcast, a place to talk all things software and technology. I'm Ryan Donovan, host and we've been recording all week here at AWS re;Invent, and today, to put the nail in the coffin, my guest is Corey Quinn, Chief Cloud Economist at Duckbill, and the main guy at Last Week in AWS. Welcome to the show, Corey.
Corey Quinn: Thanks. It's an absolute pleasure to be here. We are now on month four of the event. The sun has not risen since late October, and we are hopeful that when the spring thaw eventually comes, we'll be able to get the wagons out and continue our perilous journey quest.
Ryan Donovan: Yes, well said. My dog's a barking. So, like you've intimated, it's a long conference with a lot of things going on. What is your general feeling—and we'll drill down—of the announcements in the conference?
Corey Quinn: The general sense is that for a few years, it's been a problem, where AWS is focused, almost to its own detriment, on large enterprises. That's where the money is. It got to the point where independent developers, folks who are [like], ‘I'm curious about this thing. I want to try it,’ sort of feel like this is not really a place for us anymore, historically. My customers are giant companies. I build things by myself. Some call it vibe coding. I call it cyberbullying or robot. It's brute force mixed with enthusiasm, and you can build mountains. But it's nice to see a lot of things that are coming out that are aligned with both sides of that spectrum. And it makes sense. AWS is far large enough now to walk and chew gum at the same time.
Ryan Donovan: Yeah.
Corey Quinn: But now we're just talking about AI.
Ryan Donovan: I'm here covering the startup program, and it's definitely nice to see these smaller, early-stage [projects], and AWS supporting them with cloud credits. Do you think this sort of startup support is new? I mean, obviously it's self-interested, but do you think it is an overall good thing for the ecosystem?
Corey Quinn: Yes, and I've talked to too many startup founders over the years who will effectively follow the free credits, like they're migrating caribou or something. And– right, well, we're gonna build in a way that's agnostic and runs anywhere, and that doesn't tend to pay very often. Pick a provider—I do not care which one, you might care which one, I don't care which one—and lean on and integrate with their environment. Otherwise, you're trying to build the least common denominator story, and it doesn't work. I think folks worry too much about credits in some cases, and in others, they don't worry enough about them, where, ‘oh, it's free, it's not real money. It's like some sort of weird cryptocurrency called AWS credits,’ and then suddenly it turns into real money when they run out, and suddenly the screaming begins. Right? But it's also something I think not well appreciated in many circles. The infrastructure, then, is almost always taking a backseat to people's time. Yeah, you can theoretically run a managed database service like RES for more money that would cost to just run the open-source database on EC2, provided your team's time is free and that there's nothing better suited toward running on that. So, it's a nuance, and it's a mixed bag. I think that there's still a lot of people who look at these big deals that companies are signing and saying, ‘oh, well, they could run a data center for half of that. They must be stupid.’ Right, well, do you think they signed a billion-dollar contract just out of signing inertia, and it landed on their desk at the right moment, or do you think there were maybe a few discussions about this, and is this right for what we see our business doing? Believe it or not, compared to what I say on the internet, most big companies do not hire idiots into executive management. They have context and constraints we don't get to see.
Ryan Donovan: Yeah, and you know, most test software companies don't wanna be a data center company. Obviously, you hinted at it – everybody here is talking about AI. He is gripping the bridge of his nose, for our listeners. What do you think about the AI agent announcements?
Corey Quinn: If you think back many moons ago, back in high school, it seemed like everyone was talking about hooking up. The number of people that were actually doing it was a lesser number. Everyone is talking about AI. Everyone is selling AI. It is not that everyone is buying AI. There is value there. There is a use case. It’s not cryptocurrency. We're not seeing it as one giant hype train with no substance to. But at the same time, I start to wonder if, at some point, it's an assistant, it is not the end-all, be-all. Generally speaking, you can assume in most cases, play with the numbers to your taste, that it'll get it right 80% of the time. So, what use cases are there where that 20% ‘wrong’ factor is going to not be disastrous? Maybe you don't put it in charge of writing your corporate columns without supervision. Maybe you don't have it write your quarterly reports and not have a human to review it, but maybe you can have it start gathering and curating stuff, and starting to act as an assistant on the human side of it on your side, and exposing it directly to your customers is kind of asking for trouble in a lot of scenarios.
Ryan Donovan: Yeah, I mean, especially in the customer service interactions getting mean. We found with our last developer survey, basically the more developers use AI, the less they trust it. And I think that is a good thing because they start seeing the mistakes, they start seeing the limitations of it.
Corey Quinn: Well, I don't know about that. I mean, I've been vibe coating nonsense for a while. I have experienced in the backend and DevOps side of the world, which with that stuff, I have to very often jump in and slap the gun outta it’s hand, metaphorically speaking. But on front end, which I know nothing about. It's always clearly perfect, and it knows more than I do, and I'm starting to wonder, ‘hmm, maybe I have a bias here that needs to be addressed on some level.’ This is always the truth of AI, where the more you talk about a subject that you know well and you start to realize that, in some cases, it is a veneer of [BLEEP] over no substance. That's part of society in some cases. I mean, tech resumes, those are a form of stylized, ritualistic [BLEEP] that AI is very good at structuring into. But humans don't tend to think in most of those ways, most of the time. It's useful there. I mean, it writes like a middle manager, and part of the problem is that we start to think that it's, ‘oh, it's like a person,’ instead of realizing the truth, which is that middle managers generally have no souls.
Ryan Donovan: I mean, it's a statistical mean, right? We published a post a while back that was [on] data-driven decisions, innovative decisions, and I think data-driven language is not innovative language. It's just like, this is what everybody's doing. Here's the sort of common middle ground of this answer, right?
Corey Quinn: Yes, and I think people are lazy at phoning it in when it comes to prompting, for example. Like, you can have a lot of fun with it. Like, you are my Compsci 101 professor with a fairly disguised drinking problem and a massive chip on your shoulder. Explain recursion to me. Or you are an old-timey Victorian doctor prescribing leeches, who also has strong opinions about TypeScript, and then ask it for other things, and you start to see it really come out with some creativity. I love using it for titles for blog posts and things like that. But the way I found success with it is, ‘okay, give me 10 options. Right? Give me 10 more.’ And the answer is, I'll pick a couple and melt them together and do something else. As an assistant, it's great, but write a blog post as me? Yeah. That'll go well.
Ryan Donovan: Yeah. There's somebody who already does that. That's you.
Corey Quinn: And they trained it on me, so there is that.
Ryan Donovan: And the AI stuff that he announces specifically around agents, do you think those are wave of the future, super good stuff? Or is this dubious marketing?
Corey Quinn: It's dubious marketing speak until in a rush you can arrive at a consistent definition for what an agent is, because right now the consensus is the thing I have to sell you is agentic AI, which okay, great. Like the old line of, ‘you can't buy DevOps, but I sure would like to sell it to you.’r The same type of approach where there's a fence that having agents do things and interact with each other, it does lead to great outcomes. I have a virtual Corey bot that can write snark the way that I do with a surprising degree of accuracy, and the way to do it, in my experience, has not been with a prompt, but rather different agents, each with a different tone, working collaboratively, very similar to the style of a comedy writers' room. And then there's improvement there orchestrating those agents is interesting. They’ve not got all these enterprise options to do it. I've had great success with treating Claude Code as an agent orchestrator. I think people are overcomplicating it, but that's not enterprise, which is code for, ‘that's not expensive enough.’ And well, you know, that's okay. But one thing I will say with all the marketing for AI is that they seem to be targeting a customer persona of walking around, going, ‘my biggest problem is I just can't spend money. Would someone help me with all this money piling up and up?’ That's not been the actual stated pain of any customer I've ever spoken to.
Ryan Donovan: No. I mean, I think the real pain that Babe mentioned in a session was that people would come saying like, ‘how do I cut 30% of my workforce with AI?’
Corey Quinn: Yeah. That's the challenge, as well – ‘okay, you know all these people we have here, they’re doing good work, and how do I get rid of one out of every three?’ Like, that's not the kind of consultant that I am. That's never been something that's been of interest to me. I think that you'll find that you can use it to accelerate what individual folks are capable of achieving, and it's great. AI-powered coding assistance is fantastic when used in the right way. But does that mean I'm now going to replace all my junior engineers with it? Well, hang on a second, because even if you do, hypothetically, let's assume that it's perfect at doing a junior engineer thing. It's one that you don't have to pay and that you can actively mistreat. It's like a new grad working at Amazon. Now, the problem that you have then is okay, where is it that you believe senior engineers come from? Do you think that they just spring fully formed in the forehead of some god, or did we all start as junior engineers and now that it feels like the ladder is getting pulled up, where does the next generation really come from? I don't know what I would do if I were in school right now learning, alright, I'm gonna go work in computers because I hate the world and myself, but I don't know what that path looks like. I'm very concerned of anyone who’s always offering a certification program or a training course on how to use this. It's evolving so rapidly that it feels less like, ‘I have knowledge I am sharing with you,’ and more like I'm gonna grift.
Ryan Donovan: Mm-hmm, and I think that sort of apprentice program, it's in my job as a– it's in other jobs as the mechanic, whatever. If you don't have somebody saying, ‘you screwed up right there, here's how you do it better,’ Instead, you have a junior just doing LLMs all day. The LLMs don't tell them how they screwed up.
Corey Quinn: Right, there's a learning experience. The idea of mentorship, the idea of, ‘great, this is the purpose of code reviews.’ Rather than for the seniors to flex on the juniors, they've really decided it's like, ‘yeah, your goal is to help train the next generation. Like, okay, I'm not just saying that this is wrong; here's why I think it's wrong. Here's how I arrived at this. Here's the broader context behind why that's wrong.’ Like, ‘oh, you wound up addressing a database by a single endpoint. You can separate out the reads and the rights.’ Yeah, as a small toy app, and it doesn't matter until it scales, and suddenly it matters, and you're gonna spend a week hunting that down and fixing it out. People are very good at things like that the second time they run into it. So, it's carrying that experience forward. The problem I have is every time I correct a robot, or incorrect the robot, ‘you're absolutely right.’
Ryan Donovan: Yeah.
Corey Quinn: It's like, don't do that. Maybe I'm wrong. It's anathema to suggest the user might possibly be wrong about things, but I'm wrong constantly.
Ryan Donovan: I mean, it's one of my superpowers. You know, as a tech writer, for a long time, being wrong got me answers.
Corey Quinn: Oh, absolutely. And I find that people also tend to wanna hold onto the idea of, ‘well, I was wrong, and I said something, but I can't lose face by admitting it.’ It's, ‘well, I was presented with new information and I'm gonna walk back the thing I previously said. Why is that considered a bad thing? Otherwise, it's like, nope, I'm gonna double down and be the world's most odious blowhard. Let's go.
Ryan Donovan: Part of the conference is the announcements, a small part, but the larger part of the conference is just the sheer number of developers. People voted here. I assume you've been going around and talking to other people about their experience in the software world, at this point. What is your read of the temperature?
Corey Quinn: I think that AWS services, at least the ones that are [BLEEP], have evolved to a point where they are largely feature-complete. Like it used to be the big thing – ‘I really hope we get RDS replication multi-region,’ because I had to build that three times previously, and when it finally came out, it was amazing, except for the fact I'd just done it the week before and I wanted to basically have a conversation like Tam and a bone saw for a couple of days while I cooled down. Then it was, okay, now the enhancements are less earth-shaking and more quality of life. More have very specific corner cases, and that's great. That's important. Customers care about this, but looking at all of this, there's remarkably little that is going to be transformative to how I do things when I go home. I used to warn people, ‘don't make big sweeping changes for about six weeks before re;Invent.’ This year, the only thing that I talked to people about on that was maybe don't buy a bunch of database reserved instances just now. Ask me no questions; I'll tell you no lies. When I'm talking to folks, I’m worried about this, because I knew what was coming. But also, I'm not gonna break confidentiality. Maybe it'll make huge sweeping financial commitments before you go and have your EBCs and talk to your customers. If you're insisting on doing it now, talk to your account manager, please. You wanna make sure that people don't have situations they regret, but you also don't wanna leak information. And that's important.
Ryan Donovan: Right. The greater flexibility they're doing in some of their EC2 instances and AI instances, because I think we at Stack Overflow ran into that when we were moving. We needed something that had tons of memory, but less compute. And it wasn’t available.
Corey Quinn: And now, of course, they do. In fact, people don't talk about this much, but I believe there's an instance now they have 24 terabytes of RAM, and it's like first, well, what are you gonna do with that? Using two slacks at the same time. Beyond that, though, what does it cost? It doesn't matter, you know? It does not matter because people use that to run SAP HANA, and compared to what that thing costs, it's like a few million dollars a year for an instance. Drop in the bucket. ‘We run SAP HANA,’ ‘what does it run on?’ ‘Burning fields of cash.’ Why? Because it powers our entire freaking business. This is the problem I've gotta with the transform narratives. ‘We're gonna take your [BLEEP], old mainframe and put it in the clouds.’ People have been trying to get off the mainframe for 40 years. Do you believe they've not succeeded 'cause they're just a pack of idiots who've never seen this magic thing at the AWS booth? They had a big banner that said, ‘Eliminate Technical Debt Forever.’ AWS transformed. My god, it was counterpoint. I'll bet you a house it doesn't. We've all seen these magical solutions before. Is it better? Sure. Will I believe the transform is useful? No, not until they first stop emailing me 400 times every time they deprecate a Lambda version. Like, great. If you're gonna go from Python 3.13 to 3.14, great. You should be able to seamlessly update my code for me; that is the easiest of AI-powered pull requests. Run some tests, and you may have to have a test framework first. Fine. Make that a requirement for the feature. Awesome. Or have AI do it. But once that's done, just make it do it for me instead of giving me busy work every six months. I have enough of that on my own, called expense reports.
Ryan Donovan: I talked to somebody at AWS who dropped the Java conversion number where it was like 6,500 developer years of time. And it's like, yeah, that Java upgrade – I've been to companies where they sit on Java four or five point releases. Four or five whole releases, because upgrading is a pain. And if that's all AI does, it's probably a decent solution.
Corey Quinn: That's a perfect example. Another reason they don't upgrade Java for years at a time is that they don't wanna play slap and tickle with your lawyers when they start playing games with the new version’s licensing terms. AI was like, ‘oh, terrific news. I didn't upgrade for you. And by the way, you have legal demand incoming now. Don't thank me. It's my job.’
Ryan Donovan: Warm reminder that all opinions expressed in this podcast are those of the speakers, not the Stack Overflow Podcast. Another of the things he announced was the database savings plan. Obviously, there's jokes on the internet about a genie who islike, ‘you have $10 million to spend it in a day, and then you get to keep 10 million. I'll just use AWS, right? There's another–
Corey Quinn: This came out years ago, someone mentioned it on Reddit. There are three rules: that you can't do this; give it away, waste it like Bruce's millions approach, and if you could spend 100 million, I'll give you a billion dollars. And I looked at that for a second, and I'm like, okay, and I wrote a blog post about it called The Cloud Genie. And I'm like, I think I can get the whole billion spent in the month. And it's some of the boring stuff – I'm just gonna spin up a bunch of big instances. It was circular data replication through multi-region through S3. And the question then just becomes: what is the rate-limiting on the network transit there, because it is fiendishly expensive when the Lambda invokes itself-style patterns. But there's better news for that now because the other big news is that S3 objects can now be as large as 50 terabytes, which unlocks two or three great uses, and thousands of monstrous ones that I can't wait to have fun with. For those playing along at home, at retail price, US East one, each one of those 50 terabyte objects in standard storage will cost $1150. The fact that I know that off the top of my head says something profoundly sad about what it's like to be made.
Ryan Donovan: I mean, your cloud accountant, so it's your job to know the numbers, right? So, with the business that goes on here, there's also the ‘Vegasness’ of the event, and everybody has their branded restaurants, cabanas – do you think that is an important part of the re;Inventing experience?
Corey Quinn: In Las Vegas, the theme and the motto of re;Invent is ‘closed for private event’ because every time we try to go somewhere, it's closed. It's an experience hanging out here. It's one of those things where I think everyone should experience once, and then never again if they can possibly avoid it. And I'm trying to make it less onerous, repeat. And we have two talks. The other one I just stepped off the stage from before we had this conversation – it was about ‘thal key’ and at the end of it, I launched a demo, which you can look at yourselves if you want. It lives at [BLEEP] hosting.ai, and it has a list of all of the announcements of re;Invent snarked on by an agentic system, like the way that I snark on them. And then it generates a trip report as you, in my voice, that talks about the nonsense that is happening here. And that way you can submit that to your boss on having to write it yourself.
Ryan Donovan: You're putting yourself out of a job here. So, something I've been thinking on that level is: are we headed to our future where people are gonna put artisanal on their snark, on their blog posts, on their marketing notes? Is that something we're gonna have to do in the future?
Corey Quinn: So, I have done an awful lot experimentation. Can I get this to right as me? And it can nail aspects of the tone of snark very well, but it misses a lot of the insight. And what I've learned is that— maybe I'm flattering myself here—without the insight, maybe my snark is just annoying and doesn't help anything. I feel like I'm teetering on the brink of a revelation here, except I'm not self-aware enough to realize it. So, let's just keep it that way. And again, everyone's like, ‘well, it won't take away my job,’ but here's the reality. What you were doing 10 years ago is not what you were doing today in almost every case, and it is incredibly, like, whatever your job is today, the odds are, if you listening to this kind of a podcast, your grandchildren will not recognize that job. If you want to do the same thing more or less with minimal changes, year in, year out, there are paths like that. An accountant hairdresser… the tools I was using 10 years ago are basically antiquated now. I would argue that despite being a platform engineer, whatnot, you're still a sis admin. Your job is to keep the site or estate up. The tools, the processes, the way we talk about it has changed, but the core responsibility has not. ‘Keep the computers working’ is sort of a thing that's going to be there forever, but increasingly it's gone from, ‘that's where the best of the best architects do,’ to ‘oh, that's like a white collar janitor.’ People look down at their noses when there is rat working in the data center, which I think is unfair. Without very good people doing those things, none of the rest of this house of card stands.
Ryan Donovan: I mean, the things that AI and agents are doing are fundamental CS concerns, like greater abstraction, automation… This is what CS people have been doing forever. You no longer have the punch cards and such, right?
Corey Quinn: Of course. And I've been using it for a lot of things for weird, little, tiny business apps, because I write a newsletter. I built a publication system about six years ago, and then I've been meaning to get around and fix up a whole bunch of it for ever since then, but I don't have the two weeks to spend doing it. And hiring an engineer to do this is a really sizable endeavor. You can print business app; you can generally ballpark it starting at $20,000. And so, okay, there is value there, but there's other things that are more high priority than this. Now, I can effectively break it into microservices that are easily manageable and start vibe coding aspects of it, test suites, and the rest, and it's leading to interesting outcomes. There are tools that are automated, things I used to do by hand. A great example where you can get it wrong one time out of five is: this week so far, I have well over 150 announcements I need to go through for what AWS has put out for maybe 15 or so things to include in the newsletter. I now have AI doing an initial curation pass where it gives them all a score and shows me a ‘keep and toss’ pile, and I go through that, and I make the changes. Then it looks back at the last 10 issues for what I've done and tries to learn from it. I still make the reviews because it gets it wrong sometimes.
Ryan Donovan: You still review every single–
Corey Quinn: But, Hmm. Wow. Some partner-promoted blog post about a thing I've never heard about and care even less is probably not something that the rest of the world needs to hear about. Oh, this database savings plan. Yeah. I'm telling everyone about this, like, at the dinner table my kids are very tired of it. I don’t care.
Ryan Donovan: What announcement do you wish they had made at re;Invent? What's the thing that you're still young about – the years ago– the database savings plan, whatever.
Corey Quinn: META net gateways, Cross Easy Data Transfer is stupid expensive. Egress is really expensive. But my dream, one thing I wish they would announce, is that next year re;Invent will not be the week after Thanksgiving, because it would be nice to enjoy it without dreading the week to follow. Because it's Cyber Monday. It's a big eTail holiday here. Good thing no one here runs a website that sells underpants or anything.
Ryan Donovan: If it breaks down on Cyber Monday, everybody's here, right? Everybody in AWS is here.
Corey Quinn: Surprisingly, no. Something on the order of a very small percentage of AWS folks actually come out here. And the AWS folks versus the Amazon folks tend to be a little bit different as far as areas of focus and the rest. And everyone talks about like, ‘oh, don't deploy on Fridays.’ Okay, fine, whatever. But also, there are change freezes at AWS around re;Invent, and Prime Day, and the other big events just because ‘he, he, we don't like seeing our name on the news in that context.’ If you enjoy my sense of humor, such as it is, and care about AWS, my condolences. You can follow my insane newsletter every week at lastweekinaws.com.
Ryan Donovan: And I of course, am Ryan Donovan. I host the podcast and edit the blog here at Stack Overflow. If you have comments, question, concerns, email me at podcast@stackoverflow.com, and reach out directly. You can find me on LinkedIn, and we'll talk to you next time.