Inside Amazon's AI video generator | Kabir Bedi

0:05Welcome to Dev Interrupted. I'm your host, Andrew Ziegler. And I'm your host, Ben Lloyd Pearson. This week, we're talking about Meta making the news for business practices and privacy violations, Cursor's own hiring practices, the joy of toy software, and how API sprawl is slowing you down. Ben, what's on your mind? Well, I can't pass up the opportunity to highlight all of the horrible privacy practices over at Meta, so maybe we can just start there. okay this is the story that broke the news in the last week talking about meta attempting to approach ai engineers from companies like open ai uh and doing so with huge bonuses and paychecks and uh sign-on agreements some of them topping a hundred million dollars could you imagine a compensation package from a competitor with that kind of price tag attached to it my mind literally cannot fathom what that kind of money feels like or looks like.

1:05Yeah. So obviously, Sam Altman and Zuckerberg, they took it to the news and there's been a whole news cycle about this. And Altman proudly said that none of his employees took that offer. They all turned it down and his employees believe in OpenAI. So it obviously became a posturing conversation between the two companies. But it's really interesting to see because only a handful of companies can really play at this level. I mean,$100 million sign-on bonus to attempt to poach competitor talent. What about the rest of us? There was a really good post that I saw on LinkedIn talking about this from Mohit Malik that said, real strength is not poaching.

1:43It's in growing, retaining, and respecting the people you already have. And so that was my positive takeaway of the story of, you know, there's always opportunities to take those$100 million poach deals and invest them internally on making your own engineers the best in the field. The thing is that Mark Zuckerberg in particular, I feel like he has a well-established history of using the throw money at it strategy to try to solve problems. So on one hand, it actually would not surprise me at all if this was true. If you look at Meta, they really were like an early leader in LM research and development.

2:20And they made a lot of great advancements in open source contributions to the cause. So I think they really do deserve some respect and recognition for their contributions in the field of AI. But it really does seem like they are struggling to build AI products that people actually want to use. We've seen Zuckerberg talking about AI social media bots that interact with you and become your friend. Or Meta's out with their new chat interface for AI that some people have been using. But I just feel like there's so many more interesting problems out there to solve with AI that social media just isn't – it's not really it, you know?

3:01So I kind of imagine that Zuckerberg feels a little bit behind in all the AI hype. But when I really dig into this, it feels like something's off. Like if this is true, it is truly a ridiculous sum of money. Truly. If it's true, it feels like an attempt to disrupt a competitor rather than any sort of real investment into AI. I personally would not trust this offer unless there was like some sort of substantial compensation that in the case that like meta revokes this offer suddenly because I feel like that's a real potential here. But, you know, honestly, I just don't really take the personal drama of random tech billionaires too seriously because at the time of this recording, Zuckerberg hasn't responded to any of this.

3:41So it's just the conjecture of a single person and Altman could be playing his own games here. So, but I will say this, if I were an open AI employee and I heard that my peers are being offered this kind of money, I would start wondering, hey, Altman, am I also worth that much? You know, you might turn up your nose at the Real Housewives of Silicon Valley, but I'm definitely going to stay tuned to the gossip and we'll let our listeners know if any new developments happen. Yeah, so I mentioned the meta privacy stuff. So let's dive into that, too, because let's just keep this meta hate train going a little bit.

4:14Well, you know, it's a nuanced hate train, but the next part of it is quite interesting because I think that if Meta convey$100 million for signing bonuses, they can probably handle whatever fine they're going to get for breaking GDPR and DSA and DMA and all the other alphabet soups of privacy enforcement out in the world. because meta there was a news broke another one this week about them about how they devised an ingenious system called local host tracking that bypassed even things like android sandbox protections that protect your identity while browsing an incognito mode an incognito mode is something that we all take for granted and obviously there's always a way for your identity to be compromised through while using incognito mode but how this one worked is basically your Facebook and your Instagram apps, they would set up like a hidden kind of like a intercom basically inside of your device.

5:06It was a port that would listen. And so when you would use incognito mode and you would go to websites that had a tracking pixel, that tracking pixel would actually reach out on your local device to that open port and then use that to identify you on Meta's server. So this effectively bypassed all of what incognito mode is intended to do because incognito mode, it protects you from yourself. It doesn't save anything on your own device. It works on your own device, right? But this works by actually sourcing Meta's tracking pixel from the content you navigate to. By any other definition, you would call this like a virus or a bypass, right?

5:43It's doing something that you're not, that is not visible to you on the surface. So that was really interesting development that came out this week about how some of the underpinnings of their own software work. I'm sure there's probably going to be more conversations about this massive privacy violation. Yeah, you know, this used to be a controversial take, but I think it's becoming less and less. You know, I deleted all meta apps from my phone more than a decade ago, like long before it was even called meta. And I never looked back, to be honest. They have a very long history of just abusing and violating privacy requirements.

6:16And even if they're not breaking the law, there are still definitely many examples like this where they are doing things that are at least somewhat hostile to their user base. So my advice is honestly, just don't trust them with access to something as sensitive as your mobile devices. And I do have a bit of a hot take on this. I actually have this theory that if you just install the apps from four different companies off of all of your devices, your life will get significantly better. So Meta is at the top of that list. I put ByteDance on this list as well. I think X deserves to be there. And I've got a forth you want to take a guess andrew what that might be so just so i understand these are the four apps that you should uninstall from your phone in order to have a better life that you should install uninstall all of their apps well i have to brainstorm for a second because i'm like what's the company behind merge mansion because definitely uninstalling that one is going to give you a lot of hours back at least if you're me but hmm knowing you ben definitely read it that's another one i think you would uh you know i've gone back and forth on that one a lot actually it is currently still on my phone though so oh okay so we don't know maybe that's even the secret fifth one we haven't gotten to what is it what is it ben pdd holdings what is that so they're the the makers of timu you know it's i feel like that kind of instant gratification style app like get something instantly whether it's like social validation or a piece of clothing from china like you know there's a lot of apps out there that are like this that met is not the only the company, I guess, is what I'm trying to make.

7:49They're just one of the more egregious violators. Yeah. And after you uninstall all of those apps, then you have plenty of room to install Roblox, which I recommend being on everyone's phone. Yeah, absolutely. But more importantly, I think this is yet another example of how U.S. regulations really have not kept up with technological advancements. We have this expectation of privacy in the physical space. It's built into a lot of laws here in the U.S. and even around the world in some places. But we haven't really adapted this concept to the digital space as well. So as a result, we rely on these U.S.-based companies primarily to become like the de facto policy enforcement organization.

8:28And not just for the U.S. Like it often, like their reach extends into other parts of the world as well. And without any sort of like adequate privacy protection in place, like companies like Meta get to decide how private data should be handled. And, you know, you mentioned that there are some potentially European data regulations that Meta has violated as well as some others. I'm sure this is not the last we're going to see of this story. But, you know, I just want to say the privacy researchers who discovered this like deserve kudos for this work. Like I really love when when people uncover this type of stuff and share it with the world.

9:03So great job to them. Yeah. So, all right, Andrew, let's talk about Cursor now. Yeah, so Cursor is a hiring playbook. It was discussed a bit in the news. I think it's just worth calling out for just a brief moment because as everyone knows, Cursor has been under like a magnifying lens for the rest of the industry. Everyone's been trying to figure out how this like ARR hyper unicorn grew from 100 million to 300 million ARR in just six months, right? So everyone's looking closely at fundamentals of a company like Cursor. So in the last week, we had a closer glimpse at how they hire their engineers.

9:37And it's a good look into how an AI-first company is looking at evaluating new talent. So on day one, these candidates are onboarded within the company, the engineering orientation around how they work. They're briefed on a project they'll be working on together with a staff member and introduced to the team's workflows, tools, and goals. So effectively, you come into shadow to work when you kick off your interview process. And they actively contribute towards brainstorming solutions, attending meetings, and they effectively collaborate as a full-time member by this end of the second day. And that entire process has a feedback loop where they're constantly talking with their hiring manager, with their potential future teammates about the process and what they're working on together.

10:23So it's actually really fascinating to think that just for interviewing and for bringing in new talent, that they have such an open door on how they work and what they're working on and when they're evaluating new talents. That was something that really stood out to me. The other thing that really stood out is that at no point in this conversation is there an arbitrary take-home exam or a coding, leak code test or any other evaluations. because these engineers, they do code while they're there, but they're more focused on evaluating their communication and problem solving skills. Yeah. And I find it interesting that they don't allow AI in any of their interviews, which is the type of thing that to me really feels like it only works for companies like AnySphere, which is the company behind Cursor.

11:08You know, they don't necessarily need to test candidates on AI skills because it should naturally come out within the interview process. And what I mean by that is, Andrew, you and I, we use AI a lot here at Dev Interrupted now. It's really unlocked a lot of new workflows for us. It's accelerated a lot of routine work. It's been just a huge change to how we get things done. And as a part of that, we've adapted new language, new practices, new tooling. We've become familiar with this whole new space of operating. So if we were going to interview somebody to join this team, for example, I think it would become pretty relevant to us or pretty obvious to us pretty quickly whether or not that candidate has, you know, sort of built AI into their lives because they would either have the same understanding in the language that we use or they wouldn't.

11:57And if they didn't, then, you know, it would be it would come out pretty easily without having to, like, ask them to use AI as a part of the interview, for example. Yeah. But what I really, really love about this is there's so many stories out there about AI taking over the hiring process. It's actually like really comforting to see an interview story that is distinctly human. Like they're focused on the human side of the experience rather than like, can you produce more code with AI kind of challenges? So it's pretty cool to see. Totally. All right. What's next for us, Andrew? Yeah, so I mentioned at the top of our new segment that writing toy software is a joy.

12:38And there was an amazing article I read last week from Joshua Barreto. As someone myself who was a hobbyist dev turned engineer, like many of these joyful game dev projects he listed out were things that I certainly tried to do or worked on as part of building my problem solving skills with code and experimenting with one-off projects. And so he gives a recap in this article, some of his favorite coding projects. And he gives a list of ones that he recommends any kind of hobbyist game developer or just code writer would try. And what's great about these is there's so many parallels in them to solving high level problems and software engineering that when you look at these steps, you really do see how the evolution of someone's skills can evolve through projects.

13:19And there's a few in here that were really fun to me that I've actually done before. Like some ones I'll call out were like a regex engine. Like I've written one of those before. A physics engine who hasn't if they've ever done game development. Your own text editor, definitely hacked around with that. A rasterizer, you know, I love working with shaders in the browser. And the last one that really made me laugh was an ECS framework, which for those that don't know is an entity component system. And really, that's something that a lot of hobbyists, game devs aspire to have is a perfect version of one of those systems.

13:54So great list. Highly recommend you all check it out. Yeah, it's pretty awesome. Yeah, I really love the personal project. I am going to disagree with the author on one key point that they make. I think somewhere in this, they make a point to not use AI for these projects. And I actually have the complete opposite perspective on that. Like almost relearning how to code because I've sort of been out of the practice for a little while just with the type of work that I've been focusing on. But with all these new agentic AI tools, I felt the need to get back into it and to start not really learning from scratch, but in many ways, like starting from, you know, not square one, but like square two or three in many cases.

14:33And, you know, I think the important thing is like, yeah, when you let cursor or copilot do everything for you, you're not going to learn anything from that. But it can also be an immense help when you are actively trying to learn. So, you know, I think my next experiment is actually going to be working with like cursor rules that help make cursor more of a guide that helps me solve problems rather than it writing all the code for me. Because I have really been focusing on like the fundamentals with this because I feel like that's important when, you know, we're adopting this new technology. Yeah.

15:06So what's our last story, Andrew? There was a new report that came out recently that actually it dropped today. and I've been looking at it. And really what's fascinating about this report from Apollo GraphQL is it looks at the state of API orchestration within organizations. And there were some really profound stats that I saw this morning when it came across my desk, which is why I included it in our discussion. Things like 76 % of enterprises can't make quick changes due to rigid infrastructure. Like, come on, that's something we talk about on here all the time. I don't think that's surprising at all.

15:37Yeah, less than 30 % can test their software without cross-team coordination. That's a reality within many teams. You know, you get these disjointed workflows. And then 52 % of... It's also a commonly cited bottleneck for companies that are adopting AI, right? Like they still have these testing workflows that get bogged down. Oh yeah, absolutely. Like if you generate a lot of code, then you're just create more throughput later on in your pipeline. You have to have the means to support that. We've talked about that extensively even in our webinar we had recently. Yeah. Another one here, too, was the 52 % of organizations struggle with large scale code changes.

16:14That's a reality. You know, large companies are having trouble doing this digital or rather this like AI transformation and modernization of their software, their infrastructure to handle how technology is being built now. So I really recommend everyone checking out the state of API orchestration report. I'm going to include it in our in our newsletter. And it's something that helpful to keep in mind that as your workloads scale, you know, this kind of thinking of orchestration. It's not just helpful, it's mandatory. There's a lot of research in here that highlights how you get these performance gaps if you don't prioritize it.

16:47So I know our audience loves these kind of data deep dives. We make them all the time. We read them and cover them here all the time. So if you want to read this report, we'll include it in the newsletter. And as a final tidbit here I'll leave you with is that we're actually bringing the CTO of Apollo on to DevInterrupted in the upcoming week. So stay around for Matt DeBurgulis. He's going to be coming up later this summer. So stay tuned for that one. Yeah, it'll be really great to learn from Matt. And what I really love about this is it actually aligns with a lot of the research we've been doing here at Dev Interrupted in partnership with Linear B.

17:20You know, as I mentioned, there's lots of bottlenecks outside of code production that AI really is not unblocking yet. And I think it's really important that engineering teams just don't rush off to generate more code without fully understanding the impact of it. you know we just published this six trends and ai driven software development guide we cover a lot of the biggest bottlenecks that teams are facing when driving ai adoption across the organization and they align quite well with what we're seeing in this report so our guide is also full of some really great advice on how to address it as well so check the show notes we'll we'll share a link down there yeah so andrew tell me about who our guest is this week yeah this week i'm sitting down with Kabir Betty at Amazon Ads, and we're taking a look at what video generation can mean for mom-and-pop businesses, how AI engineering culture works at AWS, and how Kabir's team built an AI product and brought it to market, while the fundamentals constantly kept evolving around them.

18:21AI is creeping into every part of the SDLC, but how far have teams really gone? Linear B surveyed over 400 devs, many of them like yourself in the Dev Interrupted audience, and found that 67 % are already using AI to write code. But is that creating opportunities or bottlenecks? Our new DevX guide breaks it all down, including adoption patterns, pitfalls, and the AI collaboration matrix that charts your own team's journey with AI. Check out the guide and maybe even the panel I hosted about it with Atlassian, AWS, and ThoughtWorks. I'll drop the link in the show notes. We're diving into how technology is reshaping video generation and how generative AI is a huge part of the evolution.

19:07We've seen some amazing developments recently, and today we're talking with Amazon. We're joined by Kabir Bedi, the head of product for image and video generation at Amazon Ads. And we're going to understand how generative AI is being used at Amazon and what it means for the future of generation. and also how engineering teams have come together to build this new technology that's opening up a whole new door to video creation for people just like you and me. So, Kavir, thanks for being here to dive into some of the advancements coming out of your team. Thank you so much, Andrew. It's great to be here.

19:44I'm very excited for our conversation and to discuss the latest AI launch that we just had, which is an enhanced video generator for brands to create videos. You know, video generation is a really hot topic right now. Everyone is kind of tackling this topic, both with their own technology and trying to use it in a consumer space. So there's a lot to explore and to learn from. And for the folks listening to Dev Interrupted, we're going to see a lot of parallels in this conversation to things in our own world in software engineering and how tools and things available to us are changing and democratizing the way that we work.

20:21So Kabir, if you were to give us like a big overview of some of that announcement, what were some of the biggest things that stood out to you? And what role did your team play in bringing those to fruition? Absolutely. So let me take a step back and first talk to you a little bit about what's the general problem that we were working to solve. You said it very, very rightly right now. Video is everywhere. Video is super important. we interact with video on pretty much most platforms that we all are on. Same goes for video ads. But the challenge that advertisers face is that video ads are really expensive to make.

21:00They take time, energy, effort, money. And so as a result, you have so many brands that were just not able to create video ads. So we set out on a mission to democratize video advertising so that all types of brands, whether you're a mom and pop shop that's selling one product on Amazon or a large brand selling hundreds of products, are able to use video ads. We launched Video Generator in beta last November and over the last six months have been receiving really good feedback from advertisers and just now launched an enhanced version of it with a bunch of new capabilities that I'm super excited to talk about that I think will really meet the need of those advertisers that have never been able to try video advertising before because it was too costly or too time consuming.

21:56And, you know, using the power of AI to really sort of bring their products to life in a way that they've never been able to do before. Yeah, this is like a great introduction to the problem. And I like that zoom in on it being accessible to, you know, more mom and pop shops and places that didn't traditionally dive into huge scale asset creation and having the photographer come out and set up all of the stock imagery backgrounds or to, you know, to buy all of those stock images online. And there's a huge amount of effort that goes into constructing creative works for that type of business. Not to make it more accessible to them, but without sacrificing quality, it allows them to kind of open the door on how they position their own products.

22:41And right out the gate, too, you pointed out something that stood out to me as a really big parallel is just making it accessible to as many folks as possible. You see the same thing happening right now in engineers that are adopting AI tools. It makes coding, like vibe coding, as you might call it, more accessible to more folks to try out to build things for the very first time. Absolutely. And I think that, you know, we had heard that consistent feedback when we launched in beta that we love that this allows us to create a video that we've never been able to make before. And our most recent data really validates that story.

23:20You know, what I've seen is that over half of the videos that are now getting created using video generator are for products that have never been advertised using video on Amazon before. And so right off the bat, we're seeing that advertisers or brands that have not been able to advertise their products using video are using this tool right away. and with this launch, we launched a bunch of new enhancements that really has sort of gotten folks excited. One of the pieces of feedback that we heard when we launched in beta was that back then, that technology was really cutting edge, was to show how products looked in a video.

24:03But the feedback that we were getting from brands was that actually, I want to go beyond that. I want the video to show how my products will be used instead. And so that's one of the big, big pillar innovations that we've just launched, which is we've added a lot more realism that now shows products in use. A great example could be previously we could show how a watch looked on, say, a table. Now we can show how that same watch looks on a human's hand, showing how the human moves their wrist to see the time. So previously showing things on a flat surface really worked well for certain categories like kitchen or say home.

24:45But for things like apparel or electronics, brands wanted to show how their products are going to be used. And so that's that's sort of one of the pillars of innovation that I'm super excited about. The other one is also, you know, as we've invested in innovated in new technology, we've been able to create multi-scene ads. So now you don't have to have just an ad that has a single scene. Now you could have three to four scenes that makes the ad a lot more comprehensive to show you really the product from all its different angles, all its different use cases. And so you're seeing this is sort of latest, I would say greatest in technology.

25:22Similarly, plenty of other innovations like the ability for advertisers to now upload an image and convert that into a video or upload a video and have us summarize that by actually identifying the key frames in it to create a video ad that suits that brand. So a bunch of new innovations have come out and we're getting very strong feedback that I'm very happy about. That's an interesting way of tackling the problem. And it actually brings me to one of my big questions I wanted to tackle in this conversation with you. So creating this tool, you know, and really changing the level set, right, in which folks can actually approach a video asset creation.

26:00But in doing so, you're getting a whole new generation of folks who have never used this tool before. So you have to think of all of the different ways they're going to interact with it. And this one thing that you hit off the bat is really fascinating. this like multi-modality of folks being able to input like a video or input an image or input like whatever they had, right? So what kind of assumptions did you have to throw away about how maybe your traditional audience use these tools, more tech savvy audiences, in order to make it something that would be massively adopted by other folks? Yes, yes.

26:34So one of the sort of assumptions that we worked backwards from was that we have to make it easy to use, which means how can we reduce the amount of input that's really needed from the user to give them a high quality output. So that's why one of the features of this product is that it's one click, which means that all the advertiser provides is what product they want to advertise. And we handle everything behind the scenes, which means prompting, generating the video, et cetera. And I was recently on an advertiser call where they really admired that. This was an advertiser that had been playing around with, you know, generally AI and concluded that they felt that AI required a lot of high quality prompting to get a high quality output.

27:22And he told me that he loved what he loved about this tool was that he didn't have to do any of that. He said that you've made it so simple for me that I don't have to think about what should be my input prompt? What should be the guardrails and the requirements for me to get a high quality output? And so that is, I would say, one of our biggest pillars that we really sort of wanted to hold true, which is how do you make it easy? Because a lot of the brands, smaller mom and pop brands haven't, like we already discussed, haven't created video before. They also have not with AI before. And so you want to really make it simple for them by actually reducing the amount of decisions that they have to make to also reduce their cognitive effort involved.