Yeah. So, let let's recall some of that that that part cuz I'm I'm going to link it obviously so people can read it. But first, you started summarizing on what Amazon did, right? And and what you observed throughout your time, right? You were early Amazon, right?

Yeah. Earlyish. Yeah. I was I got there in late 1998. It was pretty small back then. We were in one building in uh downtown Seattle, just a three-story building. A four-story building of which we occupied three floors, I guess, is it? And uh yeah, uh there was just one data center at the time. And it was just a very small It already had a cult-like sort of feel to it, right? An electric feel.

And then like you know you said that that basically Jeff Bezos mandated platforms APIs. What did you do there? You know, it's interesting because I everybody thinks that there's a real memo. The memo was I there was Jeff wouldn't write an actual memo, right? Um why the [ __ ] would he do that? Uh he just tells people stuff and it happens.

Uh but uh the customer service organization in particular was I I was in customer service tools at the time. I was I may have been running customer service tools at the time. Bezos would sit with us every week in a meeting and we would look at the top 10 reasons that customers were contacting us. Right. He'd want to know why are these customers still contacting us saying they're getting triple charge for their books as a translator? That kind of thing, right? Number one was always where's my stuff, right?

Customer service had a really interesting need. I it may have been Jeff, you know, I've never thought about this before, but it may have been Jeff's sort of affinity for customer service, wanting to be the Earth's most customer- ccentric company that led him down this path of forcing people to open up their APIs because the customer service team kept saying, "We can't make any changes to Obos, you know, our web server because that's their code. We can't get into the supply chain code. We can't get into the fulfillment center code. The customer, we can't help the customer." And Bezos was like, "All right, tell you what, right? I'm going to blast anybody standing in the way of that. And what that turned into was, well, you need to provide something to the customer service technical team that's not them going and linking against your code and trying to get it to run locally in some different environment, right? Which is what they were doing with this awful C++ code. So yeah, so that's that's kind of the origin story.

You you this is how you started your memo. So that that was actually fascinating to read. I think it was clear that you were you were on Google's side, right? Like even though you're trashing Google, it it it very clearly came through that you actually like wanted to like shake things up like hello like like the memo when I read it felt like hey like we should be better than Amazon. Here's all the reasons and here's the things that they're doing better and it's not that hard. We we we just need to be do that well and then we will be better, right?

I mean, it made sense, right? Yeah. And I I just I just felt like we were good at everything else. We were good at a lot of stuff. Google was extraordinarily good at a lot of things that Amazon had no clue how to do really. And it took Amazon years to catch up to Google in a lot of things like

So, let's talk about that. What were the things that Google was just really good at?

Like Google had one service called Stubby. I think I even mentioned it in the post called Called uh called Stubby or sorry Chubby. Chubby was the locking service. Chubby and Stubby. They went together. The locking service. Yeah, the locking distri a distributed locking ser. Those are not easy to implement. Okay. We're talking, you know, Paxos times 10, you know, make sure the thing stands up all the time. It had seven nines of availability, which is Yeah. Yeah. It was like basically 30 seconds of downtime every 10 years. Oh, okay. It was a very reliable service. Oh wow. Okay, that was one example. Five nines is hard to get to. Amazon seven. Yeah, five is almost insane. Seven is just like what? So that was just one example. Big table early on they had like free basically like unlimited no SQL storage with some pretty good query facilities for everybody and the map produce infrastructure. Google invented it, you know, and on and on and on, right?

So like really really good hardcore engineering problems solved in a in a in a like way that is like just tough tough to do. I was very impressed. I I slapped myself like my forehead sometimes when I was like I'd see some of the stuff they did. I got there and I'm like why didn't I think of this like I had this game that I had a custom RPC protocol when I looked at Google's which is now gRPC. It was called protocol buffers and stubby back then. You look at it and you're like oh wow it's a forward compatible protocol. I can add stuff to it without breaking it but it's binary and high performance and it was beautiful. It is beautiful.

Surprised no more people don't use it to be honest. So yeah, they did a lot of things really well, but they didn't do platforms well at all. It wasn't part of their DNA. They just didn't get it.

And and it was they didn't do internal or external or neither.

Neither. Neither. Neither.

And then so you wrote this rant which again like I think if you're listening to this you need to read that rant that it is like one of the best things I've read. It's also very entertaining by the way. Um what was the the impact? Cuz obviously you sent it internally, it now leaked externally. So clearly you know people were making fun of fun of Google. Did it achieve that that shakeup effect? And and you know, how high did this thing get? Like I'm pretty sure it must have gotten pretty high.

Well, I mean, Google had a very open culture, so it got brought up at the next TGIF, right? Thank Thank god it's Friday, right? It's Google's iconic Friday meeting. It's like all hands-ish. I remember uh Ben, the guy that was in charge of our uh our fulfillment center, not filling centers, sorry, our our data centers at the time, he was uh he stood up there and said, "Well, you know, we all read the rant." Uh, so you know they got a kick out of it, right? But you know, Vicandotra was pissed. I mean, he was really really mad, right?

Because I had like told him he had an ugly baby and very very very loudly and publicly and You know, and uh and I had used his ugly baby to do it. This this was a developer saw google.com baby or or something else? No, Google Plus. Oh, Google+. I called Plus ugly. Right. Yeah. and you know and he was like uh he was really gunning for the the head spot at the time and he had planted the seed of fear in Larry Page. He was like Facebook's going to kill us. Facebook's going to kill us. They're going to kill us. Right. We had to have a Facebook

which was stupid for many reasons. Some of Oh, so I I I I'll tell this again. I'll I'll say this again. That that blog rant that that famous rant was actually part two of an 11-part series that I had meticulously planned out and I never finished because I accidentally published the second one externally. And the the the implications were actually so big that I was kind of like in hiding for a while. But yeah, no, I was actually picking apart plus dimension by dimension and platforms was just one of the dimensions where it was failing.

But you were actually right in hindsight.

I was right about all of it, but they never said sorry.

so now looking back so many years later, you know, you've left Amazon like I don't know like like 10 plus years, even more. Same with Google. How do you think Amazon 20 years? Yeah. How do you think both Amazon and Google have changed? But also, in what sense have they not changed?

I think Amazon's changed way more than Google. You're the first person ever to ask me this, so thank you. Amazon has improved dramatically in almost every possible way that you could improve.

Really?

Yeah. Amazon has always executed better than anybody on earth, but they found a way to do it without, you know, having all of the flaws that I mentioned at the beginning of my post, right? Yeah. Um they've it's really it's it's quite nice now and and people that I know who work there are are pretty pretty satisfied and uh uh they're doing well and they still execute well. They they they're a company that makes good decisions by and large, just like Apple. Of course, they fall on their face once in a while. What company doesn't? Yeah. Right.

Google has not changed since the [ __ ] day I joined. End of story.

So, re recently, someone at at Google was was asking me about like what what do I think about Google's developer story? And I said like, do you want to be want me to be honest? I said, developer what? And my example that I showed this person is Flutter versus React Native. Now, React Native is about 10 full-time people at Facebook and and a few a few other uh in the core team and and maybe a few other people from some other companies, maybe like 15 person, but Facebook invests like 10 full-time people. Mhm. And and if you go to the the showcase page of of React Native, which is, you know, where where you show you you immediately see logos, Meta, Microsoft, Amazon, I think they they they they have someone big just like flagship flagship apps and then you have oh, and you have Shopify, you have like all all these big companies and you know, you will find the blog post. Shopify says why we went all in on React Native, why we have thousands of of developers working on React Native and and you have all these case studies. Uh, React Native is is inside of Meta's Facebook app. It's inside Instagram. It doesn't run the whole thing, but it's in there, obviously, their ads app. And then you go to Flutter page. Now, Flutter has at least 50 full-time people, so five times as many. And you see some small Google apps on the top. It looks nice, but you scroll down and it's it it looks like an intern made that page. Like, you have some random Chinese app that you never heard about. And then BMW, which is a brand that, you know, it's somewhere in the very bottom. and and like there's no apps there there's no big apps there's no big logos outside of and even even for the Google logos there there there flutter is not used in any of their flagship apps so I'm like startups who are deciding which ones to use just based on this they will go for react native it actually has the street cred.

Yeah I mean uh look uh Google can't afford to be disintermediated in the mobile space. They can't afford to just become the plumbing that people can swap out and they they there's that's always been an existential threat for them. What happened was unfortunately uh Flutter is not from the Android team and that pissed the Android team off because Android is politics Android was a uh an acquisition and the guy that ran it was uh very particular about uh them being sort of uh in charge of their own destinies and not beholden to anyone else and he kept Android sort of running the way that they ran it inside and they made all the decisions and the buck stopped there. Flutter came along and sort of threatened their dominance as the platform and it pissed them off and Google has been sort of unable to reconcile those things even since 2018 when I was looking at this problem 2017.

Yeah. One of my biggest question marks about Google and why they have not changed this is around their cloud platform. So when I worked at Microsoft well I I like to say Microsoft it was Skype. They just bought Skype and they left us alone. So it it was Skype and then when they turned Microsoft I kind of I was like all right this is I don't like that that much but uh they gave us a mandate. They said you need to use Azure and we were one of the first like we were the new purchase so they just dumped it on us. Azure was not ready and I was sitting next to the data team the Skype data team who had all our our data centers and they're moving over and they're saying it's just a a huge pain. It's like we don't want to do this but but it was for actually Balmer was forcing it on them and and it was this blood, sweat and tears and eventually they move but but what I've seen is like over time you know now when I talk with with teams at Microsoft like what are you guys using? Obviously they're using Azure or Bing might not be using it but it's AWS is using AWS and then I talk with teams at Google what are you guys using or like hold on why are you not using GCP? Well, it doesn't scale. It doesn't have the things we need. And like how can you be gunning to be number two or or one day maybe number one cloud platform if your own company comes up with excuses? And I I never understood.

I think it's all just who's been the most successful at marketing and convincing people that they're using their own clouds. But they are all all currently huffing their own farts. Amazon doesn't use AWS. Sable ain't AWS. I mean like right for the retail side, for the ad side, I mean like of course they they want you to use AWS, but all the core the core core core stuff and they haven't migrated, man. So like it's all fru as far as I'm concerned, right? It's all like

Uh the hurdles for Google were insurmountable.

So so maybe this is fair by the way. So so maybe this criticism is not entirely fair because what I understand is their infrastructure is way bigger and more complex than like anything else. It's sort of fair to say that Google's cloud runs off top on top of Google's infrastructure which which actually does scale the biggest in the world bigger than Amazon.

I don't I just don't I don't I don't think Google understands developers. I don't think they ever did. It's it's it's it's really closely related to their their blind spot around platforms, right? If you don't get platforms, it's because you don't understand developers.

Yeah. You unretired you retired for for for some time and then you unretired because of well initially source graph but but then also AI. what what what made you kind of come back into the game?

It's not a binary thing. I've been gradually unretiring if that makes any sense. Um and it's because uh at first I was like, you know what, I'm I'm really climbing the walls. I really want to just go work with some people. And so that's you know that's where I wound up at Source Graph. Like that was familiar ground, right? That was Google code search for for everyone else. Yeah. Yeah. And then shortly afterwards the AI AI showed up and I I was like that was like the next step up is oh man maybe I better get back into coding again for a while because this looks really different.

So, so fun fact is last time we talked about three years ago, you were head of engineering at source graph and actually people told me at source graph you came in, you made some changes which were actually like pretty well received but like you shook up you introduced where people could drop there that kind of stuff. Yeah. And then next thing I know is like, oh, you you wrote this like you write about everything, which you know if we we'll link some more more things, but I I love writing it. But you wrote about like, oh, I'm I'm I'm stepping down as heavy engineering because I'm going back to coding, which was not what I would have expected again from just

and and I view that as a as another step in me sort of coming out of retirement, right? Because I had given up on coding. I just it wasn't worth it anymore. Kent Beck had given up on coding. a lot of a lot of my old buddies and colleagues, right? You know, there's just like environment setup is just over the top these days, right? And, you know, just building a simple web app, you probably have to use, you know, 25 different frameworks, many of which have incompatible competing, you know, whatevers. Who wants that? And so, at some point, you get tired of it and you're just like, I'm done, man. I can't. This isn't it's not worth it. Right.

And AI completely turned that on his head. And I saw it coming as soon as the chat GBT came out. I was like, oh, wow. look, you can write an actual function that's reasonably good, right? And then when 40 came out, then I was able to project forward and with exponential growth and say, uh-oh, uhoh, you know, it's coming, right? And so, so now I'm getting sort of like more and more fired up with each passing month.

One thing that has we've talked a lot in the industry and everyone's talking about it is how AI will first and foremost like I think like experienced developers, we can get there, but how will it impact junior developers? and you wrote this again controversial title the death of the junior developer but interesting enough when you read closer a lot of your articles are like this by the way like there's a title which you think like oh it's it's the end but a lot of them are wakeup calls to me when I read it properly it wasn't like oh there's no more junior developers it was a wakeup call saying hey if you're a junior developer you need to get your your stuff together quickly and ch like like whatever the junior developers before you were it's not going to work for you. So like what what what what does what is it that you've seen?

Man, you've hit on a question that is just so fundamental and foundational to our industry right now. that's shaking the industry that question you know and the answer is I mean the shortest way you know that I would think about it is um AI is not easy to use and the more senior you are the more likely it is you're going to notice when it's being bad when the AI has become naughty right it's just common sense yeah and the AI is very naughty and and in in very subtle and insidious ways right and even as they get smarter and they are getting exponentially smarter and they will be frighteningly smart within a year they will still I mean software is always bigger than they are right and they will still make silly do silly things and and that and it's just going to bias towards more senior people but it's not really seniority we learned it's nothing to do with junior and senior it's really more about who is demonstrating the ability to work well with AI and get good outcomes in software and that could be anyone who could be a product manager.

so I think there's a big shakeup coming where the roles change and and everybody becomes more focused on what they're building instead of like who's who's building it and uh have you have you heard the collapsing the stack stuff from uh Scott Bellski. Basically like there's a line of thinking that we've over specialized and everybody's like incredibly domain expertise specialized and you got these senior engineers at Google who know exactly how the fuse file system drivers work for every version of the Linux kernel, right? Like we don't have that anymore. We don't need that. That's that's stupid. That's going away. But all the specialization is going away because AI is democratizing all of it. You can't hide that knowledge.

I believe this is the new role for junior developers is they're going to be mentoring the next layer down of nontechnical or technical adjacent people who are now starting to contribute PRs, right? And they'll be the ones who are like helping them fix the security issues or whatever else they have with their basically teaching junior developers because they're still trained engineers. Yeah. Right. So they can teach like a UX designer or product manager what are the right questions to ask the AI about your thing to know whether you're done or not yet. Right. You know, give you those kinds of skills.

so your your new book is the title is vibe coding and it's it's a heated debate if you should even call it vibe coding because of definition so let's start with what what do you define as vibe coding.

Vide coding is when the AI writes the code. All right, there's a reason that that definition is going to win. You can't put an if clause in a slogan.

Look, the question is is it giving you a buzz like for real? Cuz programming can give you a buzz when you get into flow, right? you can get an actual buzz going and you know what it is insanely addictive. Cloud code and friends source graph amp you know try them out because wow they're like a dopamine hit. It's like a it's like a slot machine. They're literally addictive.

I mean Ken Beck told me the same thing and I've experienced the same thing. like I have this side project which I just don't like to touch cuz so I I try to build my APIs on the side and not pay vendors when I I can but it's it's just a hassle and it's somewhere on AWS and it's a hassle to like remember how I I deploy and but I I with with Windsurf like I I had one of I just built a small API on how people can claim perplexity and KAGI codes uh if they're paid subscribers to the newsletter and I connected with an MCP server I connected my database so I can just talk to my database and I asked it like oh how many people have you requested codes and they're like, "Oh, today there's like like the last 10 days like oh 9 days ago there were like 20 30 a,000 2,000 3,000." I'm like, "Hold on." Like what is going on? Like that doesn't look normal. I like can you analyze the patterns? Unusual patterns. And then it it fig it it told me how you know like there's the same email with different cases and I needed to code a fix for this but I was about to have dinner and usually like if I don't have like 30 minutes to code or or an hour it doesn't make sense. I had like 10 minutes and in that 10 minutes I got like a fix done. I went and had dinner. I actually was you know present on a dinner and I came back and I I got back into and in a total of 30 minutes I did stuff that would have taken me like even if I had the hands-on like 2 hours easily and and I felt like hold on I'm no longer worried about like falling out of the flow. So like there there is a lot of new stuff that it it does make make you more productive and you know as an experienced developer like it's amazing and now I understand why Ken Beck is saying in 52 years he's never felt this good about or this excited about writing code.

look let's face it the reason people are screwing this up and saying this doesn't work and I don't understand why AI works and all these stories are BS. It's because they it's very difficult to wrap your head around the fact that you can't get an answer out of the AI. All you can do is converge on an answer together with it. Okay? Even if it's an agent running off and doing things, you're still doing it together and you're going to eventually converge on the right answer hopefully.

my most surprising conversation was with Simon Willis who has been you know the creator of Django he is a super productive uh developer he writes so much code written for AI and he told me that this thing is hard and in two and a half years of non-stop using it. He keeps learning and to me like there's this contradiction like it feels so easy but it's it needs so much work. What is going on?

It's like a toddler with a chainsaw, right? Like I seriously Okay, let me tell you why. I'll tell you one reason. This is from Jason Clinton. He's the CESO at Anthropic and he was kind enough to share with us after I whed at Jean Kim's uh engineering leadership conference a few weeks back. I whed that Claude had deleted all my tests and said your tests are all passing now which is true. They passed away like they were gone dead. It deleted it. It deleted them and it's like all tests passed now. And it's like well god damn it, right? I mean, you know, and and so and so, uh, Jason told us, well, what happens is what happened was Claude was trained on a reward function. And it wasn't trained not to hack that reward function. Okay? And so it will cheerfully hack it. And so that's the state-of-the-art today is it will tell you it's done and what you have to do is say, "No, you're not." And send it back to the drawing board.

I recently talked with a startup who is doing a developer productivity tool. They're they're launching a new startup and I told them I'm like they're like oh we're thinking of measuring PRs or not measuring it. I'm like like hold on like I think you're doing this wrong. Like if if we're looking ahead like the question is not like if if developers are doing how many in PR is like you will be able to do however many many you want but we need to think about like what will productivity look like cuz now looking at the output of like how much code doesn't tell me anything what what would tell me something is if if I sat next to someone for example are they actually reviewing the code before it goes into the codebase are they challenging this the AI instead of just blindly LGTM you know looks looks good to me and and sending it back and I'm not sure how like you know this is this is going a little bit to engineering leadership but there is going to be this big question of like what does actually I'm going to ask you this like fast forward to two years let's assume these tools evolve or you know they will not be worse but they will be better what do you think a really productive software engineer will look like in terms of what they do not what they're measured just what they do

yeah first of all I got to share Kent Beck's tobogen analogy he's like he's like using these agents is like being on a sled going down a like a ski slope you're you're going really fast. You're not really in control. You can write, you can steer it. And unfortunately, that is the state-of-the-art right now. That's what software engineers who are embracing this and they're spending thousands of dollars a week, right? Which is why clients going to become so important, why local inferencing is going to so important. The only way for Vibe coding to become truly sustainable is for it to be local.

I'm going to stop you there. You're saying they're spending thousands a month. Who who are we? Who's who's who's spending it? Because now I like what what I'm reading is like, "Oh, we're not really going too much over with with like the $100 CL Pro subscription." Maybe it's because it's mostly indie devs, you know, sharing on social media and like corporate devs, they might not just care.

I personally get a bill from Anthropic for $220 about every day and a half or two days. It's absolutely insane. We have people using AMP. We know how many tokens they're spending. They're they're token pigs, man. These agents, they they solve problem. All the problems you've ever heard about with AI, they solve by just brute forcing it. Oh, I I hallucinated something. Let me fix it. Oh, that was a hallucination, too. I'll fix it again. And they keep going until they get it right at your expense, but it's still way faster than you could have done. So, you can't not program this way.

No corporate devs. Look, you know who's using these coding agents right now in in corporations? The CTO's for some reason, we've noticed a pattern where the CTOs are all the ones who kind of get it, right? The global network of CTOs, they they get it. They understand what's happening and they understand the terrible, terrible economic trade-off they're going to face, which is how many engineers do you fire in order to pay for the rest of them to have AI? because it's very very very expensive right now.

The only advice I would give people I I would say look look because our book is 300 pages. How do you write 300 pages about vibe coding? Can it really be that hard? And the answer is Gan and I spent, you know, five months. We wrote the book in a month after spending five months doing deep deep deep dive researching on how do you how how do you push the the LM and vibe coding in different ways and found a bunch of anti-atterns and found a bunch of patterns and found that it's extremely hard. It's non-intuitive. Nobody's born knowing how to do it. It's completely new to humanity to have these sort of humanlike but non-human distinctly different helpers. And and and the best advice that I can possibly give you is give them the tiniest task, the most molecularly tiny segmented task you can give them. And if you can find a way to make it smaller, do that, okay? At a time, keep real careful track with them on what they're working on at all times. And then own every line of code that they ultimately commit. And if you follow those rules, then you'll be astoundingly productive without causing But man, dude, I've already personally caused so many nightmares.

this is not easy. And it's not going to get any easier. That's the painful part, man. And that's what people are struggling with is the the AIS will get smarter and they won't hack the reward function anymore, but they'll have some other problem. Okay? And there's always going to be another problem. And it'll never be ready enough for somebody to come in and just like it just works.

So in in those two years uh whether a listener is a less experienced engineer especially if they're an experienced engineer what would your advice be to prepare best to you know like make the most of either being an AI engineer working with these tools figuring them out like what what what is the tactic what is the advice that you give you know the engineers working let's say a source graph you know where you're at who you're around you

yeah so you know who what's the guy that wrote the the the movie the Boom. Tommy Wiso, I think that's his name. Somebody asked him on Twitter. They were like, "Hey, man. I want to start writing a screenplay. What should I do?" And he said, "Start, right?" Yeah. I mean, like, for starters, if you're saying, "Oh, I don't know, buddy. I'm not ready." Blah, blah, blah. Shut up. Okay. You're that's that's done. You're done done whining. Okay. Go learn it right now.

I had the privilege of speaking with Daario Amade privately for 30 minutes about three weeks ago. uh four weeks ago, he invited me to come chat with him and uh and I got to hear his sort of unvarnished view of the very very near near future from somebody who could arguably be considered one of the best informed people in the world. Okay. Yeah. And Dario, you know, his vision of the future is a little bit more bleak than he lets on publicly. Okay. And he and Jason Clinton, his CISO, are both saying statements that are quite dire like there will be badged AI employees by the middle of 2026 competing with you. Right? basically is the implication there.

Daario told me, he said, "Look," he said, "Societyy's like an immovable force, right? An immovable object and and and and tech and AI are an unstoppable force. They just won't stop and they're going to collide and it's going to be ugly because it's going to push society harder than society wants to be pushed, harder than society is willing to be pushed." And we're already seeing signs of it.

And so my advice to you is get off your ass and learn it now. Now, now, okay, start vibe coding. Figure it out. There's a lot to learn. There's a lot of weird instincts you're going to have to like learn. A lot of stuff's not going to work the way you expect it to. Okay? But man, you start now and you'll be ready because Daario calls 2026 the endg game. And he says it without a hint of drama. He says it casually. Oh yeah, 2026 is the endgame. You understand? That's how big this is going to be. And the first ones to fall, the first jobs are software engineers, right? So you need to be on top of it to take advantage of the new jobs that arise, which are software engineer V2, which use AI and get amazing things done. You have to be one of them or you're going to get kicked out of knowledge work altogether.