Host Intro (Gene Kim)

So anyone who has studied Amazon and/or Google probably knows the name, Steve Yegge. He notoriously gave the world an inside account of the famous Jeff Bezos "thou shalt communicate only by APIs" memo. This accidental public post led him to being on the front page of the Wall Street Journal.

I finally got to meet him a couple of weeks ago and he is as hilarious, irreverent, and as brilliant as you might expect. But we shared some stories on how AI is affecting work. And one particular theme emerged: AI seems to disproportionately help senior people in their careers regardless of profession. And this has some very real and consequential impacts on junior people in so many professions, including software.

One of the outcomes of that conversation was another epic Yegge post called "The Death of the Junior Developer," based on what his observations helping build a coding assistant at Sourcegraph.

I'm so excited that he'll be sharing not only what he's learned, but his perspectives on why he thinks Gen AI is so important. Here's Steve.

Steve Yegge

This is my mic on. Alright. Alright. Hey, thanks. Thanks for the kind words, Gene.

It's good to be back on stage. I haven't been on stage in 15 years — and it'll probably be another 15 years before I get invited back because I'm gonna say something that's gonna get me in trouble.

And I just finished my slides an hour ago.

It's also great to be back in the industry after a little hiatus during Covid. If you look at this gold box here, this represents 30 years of my career. And during that 30 years, those four companies — they all built their own software, they didn't buy. Yeah, probably most of you work for companies like that too. They make procurement really hard. And they like to build their own stuff because it scales or for whatever reason.

And so when I went to Sourcegraph, I went into SaaS and I was like, "Oh no, this is gonna be terrible, right? I gotta sell software." But it turns out to be absolutely amazing. Okay? Why? Because I get to go into all of your houses. I mean, not your literal houses, thank God, but your place as a business. And I get to see your architecture and your hopes and your dreams and your ambitions and your problems. And you all open up to me and it's pretty awesome.

I mean, I'm basically like the cable guy. Yeah. Just go in house to house. I even get barked at by security and legal teams. But it's really, really cool because people kind of open up to their cable guy — kinda like their hairdresser — and they tell me all the problems that they have, and I get to see patterns.

Oh my God, this is a very old version of my slide deck. A very old version. I'm just gonna — Anne, can you please fix my slide decks? Thank you. Alright.

So all of you are, unfortunately, telling me that this is harder this time around.

Now I know a lot of you have been doing this for a long time, right? 10 years, 15 years, 20 years in these leadership positions. And I'm hearing a consistent theme, which is that something's different, and you can't quite put your finger on it. Like, you hear CEOs talking about productivity loss and they all think it's RTO, and they're all frustrated 'cause that's not helping. And why, why is everything so hard?

I think there are some forces going on here. I'm gonna talk about 'em really briefly before we get to the fun stuff, 'cause I think y'all need to know this.

Oh my God, this is a very, very outdated version of my slide. I'm gonna have to wing it here. So what happened was, y'all got ZIRPed. Zero interest rate policy. Yeah. And a $2 trillion economic stimulus package. And what happened was, everybody had free money. And jobs, jobs, jobs. Software engineering jobs.

And in my old slide deck, I can try to find some stuff here. Here it is. This is what happened. You can actually see COVID starts, and then ZIRP happens in March 2020, and it starts to climb up. You can see when ChatGPT came out it peaks, and then ZIRP ends in March 2022, and it's tumbling down.

This had two really serious effects on our industry and on our engineering organizations. All of them.

One of them, strangely enough, from this horrific AI drawing, which is horrifically accurate — actually the first problem I actually got warned about by Walmart in 1999. Is that weird? That's weird. Yeah. Because what they said was — fully 10% of Amazon at the time was ex-Walmart senior executives. They got sued over this and all this. And they brought their culture. And their culture was basically one message. It was: do not allow your engineers to become entitled.

And we were like, "What?" Right? Because, like, at Amazon, for reasons I probably explained before, we were not very entitled. But that's because Bezos took his entire playbook from Walmart. And it wasn't until I went to Google in 2005 that I finally understood what entitlement meant. Because there are four engineers by engineers — they were princesses. Everyone got a pony. It was awesome. And I became one of the most entitled people there. I mean, I'm not proud of it. Anybody who knew me there, right? It happens. It's a real phenomenon. And it happened to the entire industry because of ZIRP, when the jobs were going up.

And then on the way down, the phenomenon that had started with the lockdown — I had PMs telling me, GPMs telling me, their PMs were going mad during lockdown because their sense of self-worth and validation came from in-person interactions. Their little glowing screen wasn't doing it. That phenomenon continued as the jobs started to plummet — wherever that slide is — and now anxiety set in.

And so those are the two gifts that Covid gave us. Oh my god, the slide's gone. But anyway, the two gifts are entitlement and anxiety. And you're all dealing with it. Many of you, not very well. Okay? You've still got people, teams telling you, "Don't scare the engineers. There's gonna be a wave of attrition."

I can tell you right now — they're gonna thank you. When you get your house in order, your engineers will thank you. Because engineers only love one thing more than everything else, right? They love to work together as a team to launch something amazing. And if you have a misalignment problem — which many of you do — hey, I'm the cable guy, I'm just echoing back what you've all told me, alright? If you have an alignment problem, your engineers are not happy fundamentally, and they will be happy as soon as you fix it.

So I just wanted to talk about that and put it in context for this. 'Cause you're all stuck at the innovator's dilemma.

All of you. Maybe some of you think, "Oh, Gen AI didn't really affect this much." But remember: in order to get disrupted, either your competitor has to build the same product as you for 10 times cheaper, or they have to make a product that's 10 times more accessible. Well, they can build it 10 times cheaper if they can figure out a way to get Gen AI to accelerate the developers. And they can make their product 10 times more accessible if they can use Gen AI in a clever way. Right? So there's two ways that any one of your competitors could disrupt you.

This is a serious problem, and I blogged about it in May. And because coding itself is changing — and this post was hilarious. It split people right down the middle. People were divided over this post. The post was a really simple premise. The premise was: within 18 to 24 months, all code is gonna be written by LLMs.

A lot of people didn't like that. They didn't like that premise. By the way, Gene helped me co-write this blog post. Gene Kim. And he's part of the reason I'm so confident about it, 'cause we researched the hell out of it before I posted it.

It's happening in multiple industries. LLMs are taking over the sort of creative work, right? The low-level creative work. And all you need is supervision. It's like a kitchen. Imagine, you're a cook, and you've just been elevated to a master chef, and you've got a bunch of robots that can do the prep cook and all the work, and you just have to turn it into a good meal. That's how it's changing. And that's why there's this focus shifting towards senior contributors. Because if you're a bad cook and I give you a bunch of robots, you're just gonna make a big bad meal, right? That, I mean, this is just common sense.

So people were calling me out of retirement. People who've retired, and they're like, "You're wrong." Right? Because the consequences would be inconvenient. That's why I'm wrong, right? Actually, it turned out a lot of them were calling me, they were angry because their kids just graduated with computer science degrees, and they're like, "You're poisoning the well, right?" Fine, I'm poisoning the well.

But this right here is how programming is changing. And many of you — most of you — are still stuck in like February. That's how fast this changes. Alright? You're probably thinking about completion acceptance rates, right? And token counting, and trying to compute what percentage of the code is written by AI versus people. A lot of you are doing that. It's dumb. Alright? It's ancient history. Completions are a gimmick.

The way coding has changed — and by the way, the people who called me to tell me, right, that I was wrong, none of them write code every day. Like, my title — you'll see on my badge — is engineer. I'm a grunt. I'm the cable guy. Yeah. The people who code every day are telling me this is happening, right? Chat-Oriented Programming. You do everything with the LLM. Yeah. It's an evolving technique. I know a lot of senior people who are using it. I don't know a lot of junior people who are using it. Why? It's the chef thing I just told you.

But there's a bunch of things that you can do with CHOP, right? With CHOP — oh my God, here, here's a picture of it. Alright? You're literally in the LLM, and you're just like, "Ah, this method's wrong," right? And it'll give you a bunch of suggestions and then it'll give you some code, and you've gotta find a way to get that back into your code base. It's an iterative dig and cycle. And you can go to different LLMs, and you can go to different coding assistants, and you can go to different sources. But eventually you get it to write the code, and you're five to ten times faster this way. It's counterintuitive 'cause you're spending a lot of time just thinking, right? But you have an infinitely fast generator, right?

So what is really changing here? Because people were really, really freaked out about this, by the way. I got calls from big companies. A few big companies where you would hear their name and you'd be like, "Oh, that's money." Money's not a problem for them. And they said that they have been witnessing this phenomenon unfolding in their engineering works. The pressure towards senior developers and senior contributors. And it's tough to be a junior developer right now.

There's a huge wave of tooling coming in. Your IDE doesn't even begin to scratch the surface of what you can do with Gen AI in code production. You'll probably hear from some of our friends at like GitHub about this. Like, the IDEs are actually somewhat limiting. The tools that are coming are all going to require LLMs to process your intellectual property.

Now, I had a wonderful, wonderful slide in here that unfortunately is not here anymore. It was a picture from Raiders of the Lost Ark — the very last scene where they're pushing the box into the giant warehouse of boxes. Alright? You remember? That's your intellectual property assets. I've been talking to some of you in the audience here and you're telling me that discovery is your hardest problem. You got thousands, tens of thousands of repos. You got wikis. You've got issues and project trackers. Nubank was ingesting what, 50 petabytes a day. All of that data is gonna need to be processed by LLMs. And almost none of you are ready for it.

In fact, some of the companies that are on the floor aren't doing any Gen AI. I'm not gonna name and shame — you know who you are. Okay? But there's a massive — as your cable guy — alright? I can tell you, there is a massive disparity between how the companies in this room and the companies in our industry are dealing with Gen AI. We've got the entire spectrum. I've met some companies that are like science fiction. They're out there. They've got massive Gen AI teams already and they're executing on 5- to 10-year plans. Okay? Seriously. I mean, they got their act together. And then I got other companies that are like, "Yeah, we had to push off our AI evaluation 'cause we had a migration, blah blah blah." Right? They're gonna die, right? This is serious. Coding is changing out from underneath us, and they're struggling trying to figure out whether they can even use it or not.

So what are you gonna do about it? First, get your house in order. Get your engineers aligned. Start using coding assistants. I know a lot of you have tried them in February. Alright? They weren't very good back then. Okay? CHOP — Chat-Oriented Programming — wasn't actually practical until ChatGPT-4o came out. And then the foundation models after that — Claude 3.5 Sonnet, Gemini 1.5 — now they're good enough, and they're gonna be getting better exponentially over the next two years. So we're gonna see more and more and more code get written. But I can tell you, CHOP is a lot easier with a coding assistant.

Yeah. You can do it with just GPT. Now, a lot of people were like, "Well, that's never gonna work for us because security, legal, blah blah blah. We don't want our code sent to LLMs, third parties." Yeah, I guarantee you self-hosted models are coming. It will be accessible to all of you in a way that makes your security and legal teams fine with it. Okay? So you might as well start prepping for it.

'Cause that Raiders of the Lost Ark picture that I didn't get to show you, right? You need to start getting ready for retrieval-augmented generation — RAG. All of you are now RAG shops because of this. This is the killer app for Gen AI, right? And these coding assistants are gonna come in and they're gonna start writing all of your code, and they're gonna start doing all kinds of stuff — reviews, right? — and auditing and all that cool stuff. All those — you basically need batch AI, and then you're off to the races.

Get a coding assistant, because it's all about context slinging, right? In the end, you have a bunch of data. The LLM can only attend a token window — that's yay big. I don't care if it's a million tokens. If you're taking in 50 petabytes a day, it ain't gonna fit in the context window. So you need an intelligent system that can go in and produce indexes of all of your stuff. Search indexes, knowledge graphs, LLM semantic indexes — all of that stuff. You need it, because otherwise your engineers are gonna ask questions and they're not gonna get the answers that they need because of your discoverability problem.

So, congratulations, you are all RAG folks now. And by the way, which coding assistant — I have not had my Cody hat on today. I'm working on a coding assistant and we really like it. There's a bunch — I, the box, the box sizes are meaningless. Cody does have a bigger box, but you know, it's meaningless. But seriously, any one of these is good. Okay? They're all good. Some of them are very slick and polished and they're turnkey experiences. And maybe that's what you need.

If you're the kind of company that likes to tinker — you know, you build your own CI/CD, you build your own developer infrastructure, you're taking your developer productivity into your own hands, you're very interested in Gen AI as a platform — then I would take a look at Cody, because that's what we're doing. We're doing platform plays and context slinging. That's been our focus, not agents. Agents are coming, right? And you probably hear about some of 'em today and tomorrow.

But CHOP — Chat-Oriented Programming — is what programmers are gonna do from now until the agents actually work. And actually, beyond when the agents start working, you need a human being to supervise it, to push it along, even with the agents, right? To verify it, spot-check it, all that. So CHOP's not gonna go away. Not for a good long time. So I would recommend using any of these.

Alright. Hey look, they finally got my slide deck from the last slide. So, "Help, I'm looking for." Yeah. So you need to pivot. Stop thinking about code completion acceptance rate. Stop thinking about completions, right? That really was a gimmick. What you need to do is start thinking about how you're gonna get — 'cause it's a — right, we're gonna do this together. We're, you know, we or whoever — whoever you work with when they index your code, it has to be a joint effort, because you all have really complicated systems if you hadn't noticed from the talks.

Yeah. So, you know, call your cable guy.

Hey folks, it's been a great time up here. I've been really happy to chat with you all. I look forward to meeting some of you afterwards. I hope that this was somewhat interesting and entertaining. Feel free to come up and grab me after and introduce yourself. Yeah, thanks all.