Quinn Slack: [00:38:11.25] Yeah, Cody does make it so much easier. And yeah, going back two years ago, we had a fork in the road. We could have either made just code search, something that clicked with so many more developers, and overcome that kind of question which is “You know, I’ve been coding for 10 years. I haven’t had code search. I have it in my editor. Why would I need to search across multiple repositories? Why would I need to look through different branches? Why would I need kind of global [unintelligible 00:38:39.15] definition? Why would I need regex search that works?” We got a lot of questions like that. We could have just doubled down on that and tried to get, for us, way more devs using it for open source code, and within our customers 100% of every developer, and all of our customers using code search. We could have done that. What we decided to do was go deeper into the intelligence, to build things that were exposed as more power user tools, like the code insights. Code Insights is something that platform teams, that architects, and security teams, managers - they love, it has incredible value for them, but for the average application engineer they’re not really looking at code insights, because they’re not planning these big, codebase-wide refactors. Same with batch changes. Platform teams love it, people that have to think in terms of the entire codebase, rather than just their feature, they need it. And I think we got lucky, because given that right around that time, that’s when developer hiring began to really slow down. It was really helpful for us to get some really deep footholds in these critical decision-makers, just from a sales point of view, in companies, to have like very deep value, instead of kind of broad, diffused value.

So that ended up being right. It also ended up being right in another way, which is we got deeper in terms of what does Sourcegraph know about your codebase? And that was valuable for those humans over the last couple of years, but it’s also incredibly valuable now, because we have that kind of context that can make our code AI smarter. But I do really lament that most devs are not using code search today. I think it’s something that would make them much better developers, and there’s absolutely a part of me that wishes I could just go have 50 amazing engineers here work on just making it so that code search was so damn easy to use, and solved every developer’s problem. Now we’re tackling that with Cody, because we’ve got to stay focused… And to your point, they do solve the same problem. And with code search, if you’re trying to find out “How do I do this thing in code?”, code search will help you find how all of your other colleagues did it. Cody will just look at all those examples and then synthesize the code for you. And so there’s so much similarity… And we are just finding that Cody is so much easier to sell.

But we did have a cautionary moment that I think a lot of other companies did. Back in February to May of 2023 this year, if you said AI, if you said “Our product has AI”, literally everyone would fall over wanting to talk to you, and they’d say “My CEO has given me a directive that we must buy AI. We have this big budget, and security is done, legal is done, we have no concerns. We want it as soon as possible.” And it didn’t matter if the product wasn’t actually good. People just wanted AI. And that I think created a lot of distortions in the market. I think a lot of product teams were misled by that. I’m not saying that the customers did anything wrong. I think we were all in this incredible excitement. And we realized that we didn’t want to get carried away with that. We wanted to do the more boring work, the work of “Take the metric of accuracy, and DAUs, and engagement, and overall a lovable product, and just focus on that.” We did not want to go and be spinning up the hype.

[00:42:04.06] So we actually really pulled back some of this stuff and we level-set with some customers that we felt wanted something that nobody could deliver. And that was one of the ways that we came up with these levels of code AI taking inspiration from self-driving cars. We didn’t want the hype to make it so that a year from now everyone would become disillusioned with the entire space. So definitely a big learning moment for us. And if there’s an AI company out there that is not looking at those key user metrics that have always mattered, the DAU, the engagement, the retention, the quality, then you’re gonna be in for a rude awakening at some point, because exploratory budgets from customers will dry up.

Adam Stacoviak: Well, just think about it now. Right now we have not so much no eyeballs, but very little eyeballs, because the process to have such knowledge is difficult… Whereas with this partnership, this data set available on BigQuery and the good stuff, now people have a much easier way to find these insights, and then obviously knowledge is power, so in this case I’m on Felipe’s side, Jerod. I am kind of - not a naysayer, so to speak… I’m like, “Do it!” because I think about… In this show that’s gonna come out after this we talk to Peter Hedenskog of Wikipedia about site speed, and we talked a lot about automating reporting of site performance, and this is similar to your point, Jerod, where you said “Could we automate some things where pull request is automatically opened up?” I think about the automation tools that may be able to take place on the security side to say, “Okay, here is a vulnerability.” [32:02] It also opens up another topic I wanna bring up, which is not just the GitHub data store but other data stores, or code stores like Bitbucket or GitLab, having similar data sets on BigQuery and how that might open up insights to all stores, into all the major stores. But long story short, automating those kinds of things to the open source out there, that’s an interesting topic to me.

Bridget Todd: Yeah. And just to add on to that, when we were first working on the podcast scripts, something that Solana said that will really always stick with me is - when we’re talking about “Oh, how do you phrase certain things as it pertains to technology or AI?” and she said, “Oh, we don’t like to say AI does this, or the technology does that”, because that’s not actually true. It’s people who are programming AI to do this, or people who are programming technology to do this. And that really blew my mind, the way that I had to sort of believed this idea that “Oh, the technology is going to do what the technology is going to do. it’s this mystical robot that I have no insight into.” How that really obscured the humans with power to make decisions about AI. It really obscured their role in a way that I think really benefited them. And so I just really agree with all of the points that Solana just made.

Nadia Odunayo: So they didn’t come out. So I said “They’ll come out in the first week of January.” We had a wild January. Two weeks of downtime, we couldn’t handle the demand… There’s a whole thing there. We had to rearchitect a bunch of things, and we still have some plans for rearchitecting how our database works. I don’t think we’ll have time to get into that in this conversation, but… So I have only recently been able to pick up book clubs again.

I think in one of the ideas that I was throwing around there was a plan that – essentially, I felt that Plus users would be able to host book clubs. And my thinking was that – I didn’t think you could have a model where every member would pay for a book club, because people can run book clubs for free on several other platforms. But I thought “Oh, if we have it such that Plus members could host book clubs, that would be great.” However, my customer research revealed to me a few things. And this is customer research of people who use Plus, customer research of people who have book clubs… All these things. And what I learned was Plus is not working. We have product-market fit with the free product, we do not have product-market fit at the moment with Plus. Why? Because today, two years later, the Plus plan is still that grab bag of features that I made up. And so what happens is most people who take up the Plus plan, it’s because they want to just support an independent alternative to Goodreads… And they probably have one of the eight features that they use. And most of them gravitate towards the stats features.

[50:28] And so I did this whole customer research and I said “We need to completely change how the Plus plan looks.” So I’m currently in – I say currently… It’s been on pause for a while, but I should be picking it up in February… I’m currently in the process of completely changing the Plus plan to be an advanced stats plan. So all it’s going to have is going to be just all the extra charts; you’re going to be able to create your own charts… Because I discovered through my research that the power users are the ones that care about the stats. They have their own Excel spreadsheets.

Separate to that, I’ve been doing customer research with book club people, but also just thinking about the vision for the product and what works well. And it’s far better, because there are a bunch of features within the Plus plan right now around buddy-read book suggestions, where you can put in the user names of people that you want to read a book with, and we have a machine learning thing that will tell you the best books to read for this group, which - obviously that will be included in our book club feature. Those features, in terms of like network effects - it’s better that those are free, and that actually more people can use them and come on board… And then maybe the power users in the group, or the ones that care about stats, might upgrade their accounts to Plus… But now, the group – and no one’s stopped from joining. Because the other thing I realized is that with that Plus plan, someone might be into stats; someone else is into running a book club. Someone else is into running a buddy read, or whatever it is. So I definitely need to separate out the identity – the identity of the Plus user is going to be the advanced stats power user, and then book clubs, buddy reads, read-alongs, all our social features, you can opt in; it’s all free. If you’re a Plus user, you will get advanced extra stats, insights into any of the things you do. That’s going to be the core of Plus, but it will be free to run your groups. And I had to change the estimate to end of January, but it’s 31st of January. There’s a reason I chose to do this interview on the 31st of January in my evening. I thought by this point I have everything done that I wanted to get done. It’s still not done…!

Solana Larsen: Yeah, sure. Yeah, what a year. What a week. What a month. I feel that way all the time. It’s just going so quickly. Yeah, I mean, I guess people who might be familiar with Mozilla might know Firefox, and think “What does that have to do with AI? Why are you talking about that?” The Mozilla Foundation, the nonprofit arm of Mozilla, for the past couple of years has been really focused on what we call trustworthy AI, the term that you just used. And I guess you have a buffet of choices when you’re trying to talk about ethical AI, or fair AI, or equitable AI… Trustworthy was the one that we went with as an organization. It checks a lot of the boxes, especially at the time that we chose it… And it resonates in policy circles in particular, in some of the contexts that we like to be in.

So there’s this element of “What is the future of the internet?” And if we’re an organization that has a manifesto that cares all about making the internet healthier, creating tools that enable people to create and be a part of the internet, if AI is the future of that, then what’s our role there, and how can we help make sure that all the mistakes that happened for the Internet in all these 20, 25, 30 years aren’t just all being repeated again in AI.

And so we’re thinking about privacy, of course, but there’s also consolidation of power… You know, the way that big tech kind of take control of everything, and squeezes out opportunities for smaller players… And a whole bunch of other things.

[06:04] So we have a whole foundation of people, we have fellowships, we have grants that we make for different people who are innovating, and trying to think about AI in a way that big AI tech companies aren’t thinking about them, just so that we can have an alternative kind of conversation happening around these things that are changing our society, and our industry, and everything so quickly.

So part of this is we do have a big mouthpiece, I guess, as Mozilla, and one of those platforms that we have is this podcast, IRL, that the Firefox team started years ago, and that we took over on Mozilla Insights Team a couple of years ago, and we’re now doing a season seven. It’s the second season that we’ve dedicated entirely to AI. And part of that exercise is thinking about “Well, who do we want to lend the microphone to? Who do we give the microphone to? What kind of voices would we like to have in this dialogue about AI that we don’t maybe get to hear as often when we’re tuning into the US mainstream tech press?”

Beyang Liu: So the way we think about it is we’ve spent the past 10 years building the world’s most advanced code understanding tool. So we have the best code search, we have the best code graph… So the global reference graph, across all the different languages in the world. We have a large-scale code modification refactoring system, and a system to track high-level insights…

[16:12] So there’s all these backend capabilities that are really, really powerful. And what language models have done is given us a really, really nice, beginner-friendly interface to all that power. And I think you’re gonna see this across all kinds of software. Historically, building power user tools has been difficult, because the on-ramp to taking full advantage of those tools - it’s been a little steep.

Chris Benson: So I’ll speculate, as we’re kind of winding up a little bit, to kind of bring it back to how is power shifting at the corporate or geopolitical level, and the role of AI - all these capabilities that we’ve been talking about over recent episodes… Those who have the creative insights, where they can take advantage of these capabilities, and see opportunities… But the thing that humans still have right now is you have a form of very limited creativity in AI. In other words, it’s not self-aware, but you can create those raccoon on rocket pictures now, that we were talking about… Which is pretty cool, but it’s not sentient, and it’s not self aware, and it doesn’t have a special understanding of the overall world at all the different scope levels that we have.

Adam Stacoviak: …you have less detail, and you pinch and zoom in to get more detail of that space, because between those two is your opportunity to choose. And we’ve talked before about the power of choice. It seems easy just to, I suppose, maybe pontificate the power of choice… But in a practical way, in a very loving way, your choice is your super-power, and you have the ability to maybe have these insights, like you did. And as you were sharing your story, I had some (I suppose) shared similarity in my past with yours, where I’ve been in moments where I was abusing my life, so to speak, or my body, or my brain, and clubbing or doing different things in my life, where I was not the Adam that is here today talking with you… And I had those same kind of moments, like “Hey, what are you doing? This is not you, this is not your life. This is not what you want.”

And we so too often talk on this show about understanding what you want, what motivates you, what are you driving towards, what’s your purpose… Not so much just purpose, but what is your core motivation? “What are you optimizing for?” is what we say a lot. And I think that’s what people miss - that awareness.

[12:19] And what comes with that is understanding who you are, what you wanna be, and I suppose - to use back your words - who you can surround yourself with or the environment you can put yourself in to get there. And that’s what that space is. Pinch and zoom.

Nikita Shamgunov: Definitely. So let’s step back for a second and think about what different parts of the world, and different companies and governments, what do they fundamentally want to accomplish as we go through the pandemic? The first one is simple - how do we stop the spread of the virus? Okay, well maybe we cannot really stop it, or let’s say we put our actions and efforts to do that… But since it’s spreading and it’s a matter of fact, what else can we do and how we can drive our decisions based on data? What kind of decisions? Well, it could be capacity planning for ventilators. We know there’s an outbreak there, and we will likely have our healthcare system overrun, and we need to provide extra-capacity to the healthcare system… But how much capacity?

All of those questions require answers, and the answers are in data. That’s where data science comes in, and that’s where just starting from collecting the data, putting it in one place, organizing the data and feeding this information to people who have the levers of power.

The second one is “Who wants the data?” We have obviously Apple and Google, who own the data because they have a device. Every individual on this planet - not every, but most of them - are now on the smartphone, so you can tap into that stream of data and get information about who is at which location at any point in time, and then marry that location with migration patterns, and marry that location with individual tracing.

Given that we know that this person has Covid-19, who are all the people that this person came across in the past two weeks, so we can go reach out to them and say “Hey, you probably want to be tested.”

The second entity that has that data - maybe government, but I don’t know about that… But certainly telcos. Telcos have this information; maybe not as accurate, because they don’t have a GPS on the device… Actually, they do GPS on the device, but they may not be able to tap into the GPS. But they can triangulate the location based on cell towers.

So we’re working with some of the largest telecommunication operators here in the United States, as well as around the world, and I think the one that’s public is TrueDigital, one of the largest telcos in South-East Asia. And we do both; we do the migration patterns, where if you go back to March, February timeframe, we already knew that there was an outbreak in China, and there was an outbreak in Italy, and we already knew how bad that was… And looking at the flights from Italy, and tracing individuals that land, and then starting to see this pattern of people getting sick emerge, you can start driving decisions off of this. You can start putting policies in place that can stop the spread, you can start do capacity planning, you can start manufacturing masks and ventilators and distribute them into places based on the patterns that we’re observing.

[40:10] So that’s how data management solutions are helpful to companies that have the data, and also the insights that those systems generate are useful for people with the levers of power to drive policy and to drive decisions. Google and Apple - especially Google - have the technology, but telcos don’t, and that’s where we partner and give them those abilities.

Chandler McCann: It was very cool, and it was just an awesome experience from both a data science perspective. Where I’m working over here, I’m pulling data out of a table, to actually going on the ground and meeting with the people who are collecting it, and having conversations with them and trying to communicate the power of machine learning and the importance of the data that they were collecting, and how it could be used… It was just a humbling and awesome experience, all wrapped up into one.

So following these two trips, we had very positive relationships with the government of Sierra Leone. In 2019 the ministry of water kind of reaffirmed their commitment to evidence-based decision, and actually passed a national policy requiring the use of data in decisions about water services… Which is pretty cool, and also, again, humbling. If you think about it, I think the story of this project is the story of the power of data, and what it can do… And if you think about it, in 2018, the use of data at the national level in Sierra Leone to inform decision-making by policymakers, again, with very constrained budgets, was very low. And then in 2019, we do know that working with the ministry of water, they were able to use some of the insights from our tool to inform decision-making and budgeting that year. So that was definitely exciting and a near-term win for us on the project, and that helped shift things out of this R&D phase to where Brian and the Global Water Challenge are in the process of pursuing more funding. If you wanna contribute to them, feel free.

We’re looking to build a much more sustainable tool, that can be deployed to many countries around the world. So it kind of was a launching point for the project.

Jerod Santo: Yeah. That sounds really powerful, thanks for explaining that. It wasn’t connecting, but now I totally understand what you’re saying, and that sounds like – especially with power tools, they need to be malleable. If you can dream it, you should be able to do it, especially for developers, so… That’s really cool.

Back to the collaborative bit - I mean, I guess it kind of seems obvious, but I don’t think about these things so explicitly all the time. Collaborative tools are inherently viral, because you have to share them with the person that you’re collaborating with, and if they haven’t learned the tool yet, then they’re being exposed to it, right? Like “Here. Here’s a Postman collection.” “Oh, what’s Postman?” “It’s this thing.” They have to basically spread it in order to do what they need to do for their job, to collaborate… Unless they’re like me and they work in a silo.

So I guess the takeaway there is if you’re building collaborative tools, and the sharing/collaborative side is hard, then you’re not doing yourself any favors whatsoever. You mentioned the collections - was that a thing you formalized around, and made it really easy to just share a link and pass it on? Was there anything that you did around the landing page for the person who’s not a Postman user, who gets a collection passed to them? Share some insights on how you guys really smoothed that out and made it shareable.

Chris Benson: [15:50] So I ran across an article this past week from MIT - and it was actually on news.mit.edu - and they have announced that they are now going to build the new Stephen A. Schwarzman College of Computing. Initially, just hearing that, you may say “Okay, well, that’s what MIT does - they do computing”, but this is a little bit of a different approach to setting up a new college. They received a 350 million dollar foundational gift from Mr. Schwarzman, who happens to be the chairman and CEO of BlackStone. Then on top of that, they have what amounts to a one billion dollar commitment.

What they’re trying to do here with this new college of computing is to build an interdisciplinary school around AI and the various fields that it touches on, to try to kind of drive things into where we’re seeing the future go as we’re living it right now. The AI world has really revolutionized education in that space, and some of the things that MIT notes is that they want to reorient - and I’m reading from the article when I say this - MIT to bring the power of computing and AI to all fields of study at MIT, allow the future of computing and AI to be shaped by the insights from all disciplines. They have 50 new faculty positions to support it, there’s a shared structure with other schools that they’re doing, and they’re really looking at trying to produce students that can operate in this interdisciplinary approach.

Personally, I think that’s fantastic, because in my own experience, AI touches on other fields almost every time you use it. It’s never a standalone thing by itself, it always intersects other areas that you’re using it on. What do you think, Daniel?

Justin Garrison: Well, I mean, at the time too – we’re looking back to early 2000s. That was not an uncommon thing, because there was less clear defined lines between user space and kernel. It’s like “Well, a lot of things run in the kernel that shouldn’t.”

The other interesting thing that Brendan Gregg, who’s over at IBM now - he had really great insights into eBPF, the enhanced Berkeley Packet Filtering as a system would have made this safe. And eBPF runs at kernel space, but makes the code safe so it can’t crash the kernel if the eBPF program goes down… Which I thought was interesting. Because a lot of antivirus companies, especially for the Linux world, are moving to eBPF, and Microsoft and Windows is getting eBPF support at the kernel level.

And so eBPF is going to be your one program to write for the kernel space, and it works on Linux and Windows. And it’s also safe, where you get a little more elevated privileges to do things at a lower level, without making the whole system crash.

Bengsoon Chuah: Yeah, that’s a good point, because a lot of these traditional industries, at least in the energy sector that we see, we have sensors that are constantly flowing in data all the time. The data is there. We are collecting data. And then going a little bit deeper, then you find that a lot of us have been collecting a lot of unstructured data, too. So in my experience, at least what was happening was when I came in, I was pretty much the only data scientist, and so I had to make my existence justified, in a sense… So I knew that I had to do something about bringing value within this organization, and I’ll be able to have proof that “Hey, look, data science actually does work, and does bring value.” But the quickest, I guess low-hanging fruit that I’ve found, at least in the context of where I’m at, is the whole unstructured data.

[00:08:07.24] So we have been collecting thousands and thousands, if not hundreds of thousands of unstructured data, but there has never been a way to really analyze that at scale. So people have been analyzing it, they’ve been able to do some sort of a human analysis on it, but there’s never been someone who’s able to say “Hey, what’s been happening in the past 10 years that we’ve been collecting all this data? What’s it telling you?” Nobody has really been able to do that. So I thought, okay, maybe that could be one thing that we could actually bring in, is that you have all this data that’s ready for you, that has all the insight, that has all the information that is locked up, waiting to be unearthed, pretty much, waiting there for us to just extract it or mine it.

So I did a quick POC for one of the departments, a company, and said “Hey, guys, you’ve been doing Power BI, Tableau, Power BI kind of thing… You’ve been able to do stuff with your structured data, and you’ve been able to plot it on beautiful graphs, and stuff like that. You’ve been able to analyze it in that sense. What about all the unstructured data that you’ve been collecting over the years?” And they said “There’s no way you could do it, right?” I mean, you’ve got [unintelligible 00:09:20.26] There’s no possible way for you to put it in Power BI.” And so I said, “Well, maybe we could explore ways that we could actually get machine learning to actually help you to scale that analysis from an NLP standpoint.” Thankfully, they bought it, and we took off from there. It was pretty cool. I mean, it was a journey, for sure; a journey to learn.

Jerod Santo: What are some insights that I would gain as a developer, looking at your dashboard or whatever it is, your reporting tools, in order to observe? What might I find? I assume this API is lower than you thought it would be… It seems like an obvious one. But what else?

Tuhin Srivastava: [00:05:45.08] Yeah, yeah. And I think the last piece is just around – you know, the ChatGPT moment for AI was interesting, for a number of reasons. I personally think as someone who built infrastructure that is most interesting because - if you want to call that the iPhone moment of AI… I think it’s a bit different to that, because it’s so early in the journey. It’s like if the iPhone showed up when we were all using 5110 from Nokia. The world would be very, very different, because consumers and developers, their first taste of machine learning and AI was through ChatGPT and GPT APIs. The stakes are just – it’s just harder to build something good. People don’t want to use a model that takes 12 seconds to run. High-speed production inference is taken for granted when you’re using [unintelligible 00:06:34.09] and then when you kind of combine that with “Okay, open source models need to be run somewhere.” We personally think that’s like “Okay, well, there’s a massive infrastructure opportunity that –” Well, maybe not even opportunity, but just a fact that a whole new stack will be built to support models to be able to power the end user experiences. And I think that’s kind of the core insight going into Baseten.

And talking a bit about Baseten - what’s changed is that we kind of… Two years ago, we were talking about data scientists, we weren’t talking about engineers. I mean, that’s pretty key to our story, which is that - I think we came to the realization that every engineer needs to grapple with machine learning now, as opposed to maybe a smaller market of data scientists. I think going from smaller models that run in-memory, to larger models is another big focus change we’ve had. There’s a bunch of language stuff, and NLTK stuff - and you’re doing all that work, Daniel, but for the most part, everyone is using scikit-learn. And scikit-learn models, for the most part, run in-memory.

Kevin Ball: I do write some code still… But it’s pretty rare, and it’s declining. In fact, there’s an interesting aside on this, something that came up in a retro. So I still occasionally will tech lead a project, where I’m doing the architecture, and things… That’s actually pretty common in the industry. There’s this like whole role of like tech lead manager, somebody who’s managing a set of people and tech leading; lots of companies do this… And it sucks. And it sucks for a few different reasons.

So I had already kind of realized why it sucked for me, because it feels like there’s too many things and you can’t be good at any of them. But came up in this retro a couple of other reasons that sucks. It sucks for the team, because number one, you create a bad power dynamic, right? So like a tech lead needs to be making estimates; they probably need to be asking their team, “What’s reasonable?” but they kind of also need to put a line in the sand, like, “I think we can do it by this time.” That’s hard enough as it is, but if the person who’s trying to put those lines out and asking the team, “Is this reasonable? Can you do it? Is there a manager?”, now there’s a power dynamic that makes it harder for them to be clear and say no. And you can overcome some of that through relationships and other things, but that’s also a structural power thing. That’s not going away. So that’s a problem.

The other problem comes back to the calendar thing, right? If you’re tech leading your project, you want to be available to everybody on that project to answer questions, help guide, do all these things. I’m in meetings from eight till I get done with work. How is that going to happen? I’m not available if I’m also tech-leading the project.

So I thought that was a really interesting insight, and especially because – once again, I had sort of thought, “Okay, this is not a great situation for me, but it’s filling holes in the team; it’s going to help us get there”, but it’s actually harming the team when you have a manager also being the tech lead.

And so you know, realities being what they are, in many teams you don’t have somebody else to do that. Thankfully, I do have some people who can tech-lead, and we’re kind of making sure that that happens going forward. But yeah, it was kind of an interesting, like, “Oh, I was thinking about me, but I should have been thinking about the team when it comes to why this is a bad idea.”

Kevin Ball: Abso-freakin-lutely, because you brought out the JSX. Now, JSX is a domain-specific language for generating HTML. And if there is one programming concept that is under-utilized in the industry, it is writing DSLs. Domain-specific languages. You can get so much freakin’ productivity and communicatibility by moving into a domain-specific language. And what this does is it allows you to bring the full power of JavaScript to bear on your problem, as well as domain-specfic sugar, and help, and a way to talk about and think about things that make sense for rendering HTML. And as much as I do love templates, and I do love templating languages, and I love writing bare HTML, there is a real cognitive load to jumping back and forth between mental models and programming languages… While JSX essentially lets you stay in the JavaScript mental model throughout. And you can write it in all the different ways you write JavaScript. So if you like writing imperative, stateful JavaScript, you can write imperative, stateful JSX. If you prepare pure functions and composition, and really clean, and beautiful, functional architectures, you can do the same thing with JSX. And this is a huge reduction in cognitive load as you move through parts of the application. This was the key initial insight of React, was you put these things that are conceptually together, together in a place where you can work on them without having to jump context, without having to jump mental models, without having to do all of these different mental gymnastics, and it’s a dramatic unlock in terms of productivity and being able to think about it this way.

Brian Kernighan: But they’re growing up in a world where computers and communications are obviously pervasive, and that’s changing the world extremely rapidly. It’s accelerating. And I think it’s important that anybody who thinks they’re an educated person ought to know some of that stuff, about how do computers and communications work, and how do they affect people, and what can they do about things like privacy and security and defending themselves in various ways. And you could see the effects everywhere. I mean, think of the social media - mostly bad, occasional good. The advertising industry, cyber-whatever… All of these things.

So the kids in my class who were the non-technical ones in this class - they’re gonna be in positions of power and authority and influence to a degree which I’d say the kids in computer science are much less likely to be, actually… So wouldn’t it be nice if these folks knew something, so that when they’re running the world 20-30 years from now they don’t make silly mistakes about technology, or at least they are better able to assess what’s going on and make better decisions? So that’s the hope of the book.

[01:12:06.17] I got into the whole thing kind of by accident. I spent a sabbatical, sort of – you know, one semester. I was still working at Bell Labs, but in 1996 I spent the fall semester at Harvard teaching CS 50, which was this big introductory course for pretty much everybody who wanted to learn anything about computing. I did it as a visitor one semester… And what I discovered in that class is there were lots of kids who were very capable at programming; they were the ones who’d started programming when they were basically five years old, but there were lots and lots of other people who had no insight whatsoever into computing, and would probably never need computers, but had to learn something about it; it in some ways satisfied a requirement. And it was hard to have one course that would satisfy that broad a population… So when I some years later wound up at Princeton, I thought “Why don’t I try and teach a course for the non-technical end of that broad spectrum of the kids who were history majors, or English majors?” and that sort of thing. That’s the genesis of the course.

It’s been a lot of fun… I mean, it really is fun to try and explain the kinds of things that you and I would think are interesting and fun and important and all that stuff to people who come from very different backgrounds and may not appreciate why we think it’s important, or fun, or relevant.

Chris Benson: I think that’s a fantastic insight that you just made there, and that is that from an economic standpoint, people who make these large models, these foundation models, are intended to solve the problems which are probably being done by what is current mainstream, current top of the power stack, if you will, in terms of companies and countries even. And so there is an incentive to unintentionally perpetuate inequality with these. And so I think that would be a real dilemma.

How might you tackle that going forward, to where if we’re looking at foundation models as being somewhat core to most workflows in the AI space, because they’re available and they accelerate where you’re trying to go, but that they are built on imperfect data that has bias, that leaves out marginalized groups - do you just iterate and add those in and redo that? There’s a certain investment, obviously, to be made in that. Do you have any thoughts on least evil path forward to get the best foundation model with the least amount of unintended consequence over time?

Sathy Rajasekharan: [40:04] The goal is always to get that mom care as quickly as possible… So now we’re looking at what other data, what other information will help us increase our risk rating for a mother, will help us understand which facility she should be referred to… Something we didn’t even talk about is half of the organization works with providers on improving the skills of providers, and so we have a lot of data and insights from that kind of layer of the health system… So can we use these layers of data to really route those moms more effectively to care in a timely manner? That’s work that’s going on now, and just scratching the surface of it is to look at conversational history and look to see if there are triggers, trigger intents that could be predictive of a future danger sign. And it actually looks like we may have enough data to develop a model just from the initial work that’s been done.

And then - Daniel, you were asking a question about the users, and I think the ultimate goal is to support moms with what they need in terms of information, whether they wanna know “Where should I get my baby vaccinated? What kind of diet should the baby be on? How do you transition from foods?” And these are questions the moms are asking us. They want more and more information, because in the context that we live and work in, that information, that kind of support is hard to come by. So the lower cost, the more digital, the more close to the mom we can get, on her phone, or maybe phone plus, the better. So we’re looking at things like voice, we’re looking at the home record in terms of their own medical information that they wish to choose [unintelligible 00:41:50.29] free to access for them, so that it can help them provide a case history to a provider easier…

So there’s a lot of future thinking around this, but the principle is always “What does that mom need to get care quicker?” So lots of work going on in the background right now.

Asim Aslam: To the people who branded that, that was pretty phenomenal. I think it was actually a branding agency that did that. But Hailo was a taxi app founded in London in 2010 by a gentleman named Jay Bregman who had come from New York, previously founded a company in London as well that was acquired by Royal Mail; it was called eCourier. So he had this idea around this taxi app… Uber was already around… And he wanted to focus on the regulated market, and kind of said “Hey, there’s already this taxi supply. If we bring them online, then we can service all these consumers.” That was super-interesting, and it scaled phenomenally quickly, to the point of raising venture funding very fast… And they had the breaking point in terms of organizational and technical scale, and that’s why they required this complete replatforming, and that’s where I learned all these insights about the technology…

Eventually, it turned out that business model worked against them as they tried to expand beyond these markets that were not as highly regulated or suffered from problems. One example would be - you know, San Francisco is probably a pretty bad place to try to operate in a taxi market, because the taxis aren’t that great there. New York has this entire medallion system, and… You know, you can already find a cab everywhere. It’s a good thing. But beyond some of these markets it was quite difficult.

Uber had really understood, like “Hey, this isn’t about regulated taxis. This is about getting people a ride, which means we have to move into different vehicle types, which means we have to move them to Uber X, lower the prices, do ride-sharing, all this kind of stuff.” And we were kind of stuck in that model, so the business was going sideways. Eventually, they brought in a different CEO, the business became profitable by removing all those cities that were not doing well, and got sold to Daimler, the company that owns Mercedes, and it merged with another taxi company in Europe called MyTaxi. Now it’s roaming around under the name called FREE NOW.

Chris Benson: Yeah. They also use artificial intelligence to power 3D people counting cameras and facial recognition systems on a vast scale. I certainly can understand why the U.S. government – if any of that is accurate, and if they were using it in certain venues, I could understand why there might be a concern over that; that doesn’t mean, obviously, that I have any insight into their activities, or how they use it… But I could certainly understand why any government might be concerned about the potential for another government to be able to use these technologies to gain insight from an intelligence standpoint.

Tim Bell: That’s right, and I think especially with coming into school curricula – someone told me this, there’s 50 countries around the world that are introducing this into their grade school curriculum… And when I talk to people who are doing it, they go “Oh, now we’re gonna have to buy a computer for every kid” and so on. Now, again, you do need a computer; I mean, in the end, if you’re not writing programs, then you haven’t quite got your fingers dirty and you’re not really getting the whole thing, but…

[15:46] One interesting bit of research that came out recently is that they tried – I can’t remember the exact numbers, but I think it was something like a series of about eight weeks on programming for young kids, using a language called Scratch, which is really popular for younger kids (a block-based language). And they tried two variations - one was to spend the full time doing nothing but teaching them how to do things in Scratch, and the other variation is they spent about a third of the time doing nothing but Unplugged material, and then the remaining two-thirds teaching them about Scratch and programming.

And at the end, they found that both groups were equally good at programming, but the group that had done Unplugged first used a wider range of blocks or commands from the system, and had a much better understanding of just some of the deeper elements of what they were doing… Which actually was, again, not an intentional thing, but it kind of makes sense, because the Unplugged thing gets people thinking about “What’s the big purpose? What are we trying to achieve here?” and I think often when kids get put in front of a computer, they just go “Okay, tell me what I have to type. Show me the model. I’ll copy it onto the thing, and it’s finished, and I’ve done what I’m required, and I’ll move on.” But by getting them away from the computer, you’re actually getting them to think about the big ideas.

I think one of the key things with Unplugged too is that – people go “Oh, it will be so cool. One kid could be the CPU and one kid could be the input/output unit, and they could pass this onto their –” and that is so boring. Being a computer is quite boring; I don’t know if you’ve ever tried it, but it’s quite a repetitious sort of job… So the Unplugged stuff really turns things inside out; it will be a magic trick, or… It’s things where there’s an intriguing puzzle, and very simple instructions too, which is one of the really important points.

For binary numbers we just say “Here’s some cards. They can either be up one way or the other. You can’t have them half-way up. Now I want you to show me this.” That’s the entire instruction set. So it’s a relatively simple puzzle, but it gives the students a lot to think about as they work through it.

Searching is probably another good example, since I mentioned search engines before. One of the ways that we’ve done that is we get a bunch of caps (paper caps, or something) and we put numbers underneath them. Then we basically lay out about 30 caps on the table and just say “Okay, I want you to find the cap that’s got number 83 underneath it.” And of course, if the caps are in random order, the kids lose a point; in fact, sometimes we give them a candy that they have to pay for each one they look at… And they go “Oh, we’ll never find it… There’s 30 there, and you’ve only given me four candies to buy a view underneath a cap.” That shows the frustration of a linear search.

Then we simply say “Okay, here’s another set of caps. The smallest ones are on the left, the largest ones are on the right. They’re in order. Here’s your four candies. Are you feeling lucky?” And we don’t say anything else, we don’t tell them how to do it, or anything like that. Some kids will start at the left hand side and they’ll go “Oh, it’s number 3, it’s number 6… Oh, hang on a minute…” Then they realize that if they’re looking for 83, maybe they should look more towards one end than the other. But very often, particularly with kids around 8 to 10 years old, they’ll go “Oh, hang on, the middle one would be a really good one”, and they look at that and physically they really they’ve just eliminated half of the caps at the cost of one candy. Then they get to divide it in half in their way, but they have invented the algorithm themselves. I mean, it’s a well-known algorithm, but suddenly they see the power of it.

[19:56] And the next question is, “Well, that was 30 caps. What if I’d given you 60 caps?” They go “Oh, that’d be twice as hard.” And then they think about it; the very first candy reduces the 60 caps down to 30. So then we start looking at that exponential growth, or actually, technically, logarithmic power of a binary search… “What if I give you 1,000 caps?” “Oh, that’s… Oh, hang on a minute!” and then they start to realize that that’s only 10 candies, and a million caps is only 20, and suddenly, a billion… And there’s a lot of humor here, as well; like, imagine a billion caps… It’s quite a lot, but it doesn’t take too much to reason that you just need to look under 30 of them and you’ve found something. And now we’re suddenly showing how you can search a billion things, like a billion web pages - it’s not quite that simple, but you could search a billion things by only looking at 30 of them… And you suddenly realize that this is quite a different kind of thing to just doing a linear search or having to go through every single one… And the kids have come up with the idea.

I think a lot of the point of it is that these systems seem magic too, and especially through the early 2000’s, people were kind of locked out of knowing about things. You know, shiny gadgets came out - iPods and iPhones and things like that, that if you looked inside, it would void the warranty, and I think with a lot of them you couldn’t even write your own program for them without getting it approved by Apple. So they were kind of saying “Just trust me, there’s magic in here. You’re not allowed to look, you’re not allowed to change it. We’ll sort it out for you” and it sort of turned people into users, instead of developers… Whereas back in the ‘80s, if you bought a computer, you switch it on and it says “Type a command.” It’s the first thing you have to do, and it’s quite a different kind of invitation.

So we end up with these devices that seem like magic. I mean, they are kind of magic, but if we can undo some of that magic and kids look inside and go “Oh, I can understand how that works”, then that gives them some of the insight. And I think especially for teachers too, because more likely as teachers get into it, they go “Well, this stuff is a great mystery. Only a few very clever people like Bill Gates and Steve Jobs understand it, and they’re out of the business now anyway, so all the magic is locked up”, whereas it’s actually “No, it’s stuff that lots of people can understand at some level.”

Gaurav Seth: Yeah, I think one of the big pieces that we have not talked about is stuff that’s happening outside of ChakraCore, and the work that we are doing for Node in general. One of the things we’re doing is, given that – Node ChakraCore is one piece, but one of the efforts that we’ve been working on is like, for today’s Node developers, you wanna have a great experience of allowing them to use Node on Azure as a cloud platform.

We’ve been doing a lot of work, ranging from “How do you really improve your inner loop scenarios?” When I say inner loop, it’s like as soon as you start to code up your stuff right from editors such as VS Code, there’s a lot of effort that has been going there… How you use the programming language; there’s a lot going on in TypeScript that helps and works along with VS Code and powers some amazing experiences in VS Code, to kind of now enabling and working with Docker and the container servicer to bring all of these things light up on Linux, on Azure… Which people typically think like “Hey, you talk Azure? Azure maybe is just like a Windows cloud”, which is totally incorrect.

So kind of taking these things - VS Code, TypeScript, Docker is working on Linux, on Azure, either in the form of PaaS or Azure Container service, bring your own container… Then, once you have your app deployed, how do you go take the next step with app insights, and have some amazing experiences from a production diagnostics perspective? That is one.

The other thing is having the ability of full stack diagnostics experiences in VS Code, which is like “Hey, I’ve got the backend and frontend both in JavaScript - how do I target both of them in the same editor?” So there’s actually a lot of work that’s going on right now, to make sure that we kind of go nail the end-to-end experiences, so that as developer start working with Node, they have an amazing amount of productivity and they can get to the end solution that they wanna get to, in as minimum time as possible, let me put it that way.

Julian Shapiro: Libscore came about from two things. One was after marketing NameLayer I realized how important it was to actually assess how effective your marketing is. It seems self-evident, but when it comes to open source, there was actually no way to do that. Even if you are not marketing your open source project, let’s say you just get it out there on GitHub - you have no idea who’s using it. If you scroll through the list of everyone who’s starred it, maybe you can find in their bios they’re the employee of Yahoo!, but that’s not a representative metric, that’s not enough. A lot of people don’t star, a lot of people use npm, so in open source in general, I just had no way of figuring out how effective was my marketing. All I could track were page views on VelocityJS.org. So I knew I needed to close that feedback loop that any growth-minded entrepreneur would wanna close. That was the first thing. That was like the seed of the idea.

At some point, I heard that Stripe was having this open source grant program which hadn’t been done before. It was a for-profit company saying, “We wanna sponsor interesting open source developers, working on projects that we think will have an impact, and maybe tangentially related to the work Stripe does.” I got the grant, and they took me out to San Francisco.

On that grant I realized I’m actually mostly done building Velocity. It was at the point where it worked well. It wasn’t encountering much bugs, and so I thought, “Well, maybe I could actually use this as an opportunity to pitch Stripe on a brand new project, something else.” And I told them the story about how I marketed Velocity and how I wanted to track its usage across the web, but doing so seemed very intimidating. It didn’t seem like something I could do in my room.

[40:09] I felt like I needed to be with a bunch of very technical people; I needed the processing power, I needed some expertise in scraping, so I took advantage of all these people around me at Stripe to put together a small team and a little bit of sponsorship to build Libscore, with the intent of scraping the top million most trafficked sites on the web, to determine which JavaScript libraries those sites were using.

It turned out that part was actually very easy. At that point in time module loaders weren’t very heavily used to the point where they were obscuring the footprint in the browser of which libraries were in use, so I could just sniff all the top-level variables and run a bunch of heuristics to figure out, “Is this variable random-leaked stuff, from code that was messily written? Or is this the exposé endpoint of an intentionally installed third party JavaScript library?” So I would try to figure out how to do that, and then I would extract that list from the web page, and then index that across a bunch of things, put that into this big database and then make that searchable.

That was the intent, but all I knew how to do myself was that scraping portion to figure out which libraries were in use on the page. So I got together with this incredible designer named Jesse Chase. He was the creative director of Digital Ocean at the time, who built the whole UI and the graphing system. And then Jason Chen, founder of another open source project called Quill, which is a rich text editor. He built the actual backend architecture and the system that scraped the web. And collectively we built this thing that finally gave a footprint to the impact of open source projects, whereas it prior did not exist whatsoever. You didn’t know who was using the project. So the outcome was not just an idea of, “Okay, great, people are using my project”, but also “Oh, wow, Uber.com, WhatsApp, Twitter - these companies that I use, these sites that I go to every day are using my thing. That’s really cool.”

So what I saw was Libscore motivating developers who may not have had that insight into who is using their stuff to be like, “Oh wow, I really wanna make sure I continue maintaining this. I want to merge these PRs, there’s a lot of people depending on this, and now I know who that is.” And it provided that tangible footprint, which was pretty exciting.

Alex Russell: You asked a really insightful question earlier, “How did we get here?” And part of how we got here was that engineers enjoy a lot of autonomy, and a lot of power inside their organizations, in ways that I think are not – they’re not always legible. But it is the case that as an individual engineer, you have a great deal of authority over how the task is accomplished, and you can bring your values to that question. Not all the time, maybe not in the best way that you could possibly imagine, maybe not in the purest form… But it is worth trying, because we got here because some of us were confused about what that world looked like, who we were serving, how it was going to work… And we don’t have to remain confused, but we still have that power. We can help shape whether or not products are succeeding more often or not. And then we can learn to socialize the value of data and evidence in evaluating whether or not it’s working, not just to make ourselves look better, or because we did a better job, although hopefully you will get promoted more effectively if the organization also learns to value evidence and data, but because that’s how we preserve our social license to operate.

[01:09:59.14] We will continue to be able to make websites for a living, if and only if making a website for living results in something that people want to use. So we all have skin in the game here, and I think it’s really important that we acknowledge that we have some incremental - not absolute, but real influence on the kinds of decisions that our organizations are making. You can push back on the nonsense; you can call it out as nonsense when it isn’t going to work for those users.

So let’s not use a blanket kind of “These people are good, these people are bad”, until the evidence is all the way in about behavior changing or not… But we definitely need to put evidence at the center of who we’re serving and why, and I think we can do that.

Amal Hussein: Fred, that was really insightful and super-interesting, kind digging into some of the context and history behind some of the motivations of Astro… Y’all make some pretty bold promises here; we’ll list through some of those. So Astro is content-focused, server-first, fast by default, easy to use, and fully featured, but flexible, right? And I think that’s interesting. That’s a tough – fully-featured, but flexible. I mean, that’s – how do you make it easy for newbies to get started and make good decisions without having to really think about it? And then how do you not hold back power users? That’s always the struggle with API design.

So let’s get into some of this… So we’ve talked a lot about the content-focused piece… Can we talk about the server-first piece? Because there’s quite a bit here with just being server-first. For me, I kind of read that y’all are promising that this renders in the server. So you don’t have to ever worry about any state management, or whatever else, unless you want to. But I’m just curious, what does that even mean? Can you explain that to folks?

Amal Hussein: Fred, that was really insightful and super-interesting, kind digging into some of the context and history behind some of the motivations of Astro… Y’all make some pretty bold promises here; we’ll list through some of those. So Astro is content-focused, server-first, fast by default, easy to use, and fully featured, but flexible, right? And I think that’s interesting. That’s a tough – fully-featured, but flexible. I mean, that’s – how do you make it easy for newbies to get started and make good decisions without having to really think about it? And then how do you not hold back power users? That’s always the struggle with API design.

So let’s get into some of this… So we’ve talked a lot about the content-focused piece… Can we talk about the server-first piece? Because there’s quite a bit here with just being server-first. For me, I kind of read that y’all are promising that this renders in the server. So you don’t have to ever worry about any state management, or whatever else, unless you want to. But I’m just curious, what does that even mean? Can you explain that to folks?