Bruce Schneier: It’s a total, complete con. In a lot of ways, all of this blockchain-based finance is speed-running 500 years of financial fraud. You’ve got wildcat banks, you’ve got Ponzi schemes, you’ve got unregulated securities, you’ve got pump and dump… It’s all there. Front-running… And it’s all illegal in the normal banking world, because it’s bad, and it should be illegal… But because nobody’s regulating these blockchain-based systems yet, a lot of people are losing money.

Another fun experiment - go on to Twitter and type “I’m having trouble with my Bitcoin wallet. Can anybody help me?”

Daniel Whitenack: The majority of people running and operating this thing are data scientists and NLP people, so we needed something that we could support and not some thing that we would have to know a lot about HPC systems to support. So that was one thing. And we needed a way to queue jobs. And the fact that we could also integrate that with our runs on Google Colab and all that was really nice.

Now, we do use also– so this is for our training jobs. We do also use other solutions for persistent data pipelines. We use Pachyderm for that, which actually allows you to create and subscribe to data sources and pump those through to update data sets in a versioned sort of way. So we use that for other purposes, but ClearML gave us that experimentation for NLP research. And what we’ve found is we can run our experiments there, but then we can also kind of– and this is where I wanted to get with the connecting the pieces.

So we’re using notebooks, we’re using the GPU server, but then these models and data sets that we’re creating in our experiments - we can upload those now to Hugging Face, data sets hub, models hub. Those are versioned in that model hub. And then any inference we can do, we could either use the Hugging Face inference API, or we could call down the model using the Hugging Face libraries into Python code, where we could serve that model in a custom application of some type.

Jerod Santo: And we’re ready to go. First up, we did have a – we primed the pump a little bit… First of all, I should say this is for everybody to participate in. We were afraid or concerned that perhaps nobody would come up and we’d have not much of a show… So we did prime the pump and I do have a short list of people that we’ve contacted before, that we would love to chat with. As we make it through the list, maybe some people don’t show up, maybe some people do… We’d love to open it up for everybody to come and talk. And then if we run out of time during this session, we are down on the third floor, booth 72, with the mic all set up, and we will continue the lighting chats there if you don’t get a turn here during this session.

The first person we’d like to call up on the stage is Danese Cooper. Danese is happy to join us, but has a flight to catch, so she’s going first… And here comes Danese. A round of applause for Danese!

Pia Mancini: I don’t think I face particular challenges from the motherhood perspective. When I got pregnant, I sat down with my husband and we agreed that this was 50/50 equity on our daughter, and that’s something that’s a rule in my house. So we really share all responsibilities. I would say at this stage - Roma is two and a half years old and she’s pretty independent, and has her own organization, nursery school and things like that, so it’s easier. I think it was really harder on me when I was breastfeeding. That’s very challenging.

I remember we started working at this office space and the bathroom didn’t have a plug, so I couldn’t pump, right? I didn’t wanna stop breastfeeding because I was working, and I didn’t wanna stop working because I was breastfeeding, so I had to run out of the office to my doctor’s office… He would lend me a room to pump for half an hour and then go back to work, and things like that. The world is not super prepared for these situations in general, or the office space… I think that that was the most challenging thing for me, having to come back really tired, and put Roma down, and pump while I was answering emails. All of that – the first four, five months are very hardcore. After that, at least for me, everything has been much easier.

Kevin Owocki: But you know, I try to take a long-term view on this, because I think it’s really 1996 right now for blockchain. That’s how early we are in this whole new world. The pump and dumps and all the tokens that are just cash grabs, they’re eventually gonna get filtered out, just because of the Darwinian nature of capitalism. What’s gonna be left is the Facebook or the Google of Web 3.0. Well, I don’t know. Invite me back on in 2025 and we’ll see what happened.

Tim Mecklem: Yeah, so the community really understands. There are only specific versions of the pump, and they have to be older in order to loop. It’s a definite drawback. You have to really understand what you’re doing when you’re going to – you’re saying building a product… I’m not really looking to do that. I’m looking to enhance what’s out there and be able to support the community better, but I’m not looking to build something that I could sell here.

Pieter Hintjens: Yes, so the story is kind of Prometheus and stealing fire from the gods and giving it to the mortals to cook their hamburgers on. Me and a couple of my friends got involved years ago in making messaging for the financial industry - these were my first major protocols. I think at a certain point it became clear that what we were building was actually really interesting for other people, too.

The first people we were aiming at as customers were smaller financial businesses, small trading houses, so we focused on raw power and speed, and making it open source and trying to get license fees from these people who of course never paid anything because they were so greedy, they wouldn’t pay a cent, but it was great fun.

Then we kept making it simpler and more accessible, and at a certain point for me it was like “Okay, we’ve actually cracked the problem of building distributed systems. We know this stuff works.” Our software ran the Dow Jones Industrial Average for years and it worked. It was really efficient, it was straightforward. ZeroMQ is better than that. It’s 25 or 100 times faster, it’s simpler. This stuff is really really valuable, but our target audience doesn’t really understand the problem space yet, and one of the biggest handicaps is the complexity of this whole thing; distributed systems are very complex. So we’ve been working towards making that simpler and simpler and simpler, as a metaphor, as a model, the learning of it. When that works, it’s a thing of beauty.

One of the stories in my article - people have been writing on my article all kinds of comments, and one of the stories from somebody is this guy who was asked to make a prototype for the NFL. They have put RFIDs on the players and they’ve put little scanners around the field, and they can collect all the data from where people are, where every player is, and they wanted a system to bring that data in and pump it out to their machines. So this guy wrote in a weekend a ZeroMQ program in C++ that pulled in and distributed. It kind of worked, and then he went off and forgot about it.

A while later his company folded, he was fired, and his friend at the NFL said “Come, come work, come back.” It turns out that his code was being used as the basis for the NFL’s Next Gen statistic system, which is one of their products, and it’s the same code; essentially it hasn’t been changed very much, there’s a team around it but the core of that was 500 lines of C++, and it worked basically the first time and it’s still working. The same model has been working now for a year and a half very successfully.

Jerod Santo: Oh, yeah. And you can trick it in different ways. Like, my wife will stand there – or sit there, because she’s not standing… And she’ll just pump her arm like this; it’s to get her stand goals. She’s like “If I pump it enough, it just thinks I’m standing.” Stuff like that.

And I’m a completionist, so I was on the train of closing the rings, and what happened to me was my Apple Watch broke, and I had to either go get a new one, or be done. And like day one and two, I was like “I’ve gotta go get a new Apple Watch. I’m gonna lose everything.” And then like day three and four, I was like “Freedom…!” I was so happy not to be tracking myself anymore. I still do not have a watch to this day. It’s been like a year and a half now. And it’s just like “You know what - I don’t need that in my life, it turns out.”

Carlisia Thompson: I don’t even have to be necessarily around somebody like that… Just being around myself. [laughter] Literally, every once in a while my confidence goes down to a level – this happens once in a while, not all the time… Once in a while, my confidence just dips and I have to go on LinkedIn and look at everything that I’ve done, because that’s what I forget. And I’m like, “Oh my gosh, I have a CS degree… It’s a masters degree - that’s amazing!” Then I pump myself up and like, “Wow, I’ve done this project and that project”, and then I pump myself up.

I have to go and look at my resume multiple times a day for a few days to bring my confidence back… [laughter] Has anybody done that?

Alex Russell: Yeah. One of the contributors to one of the earliest JavaScript toolkits that I worked, something called [unintelligible 00:31:27.17] a guy named Brad Neuberg, who is I think spending most of his time on ML topics these days… But I was lucky enough to work with him for a while at Google. He got involved in a project with Doug Engelbart, trying to rebuild a piece of the system that they had imagined for hypertext early on… This is probably 2011-2012, but working on ideas that had been husbanded by Doug Engelbart and his group from the ‘60s. And that transclusion idea sort of made it into pieces of this work… And they are one of these things which is possible at the limit of our shared vocabulary, and our shared performance; to the extent that we can afford to represent the same concepts in a document that’s hosted on someone else’s computer, then we get to do transclusions, and to the extent that we can’t, we don’t. I think it’s fascinating, going back to – rolling back all the way to the JSON idea. One way to think about this is in the footer of every document that’s been “SSR-ed”, why don’t we have a copy of those same exact semantics living as a JSON structure, which then gets rehydrated? I think that’s probably what we mean by JSON… Because this is actually the dumbest version of taking a local data model; the one and done local snapshot of a data model is the dumbest version of that data model… And from the perspective of just going back and forth across several pages, if you just go look at the structure of - I don’t know, just pick on anyone; a New York Times web page. Just view source on a New York Times article, and if it doesn’t make you laugh, you aren’t looking hard enough… Because you get some HTML at the top, and that HTML pulls in some JavaScript, and then if you scroll halfway down, you’ll get to the bottom, and it will have the same HTML as a JSON structure. Everything, literally everything, as a JSON structure.

And what does that tell us? Well, it tells us that the way that we want to make those things interactive has to bootstrap onto some representation that is not the HTML. And so we don’t actually share those HTML representations anymore. We’ve tried to abstract yourselves up off of HTML and into some other thing. And then we want to operate on that thing when we want to make changes, and then splat it back out in some other way. And this is kind of - you know, I think back to Carson’s point about having some necessity driving this… Without the ability to upgrade the semantics of HTML, maybe because browsers get good at doing this again - although we don’t have a great track record there… Or because we have failed to bring along a system for doing it naturally ourselves, we wind up in this logic of building a parallel world, and then having to reconcile it with the displayed reality of the thing that actually pumps our rendering system, the thing that actually takes the diffs and turns them into DOM diffs, and turns them into CSS diffs against that, and layout trees, and then render trees, and then raster.

[34:10] That whole flow requires that we pump it with something, and how far away we are from that core semantic model that it’s designed to be pumped by is this tension that we feel. And that I think gets us into this question of “Well, why did we pile into the clown car of web 2.0?” And it’s interesting to me that all of the examples that we bring up about where that stuff was truly revolutionary are systems that had many clicks, or taps, or drags. They had many detailed, drilled-in interactions per session… And it’s in those places that that local data model thing shines. And it’s in the other places, which is - I’m just gonna handwave and say the vast majority of interactions on the web, where you have shallow sessions, whatever the designers intent, where you aren’t doing many clicks or taps against something that would be better if it was a local data model, that this stuff just feels like it’s… It feels like cruft, but we’re not sure why, even though we have some innate belief that it’s good for us. Because we’re not looking at the diversity of session depth.

Daniel Whitenack: [00:17:47.11] Yeah, it’s quite interesting, Chris… I think it’s actually a very subtle, but creative way of doing this control. So the process is as follows. In the blog post they kind of break this down into four steps, and there is data that’s needed, but you’re not creating data for the purpose of training the model, you’re creating data for the purpose of generating these - what they call control vectors. So the first thing you do is you say, okay, let’s say that we want to do the happy or not happy, or happy and sad operation. So you create a dataset of contrasting prompts, where one explicitly asks the model to act extremely happy, like very happy… All the ways you could say to the model to be really, really happy, and rephrase that in a bunch of examples. And then on the other side, the other one of the pair, do the opposite. So ask it to be really sad. “You’re really, really sad. Be sad.” And you have these pairs of prompts. And then you take the model and you collect all the hidden states for your model, while you pump through all the happy prompts, and all the sad prompts. And so you’ve got this collection of hidden states within your model, which are just vectors, that come when you have the happy prompt, and when you have the sad prompt.

So step one, the pairs of kind of like a preference dataset, but it’s not really a preference dataset… It’s contrasting pairs on a certain axes of control. So you run those through, you get all of the hidden states… And step three is then you take the difference between – so for each happy hidden state, you take its corresponding sad one, and you get the difference between the two. So now you end up with this big dataset of - for a single layer, you have a bunch of different vectors, that represent differences between that hidden state on the happy path and the sad path. So you have a bunch of vectors. Now, to get your control vectors, step four, you apply some dimensionality reduction or matrix operation. The one that’s talked about in the blog post is PCA, but it sounds like people also try other things. PCA is Principal Component Analysis, which would then allow you to extract a single control vector for that hidden layer, from all these difference vectors. And now you have all these control vectors, so when you turn on the switch of the happy control vectors, you can pump in the prompt without an explicit [unintelligible 00:20:43.16] to be happy, and it’s going to be happy. And when you do the same prompt, but you turn off the happy, and you turn on the sad, now it comes out and it’s sad.

Adam Argyle: I want to announce a tool to help people. So I mentioned oklch.com, which is a phenomenal OKLCH colorpicker. I have been working for the past few months to help enable people into getting into these new color spaces, to learning and seeing what CSS color can do, and especially seeing the results that you get in CSS gradients. You can now go to gradient.style. And that’s where I’m building a very much it looks like a design tool to help you build these new HDR gradients, and see the SDR version fallbacks. I’ll generate a fallback gradient for you, and generate the new modern gradient for you. You can change color spaces, you can do all this stuff. And it’s really, really nice, really visual… And you can even drag it across from an SDR monitor to an HDR monitor, and kind of see it that way as well, and just see the change, and start to feel out these colors.

I think there’s two main things on that tool that are really fun to do. Well, there’s a lot of main things to do, but the first one is use the colorpicker. So if you see a little color dot - that’s when your color stops - click it and you’ll get the world’s first next-gen CSS colorpicker, where it supports every color space that CSS supports. They’re even grouped and tell you if they’re HD, or SDR, and stuff like that, which is really nice. And they give you a little controls to go play inside of there. And it will convert. So if you have a HEX color, for example, you can paste the HEX color into the color [unintelligible 00:58:39.25] open up the color picker, go to OKLCH, it will convert the HEX to OKLCH, and then you can go bump the chroma up and be like “How much brighter can I make this HEX?” You’ll like “Oh, crap, that’s a lot brighter.” And you can do that.

Another fun thing to try on the gradients that are in the bottom left; there’s some presets where I’ve created some things for you to go explore and try on. And then there’s a Discord if you want to join and ask questions and learn more about color spaces, and about CSS color. In the top right there’s a settings cog; click that, and go to think - it’s like Help and Feedback, or something like that. That’ll pump you out to the Discord to come have conversations in there.

There’s also – just above that is a tips and tricks. You can click that, and I’ve put a bunch of hints all over the UI to help you learn about what the UI is doing, and help you get you the ball rolling into building a new gradient in these new color spaces.

Yeah, gradient.style. I’m going to probably be working on it until the end of the year. You can tell some of the stuff’s about done. It’s in beta, but it’s a really strong place to start as we transition as an industry out of SDR - which we’ve been in for 25 years - into HDR. It’s just the future.

Montana Low: [24:07] Yeah, I think Lev should tell you about pgcat and I’ll tell you about PostgresML, because I think they’re two different pieces of the puzzle. But to answer your question - like I mentioned, we weren’t able to get a lot of the deep learning models into Postgres. And so because we couldn’t get that, we didn’t even really start with all of the XGBoost models, or even simpler models that we were running.

I think there’s a big disconnect between what is flashy and hype in Academia, or coming out of OpenAI. There’s a lot of reinforcement learning, or unsupervised applications, or whatever. But in the business context that we saw across at least 20 different models and machine learning applications at Instacart - they all basically boil down to you’ve got some relational data and you’re trying to predict a single column; and you’re gonna do some joins, and that’s it. First thing is you pump it through a linear regression, you work your way up through the scikit-learn algorithms, you hit XGBoost, and your predictions are gold standard. You don’t even need deep learning 90% of the time.

And so thinking through like the convoluted data architectures, and data engineering, and everything else that we had to do to get features to models, I was like - well, the data is in Postgres right there; you can just write a join, and you can just do a select, and you don’t even need to bring a lot of it to the application layer if you can actually do your ranking, or whatever it is that you’re trying to do.

Ranking is obviously a big application of machine learning. You need to know what your popular things are, your trending things are, your relevant things are, your possible alternatives to this thing are… There’s all these different way that things are associated to things in this high-dimensional space that data scientists love. So being able to pull things, only the most relevant of any of those dimensions out of the database, the application layer… Which is actually a really expensive operation, compared to things like deep learning. People think deep learning inference is expensive; it’s really not, relative to like taking thousands of rows out of a database, serializing them, sending them over a copper wire that’s multiple feet long, reading them into a JSON blob on the other side in some dynamic language that’s allocating a ton of memory to do all of these operations, so that you can operate on this in Python or Ruby or whatever it is… That’s where actually most of the latency in the system comes from. It’s not from the models themselves; the models themselves are like highly-optimized C code that, sure, it may have millions or even billions of parameters, but it’s relatively fast and optimal.

And so I was just thinking, what if we could cut all of that complexity, all of that latency out, and keep things in the data layer? And I was talking to Lev about this, and I was like, “I’m gonna go on vacation, but when I get back, I’m gonna start on this project.” On vacation, on day one, Lev emails me and he’s got a new commit. He’s like, “I’ve got deep learning in Postgres!” Lev is really phenomenal this way, in that he’s very competitive. If you tell him about something, he’ll try to beat you at it. [laughter]

Zed Shaw: See, I think you don’t get examples. I think this doesn’t happen too often for two specific reasons. One is the companies that are doing this are HUGE. Can you imagine if you had an open source project and you decided you wanted to re-copyright it, remove your license, and that would affect the bottom line of like, say, Amazon and Google. Let’s say there’s something they’re using, that you know they’re using, you’re not making money, you’re desperate for work, they don’t hire you, so you say “Screw this! You can’t use my stuff anymore. You’ve gotta pay me.” Now, I guarantee they’re gonna pump probably 2-3 million dollars into lawyers and just bury you. So nobody tries it. They know they’re not gonna be able to do anything about it. If you tried to fork Kubernetes, they’re gonna use just the way programmers are to keep people there - the way the community is run, the way open source is sort of like considered a community thing, “How dare you…!?” They do this all the time. So I think you don’t have examples because the system doesn’t provide examples.

Then I think the second reason that all of this is allowed - I’ve gotta say, I’m not upset or really blaming the corporations for doing this, because the programmers let them. I can’t say that – corporations just do corporation. That’s what they do. If they can make money anyway, that’s what they do. I mean, they try to pretend they aren’t that way, but that’s the way it is, that’s reality.

So actually, for a long time I didn’t blame the open source projects, I didn’t blame anyone or anything like that in running projects; I said “This is just corporations exploiting people.” But then, when I started talking about how they sort of allowed this, I would get death threats if I criticized a project… And there was this whole other side to open source, and I just sort of started to see that, like, really, the reason these guys are getting exploited is because they’re just a bunch of servile fascists. They just like things being in charge.

Jason McLeod: Yeah, exactly. Exactly that. So we got on a call, and we were like “Alright, our game’s most like these categories. And without changing our game, how many other things is it like?” Because we were trying to just figure out, “Can we use the original idea and then find the twist?” But then maybe - gosh, what was it? Like 15 minutes into it, we went back to our ideas board, because Gustavo and I have a very unofficial game company… It’s just the two of us, Pixel Party Games; I had to drop that… But we went to our board and it was like “What if we took this game and we applied it?” And the game that we had kind of been noodling around with was actually something we anticipated adding to the Rune platform even before we knew there’d be more and more jams. Because Rune will let you pump games to their platform, independent of a jam. I think they would love that.

So if you guys are familiar with the game Pipe Dream - it probably goes by a lot of other names, but the idea is that you’re given pieces of like tile-based, like pipe, and then you have to place those pipes in a formation that would allow the liquid to make it either from point A to point B, or as far as you can, without [unintelligible 00:25:19.27] back on itself, that kind of stuff. So this time around, we’re gonna be making a multiplayer version of that. But the twist is there’s not just a queue of pieces that just appear; you have to run around and grab those pieces of pipe from somewhere else on the level, and then run them back, navigating the current pipe layout like a maze, and place them in a position to where the line can keep going.

So yeah, we had that [unintelligible 00:25:45.18] and thought that idea up, and at least I left that one kind of relieved, because we have a really hard time making really elaborate scope, and then telling each other that we should reduce it, but then secretly not reducing it. And I’m like “Remember that thing you said we didn’t need? Well, I went ahead and implemented it.” But yeah, this time around we’re definitely sticking with multiplayer, because I think that’s what Rune is going to fit well for us, what they offer us, that SDK for that… But yeah, the classic pipe dream game, with a twist of multiplayer.

Chandler McCann: Exactly, yes. As I was saying before, the challenge we were trying to solve was why do these water points - either wells, or taps, or rain harvesting systems - break. So the dataset that he set to normalize includes things around the location of the water point… So there’s cell phone applications, they can take a picture, capture a lat/long and geolocation; so there’s also image data, to some degree.

Then there’s information on the source of the water, as well as the technology… So whether it’s coming from a river, or groundwater, or if it’s a tap stand itself, or a pump. Then it’s got information about the country and region it’s in, as well as the installation year, who installed it… And then some interesting factors, such as “Are the communities paying for it? Is there a management structure in place in the community to maintain the well or water point itself?”

Matt Heath: Yeah, I think those things have been really essential for us… Both being able to just pump a lot of metrics into a system and then go and introspect it later, and both to use that for investigations later, but also monitoring in real-time… And then - yeah, the tracing mechanism; we use Jaeger now, and for a while we had a system which would use structured logging to propagate a request ID…

[44:12] So you could look at all the log lines across a hundred different servers to just pick that needle out of that haystack and combine them all together. And I think those were the only real ways we could understand those kind of things. But with those tools in place, I think it’s quite useful, really beneficial.

One of the questions I was gonna ask, actually - I think one of the things you get for free if you have that approach is… We have many services, and a call graph crosses many of these service boundaries. And each one of these boundaries has all these automated breakpoints effectively that are instrumented. So every call between a service has the perceived time from the caller and the server. So you have all these checkpoints all the way down your call graph, where you can see the performance of different things, which engineers don’t need to think about, they just get for free, because the client that we use adds them. How would you do that if you were building a monolith? …because you’d have to remember to add them, presumably at all these different places. And then that depends on the team, or is there a common approach?

Jerod Santo: In fact, you can take us into the next segment, because you’ve primed the pump. I’m just gonna set a marker here and say “Go ahead, Kball.” Start us off.

Ryan Singer: Well, I don’t really see it as having much to do with – I mean, if you want, you can look to the brain to find some sort of neural correlates of this stuff, but I don’t really look to the brain as the seat of these things. I think it’s much more about the circumstance. You can black-box the brain, and look at the circumstance, and then you get the right variables out. It’s a question of what level of abstraction you wanna deal with the world on.

You raise a really good point about this thing, about compassion and empathy… I think we have different things which drive us, in terms of what we like to work with and where that moment is where we do the fist pump in our own work… Like what’s that moment for you where you’re like “Yeah! It all clicked together and I achieved my goal.” I kind of sit in-between a lot of roles, because I have programming in my background, but I also do product and user-facing stuff… So I feel like I have a little bit of both. But there’s a satisfaction as an engineer of flipping the switch and then watching the current run-through, and then the gizmo does what it does on the other end… You know what I mean?

Jerod Santo: It’s very cool. Yeah, it’s funny, because I looked into Erlang probably way back when you were starting DNSimple, and took a few tutorials, and I just couldn’t get the syntax to stick into my brain… But I liked, for instance, the pattern matching that you’re describing, and a lot of the properties of it just seemed really smart. And I couldn’t quite get over the hump. And then when Elixir came out, it actually – because I was a Rubyist before that, and I knew Jose and everything, it all kind of connected. And so I’ve been writing Elixir for a long time, and enjoying some of the advantages of the BEAM, with a language that for me just more syntactically aligns with the way I think. But I think once you get past that, if you’ve written some Erlang for long enough, it’s like the matrix; you don’t even see the ones and zeros anymore.

Joran Dirk Greef: And you can jump and spin a beetle, and you do get a score. That one is quite hard.

Jimmy Miller: I enjoy – I find I’m much less distracted by having that constant background stuff going on than I am often focused in… I’ve always been of two minds about flow state. When I was younger, I definitely think I had flow state type stuff quite a bit, where I would stay up till like 4am coding, and like not realize the time. And I still do have moments of that. But I think most of my coding doesn’t happen in a flow state. I don’t know that I can achieve a flow state very well. I get very distracted very easily, and jump back and forth between things.

But I do think that there’s a lot that’s good about that general advice. Like, one of the things I’ve found that I’ve been really struggling with over time is to stop starting and start finishing. To like actually push through on a project. And one of the things that has been really helpful for that is being willing to make tiny incremental changes, even if I’m not feeling it. And I think that’s part of this, writing every day. Some days you’re going to write like a ton. Some days you’re not going to feel it at all, and you’re only going to write a little bit. And I have found that to be super-useful. Sometimes I’ll not be feeling coding at all, but I really want to make sure I actually get – I’m building a programming language right now, and I really want to actually get it so it’s good. And so I’ll just spend like 30 minutes fixing one small little bug, or cleaning up a comment, or whatever. And then I find myself like three hours later still working on stuff.

Adam Stacoviak: “I came to get down, so get off your feet and jump around.” That’s right.

Jerod Santo: Yeah. So I don’t know if I would have made it over that hump, because it does have that steep initial learning curve, had I had a choice. But I did have a choice, I made it over the hump, and my college teacher was very proficient in vim, and he would live-code in class. And we would watch him and be like “Oh wow, you can really move fast if you get good at this thing.” So that was also motivating.

So I learned vim, used it for many years. I still use it off and on all the time, mostly in the terminal, mostly SSH into a machine. That’s my go-to in-terminal editor. However, I do like a nice graphical user interface specifically for text editing, and so I picked up TextMate back in the day, eventually moved over to Sublime Text, and Sublime Text 2… Never hopped on the VS Code train, although I have installed and I have used it.

I used Sublime Text for a very long time, and nowadays, I have recently, within the last six months, switched to Zed, as I find it the only other Sublime Text alike that is fast enough, and light enough that it feels Sublime-esque. Sublime was always just faster than all the other ones. And VS Code started fast, and then has slowly gained, in my opinion, some weight… And so I’ve been using Zed, and really – I still keep Sublime Text launched, just for one-offs. You still can’t beat it, in like I’m just gonna open up a buffer, and do some stuff, and do some text manipulation, and then just get rid of it; it’s still faster. But for project work, I’m chilling in Zed.

Justin Garrison: I love debugging. Debugging is one of my favorite things. It’s struggling, but I love diving deep into that stuff, and like “Well, let’s pull out S-trace, and TCP dump, and see what we can find.”

Justin Garrison: Yeah. And at some point, you just have to ship something. You have to say “Hey, we built the thing. Here’s why it didn’t work out, and why you might want to choose something else.” So let’s go ahead and jump into that interview, and we will see you all after to talk about the Meta white paper I read. [unintelligible 00:04:28.25]

Break: [00:04:32.04]

Alya Abbott: Yeah. I mean, I can’t promise anything about your specific experience, but we have had folks tell us that when they moved to Zulip, they actually started getting much better community engagement, because it works quite nicely for folks who are not around all the time. I mean, one kind of category of folks, as you were saying, are may be people who are lurking, or who are just kind of coming by once in a while to check in on something… And if you’re coming to something like Slack, it’s hard to – you might see the latest messages. It’s really hard – it’s going to be not really possible for you to kind of catch up on what you missed if you’re checking it out every couple of weeks or every month, in an active organization… Whereas for Zulip, if you just want to sort of check in on things occasionally, folks will come in and they’ll look at that recent conversations you maybe saw when you were exploring the app, and instead of having to look at sort of individual messages and try to figure out what’s going on, they’ll just see that list of topics, and they can be like “Oh, this topic sounds interesting. Let me jump into that.”

[00:52:14.13] And so you don’t even have to feel obliged to kind of make everything be marked as read, or kind of manage your own reads necessarily, if it’s just something where you’re not following every little detail. You really can kind of just skim that list of what’s been going on, and jump into the ones that are of interest.

So yeah, we’ve had folks say that something like an open source project, that it can actually really be great for community engagement, because people can select the parts that are interesting to them, and just follow those, and jump in on those. You can even configure notification – there’s a concept of following topics. So once you’ve seen something that’s interesting, if it’s a community you’re not engaged with very regularly, you can follow that particular topic and say get email notifications when there’s more messages just to that topic. And so there’s really ways to follow specific conversations and find things to engage with for occasional users in a community.

Suz Hinton: [laughs] There are threat hunters at - yeah, like private institutions. So for example CrowdStrike’s threat hunters are actually threat hunters for hire. So they work with companies directly. And if you look at the product offering online, it’s called Overwatch. There are different tiers of it, where they’ll even give you briefings on the latest threats to look out for, and things to maybe specifically look at for your industry even. So if this company is a financial tech industry and they’re working with CrowdStrike, the Overwatch team, the threat hunters can give actual briefings on what they’re seeing as trends in that financial industry based on attacks on other companies that are similar to them.

The threat hunters do a lot of different services, and so it’s going to depend on whether you’re in the private or public sectors to like what tools you use as well. So I’m sure that there are teams that use JIRA to keep track of intrusions, and dump a bunch of data in there… But I think that a lot of these tools tend to be very proprietary. And so they’ve been designed and developed and incrementally improved based on the specific kind of work that these threat hunters are doing at their institution. That’s all I can say.