Gavin Wood: Sure. The DAO, if you’re not aware, stands for Decentralized Autonomous Organization. Essentially, it’s a means of bringing various people together into some form of bound contract, or some bound set of interests. The idea behind the DAO is more or less like a fund, a VC fund. It can back particular endeavors, and the idea is that the endeavors will ultimately make money, or create value, and push the value back into the DAO so that all of the members of the DAO get in some sense richer.

Now, you can buy into the DAO - that was the idea - by depositing Ether under the control of this DAO contract. It’s a bit of an odd thing to say, “Well, putting money under control of a contract? How does that work?” But that’s exactly how it works - you write some software, and the software states the logic of this agreement, of this multilateral agreement, and part of the logic is saying under what circumstances money should be transferred. In most the same way that you might have a joint bank account where you need both signatures before the money is transferred, so is the case for the DAO. Under the DAO, it will happen with four thousand or five thousand signatories, and there’s some much more complex thing about how many of them had to agree to it before funds were transferred. But nonetheless, basically it’s that kind of thing. It’s a shared bank account with very particular logic for who gets what money, and for hopefully integrating the eventual ventures back into the DAO so the DAO itself can reward those who backed the right venture.

Anna Derbakova: Yeah, that’s the basic concept; that’s still there, because in Bitcoin you’re exchanging essentially value, right? You would be exchanging value of Bitcoin, but in Ethereum you could be exchanging some kind of coin, but you could also be exchanging an asset, because they have the idea of a smart contract. But the idea is it’s a different type of blockchain, it’s a different implementation, and Hyperledger also is slightly different. But who owns it? That’s a good question. I mean, there’s not gonna be one blockchain to rule them all. They are gonna be different, and I think that’s a very good forward-looking question, because then people will wonder, “If we have this blockchain here for this group of banks, and this blockchain here for supply chain, if they ever need to interoperate, or something of as far as transferring records from one to the other, what are we gonna do?”

So I think that’s a question that’s definitely gonna need to be explored fairly soon. If you have two different blockchains with different consensus algorithms, how do you deal with that? These questions… I’m not sure if IBM, for example, is researching anything into that - at least I’m not directly involved in it - but if it doesn’t come up soon, it will in the very near future.

Till Döhmen: Yeah, so one of the principles of this dual execution is to reduce the amount of data that has to be transferred as much as possible. One of the use cases, for example, is I have a really large dataset on S3, and I want to join it with a small table that I have on my notebook. So in that case, an optimizer, a query optimizer, will make the decision to, instead of downloading the one-terabyte dataset to your local device and doing the join there, to instead upload your small local file to the cloud worker, and do the processing there. So that saves, in that case, a lot of bandwidth.

[00:40:06.25] The same with a filter pushdown. I query a large dataset on S3 again, and the transfer only has to happen for the filter. You can get something similar with DuckDB as well if the data is partitioned. So DuckDB has clever ways to optimize remote file access as well, without MotherDuck. But the thing you get with MotherDuck is it even filters the data if your data is not partitioned, because the cloud worker still takes care of doing the bulk of the work, and only gives you the results you actually want and need.

Arisa Fukuzaki: Sure, sure. My name is Arisa, and I’m originally from Japan, but living in Germany. Storyblok is a headless CMS, and I work in there as a developer relations engineer. Not just on the headless CMS… What makes it unique is that we have the real-time visual editor, so whenever you make content changes, you can take a look at it and see in one screen.

[24:37] We’re based on the atomic design methodology; it means that you can reuse the components, as many as you want, and you can nest these reusable components in any levels. So you can only, for example, define only the necessary components that you need, and you don’t need to duplicate unnecessary [unintelligible 00:24:57.25] components.

So that’s what we have… And here for today’s talk I was showing that, let’s say, choices of choosing the technologies; I was showing as one of the examples that imagine that if you are for example in your everyday life, when you face something non-user-friendly interface, then maybe in the end you will be able to get something you wanted to have, but the journey until you will be able to get the things you wanted to have, or the things you wanted to achieve - it’s not going to be nice. Same thing could be transferred or would be applied in the tech world.

I was showing how resilient, fast, very smooth and organized experience you can get from, let’s say, Storyblok and Remix. Because on the fly, Remix already provides a very nice way to dynamically handle the routes. Let’s say anyone who doesn’t have the developer experience - they can easily create nested routes dynamically, from let’s say the dashboard and the visual editor… And developers can focus on their own work, and stuff like that. So that’s something what I was giving a talk today.

Anna Henningsen: Yeah, shared array buffers are [unintelligible 01:02:01.00] and array buffers can be transferred.

Tim Bell: [03:49] Yeah, and the kids were raving about it, and it was really fun. Then the next week the parent to come along was me, a computer scientist. At the time and for a long time after, my main research area was data compression; how to make data smaller, and all that sort of thing. And I thought “How do you explain this to kids?” They don’t even know what data is, they didn’t own a computer… Nothing like a data projector in the room or anything to demonstrate things… So I don’t know why, but I just had this idea “Well, let’s not worry about the computer at all, let’s try and get to the heart of what I think about, what I care about when I’m solving problems, when I’m trying to develop a new program for something.”

So I sort of peeled things back and I thought, well, what are the key ideas that we’re looking at? I just looked at a whole lot of topics that exercised my mind as a researcher in computer science, and I tried to think “Well, how could that be transferred into a game, or an activity, or something like that?” I came up with three or four activities, went along into the class… I was using nothing but cards and paper and string and chalk and things like that, and we had a great time with the class. In fact, it went so well that they invited me back, which was a very pleasant surprise, because I thought it might have been one of the least interesting talks that the kids might have had. So I ended up going back regularly, and developed a whole series of these things.

And back then I came across a colleague on the internet who had been doing the same thing in Canada, Mike Fellows. Mike and I pooled ideas, we put them together and we said “Oh, we should get together.” I went over to Canada for a month, and we sort of tried to put it all together in a book, and came up with about 20 activities like this. It seemed really good, so we sent it off to some publishers and said “Here’s something a bit different… What do you think?”

It was rejected 27 times by publishers, which is the best thing to happen… The reason was that they couldn’t place it, so we sent it to a computer publishing place, and they said “Well, you’re not using computers. It looks really cool, and by the way, can I keep a copy to do with my kids? …but we can’t publish it in a computer book. Why don’t you send it to our education department?”

So I sent it to them, and they’d go “Well, it’s obviously about computers. It’s got ‘computer’ in the title.” Anyway, we got sent around in circles, and we got more rejections than Harry Potter, which I’m very proud of. But the cool thing about that is that in the end we thought – because back in the ‘90s, open source and publishing into that wasn’t a big thing… In fact, the web was really just starting to become a thing, and the public didn’t really have access, or teachers and so on… But we put it up in the form of the web, and just said “Well, help yourselves, everyone. It’s free, and we think it’s a good idea.”

And it just sat there for probably about a decade. A bunch of people around the world started using it and came across it one way or the other, but around about – I think it was 2003 at the ACM, which is the national body of computer scientists that set curricula, and things like that, started looking at K12 curriculum… And kind of almost unknown to me – they’d probably emailed and asked me about it, but they published a thing about it saying “We think there should be a K12 curriculum for schools, and by the way, here’s some examples of what to do” and about two thirds of their examples were straight out of our material… So suddenly, everyone was contacting us and saying “Oh, tell us about this method, tell us about your philosophies”, and so on.

Of course, at the time, we’d just gone back to our day jobs and we weren’t particularly working on this… So the thing suddenly started taking off, and then even more so I guess in the last eight years or so, because all around the world countries are starting to say “Well, how do we introduce computer science or computational thinking into schools?” and this became a really easy touchpoint for that. Because of not using computers – it took us a while to realize exactly what we’d done, but it means that it’s using things that teachers are familiar with. So when you say “Put this on the card and draw this on the ground” and all that kind of stuff, they go “Well, I can do that.”

[08:23] But if you say “Oh, you’re gonna teach computer science”, particularly if you think about primary school teachers, elementary school teachers, they go “Well, hang on, I haven’t been trained on this. It’s probably all this thing called ‘coding’, which sounds really mysterious… I’ll probably have to install something on a computer, and when I install it, it probably won’t work, because then I’ll have to install something else and reboot, and then I’ll need permission from the school to install it… Nah, I can’t be bothered.” But when you say “Just print out these three cards and start doing stuff with them”, it sort of becomes a very easy entry point.

So it sort of evolved from an outreach thing for academics like me to go into a school and just talk to our kids, or try and drum up a bit of interest, to something that is being offered for teachers, who have been told “This is now part of your curriculum.”

Adam Stacoviak: Something like that behind the scenes, to allow money to be transferred, to enable things to take place… And to organize and sustain things. That’s an interesting direction. Maybe speak to that, since you said that Open Collective – your daughter is roughly the same age, because of the same timeframe… What was that for you? It’s that your first child? That’s a new thing for you, it’s gotta be scary.

Alexander Neumann: This deduplication thing was built into Restic right from the beginning, because at first I thought “Oh, I’d like to have a backup program”, and then I started thinking about it and discussing with my colleagues what should a backup program do, and one of the things that you have in a backup program - you have duplicate data. Either you have the same file at different times - sometimes files haven’t been modified, so you have exactly the same data and that’s really easy to handle, but sometimes you have virtual machine images of like 100 gigabytes and you just have changed one or two bytes within the whole image, and it would be really a shame to store this 100 gigabytes twice, because most of the data is exactly the same.

So I started looking into algorithms that try to detect changes or similarities in data, and one of the ideas that have been built into this is this really old tool called rsync. There’s a PhD thesis by somebody called Andrew, I think… I’ve forgotten the name.

rsync does really interesting things. For example, when you try to copy a file to a remote server, then on the local site they have some process that opens the files, starts reading and sending to the remote site. If this is cancelled for some reason - for example, you cancel the program locally or your internet connection breaks down… Afterwards you restart this process and rsync will detect that there is some part of the file already on the remote side, it will open it and it will find where it left off on the previous run.

This is the easy part, but what happens when the file on the remote side was modified and you would like to make it pristine again and copy the original file over to the remote side again? You can just delete the remote file and start transferring again, but that’s not very efficient. rsync cuts the file into different blobs and detects which blobs need to be transferred. For example, if you just changed one byte because of a hardware error on the HDD or something like that, then you just need to detect which of the blobs changed. For example, the first blob, the first 1,000 bytes or so - rsync will detect that and will transfer this small amount of data and reconstruct the file on the remote side.

The algorithm that it uses is called a rolling hash sum. It starts by reading the file and taking all the subsets of 32 bytes from the file, and for each of these 32 bytes it computes a hash. When this hash has some property, for example the lower bits are set to zero, then it says “Oh, I’ve found a block boundary!”, and afterwards is uses a real hash function, a cryptography hash function - I think it uses MD5, or something like that - to detect if the content of the blob has been changed.

So you could have also kept the file into 1 kilobyte pieces, but then the problem is you cannot detect when one byte has been inserted at the beginning, because all your blob boundaries are wrong. And when you have these dynamically-sized blobs, then you can detect “Oh, there has been a byte inserted; the first blob is different, but all the other blob markers at the end of the file are completely the same again.” This is really efficient to dynamically slice the file into blobs, and this is what Restic does.

The problem with the algorithm that `rsync’ does is that it is targeted at really small blobs, for example 100 bytes or 5,000 bytes, and in a backup program we don’t deal so much with inter-file duplication, but with intra-file duplication. We have some files that are not exactly the same, but mostly the same. So it’s a good idea to have larger blobs, because when you have a snapshot that I’m doing of my directory now, and two days from now most of the files will be exactly the same and some will be modified, but most of the data will probably be exactly the same, so it makes sense to reduce the number of blobs that you need to handle and have larger blobs.

Restic at the moment is configured to try to have one megabyte of data as one blob, and this was just not possible with the algorithm that rsync uses. So I looked at an algorithm called Rabin fingerprinting. It’s a really old idea by a mathematician called Rabin from the ’80s, and I read the original paper and it was a typeset with a typewriter, and I think the formulas were manually inserted and painted in there, something like that. I tried to understand the math behind it and then implemented it in Go.

Elham Tabassi: It certainly makes sense. I’ll try to answer the way I understood this. So AI systems are not inherently bad or risky, and it’s often the context that determines if a negative impact will occur, and also what are the risks. An example that I usually use is that if I use face recognition to unlock my phone, versus face recognition in the airport that now our faces are a boarding pass to get on the plane, or face recognition in the context of the law enforcement, it’s the same technology, but in the different context there is different risks and different levels of the assurances that we want to have for the systems to work in a trustworthy manner.

[00:28:30.07] So what we have been trying to do as part of our work in approaching trust and trustworthy AI - the first one was to unpack the concept. Try to get into the characteristics that make a system trustworthy. That helps to answer the question of what to measure. If I want to know if it’s trustworthy or not, what are the measurements I need to do? So I listed the seven characteristics, from valid and reliable, safe, secure, et cetera. So that gives a more of a systemic approach and structural approach to what are the dimensions, what are the characteristics that together can make a system trustworthy.

And by the way, AI-RMF talk about this, that not each of them by itself make a system trustworthy. You can have a system that is very secure, but not valid or accurate. So that’s not going to be trustworthy. And a system that’s a hundred percent accurate, but not secure - it’s also not trustworthy. So that gives it, again, a more structured approach on what to measure.

Then the next step is how to measure methods and metrics for the measurement. Those types of measurement gives information about limits and capabilities of the systems, the type of the risks that can occur, the magnitude of the impact if those risks occur. And then based on this information, then we can come up with mitigations and management of the risks.

So AI-RMF, its recommendations is really categorized in the four functions of govern, map, measure, and manage. The govern is giving recommendations on procedures and processes, roles and responsibilities that we want to have in our organizations to do effective risk management. So what is the accountability line? What are the roles and responsibilities involved?

The map functions provides recommendations on understanding the context of the use. Going back to the examples of the face recognition - understanding the environment in which the AI systems are operating there, understanding the community that can be impacted by that, identify the risks in this particular context, understanding the laws, regulations, and policy that are effective in this context of use…

The measure functions provides recommendations on the how to measure. So for all of the risks identified in the map measure, it provides quantitative or qualitative recommendations on how to measure them, how to take into account the trade-off between all of those trustworthiness characteristics… And all of this information is being used during the managing the risk part, that the recommendations can go from safeguards and mitigations that can put in place to mitigate risk, to - sometimes we cannot just mitigate risks, and the risks should either be accepted or transferred, or the system is too risky that it should not be developed or deployed. So that is the process in AI RML.

Break: [00:31:47.15]

Tess Rinearson: Yeah, I think, potentially. We’re definitely starting with financial institutions, and we’re starting with financial institutions that are well-known enough that they can power their own networks and people will want to join them. That serves the power of this partnership. We just on Friday announced a partnership with VISA and the network that we’re working on with them to do business-to-business payments, and it’s really valuable to go be part of these well-respected financial institutions, because people trust them enough to wanna join their networks. So we’re starting there, but I do think that, generically, anything that holds value is a good candidate for something that can be issued and transferred on a Chain block chain.

Mark Kurtz: So on the open source side, pretty much everything that we have on the SparseZoo currently has either been from our lab… We have a few that are from Intel’s lab up as well, and some Hugging Face examples, and things like that… It’s primarily because a lot of the sparsification research is all built around a few models, like ResNet-50, BERT, and things like that, and they don’t expand it out past those models to prove out their algorithms. So we have the best of those, and then - yeah, that’s exactly what our team is working on. And we get some community contributions in every once in a while for sparse models that people have generated or transferred.

So our goal is to be able to push these up, so that as you said, anyone can come in from the community and be able to pull those down and get value out of those models. So you can think of them as sparse foundational models, rather than the dense foundational models.

Break: [25:58]

Tim Coulter: …the physical good being sold, being transferred over. Here there will always be a level of trust involved when you deal with the physical world. You’re gonna have to trust FedEx in this case. But FedEx could implement features that track the location of the good and where it’s at and who’s handling that, on the blockchain, if that actually makes sense for them.

The FedEx example of actually putting that on there - that may or may not make sense for this example, because basically we’re all fine with checking our tracking number online and using their current existent databases. Blockchains don’t make sense all the time. But there are cases where you might want to track physical goods and where they’re coming from, and pair identity - which is something we’re solving - with the blockchain itself, so you then have this person making a statement that something about the real world is true, whether it’s the location of a physical good or something else. If we can do that right and we get the identity thing down, then it can unlock quite a bit of things.

It’s gonna come down to, you know, I suppose I could lie in person to you, talking to you, as much as I could make a transaction on the blockchain. But if we can build systems that make that harder or make it very clear that I’m the one who embezzled whatever this thing was, then perhaps we can increase accountability or what have you.

Daniel Whitenack: [00:16:11.24] Well, yeah, there’s definitely the side of this which is probably the more commercial side of it, which is media production, and that sort of thing… Let’s say that you produce a video in one language, and you’re wanting to do the dubbing across languages, or something like that… Or maybe even you’re using like an avatar and using synthesized voices in your video, and the whole thing is synthesized. I mean, that’s happening quite a bit right now as well. And so the ability to bring in multiple voices, and do all that without going into a recording studio - of course, that has huge applications for like advertising, marketing, media, production, entertainment, all of those different areas, which is where I would guess, and I can’t speak to Coqui’s business model, but I would guess that their tooling is quite applicable across those areas.

For local language communities - of course, they’re also involved oftentimes in the production of media or content for their communities, so that also relevant there… But I think there’s also unique things that are relevant in those scenarios. So imagine that you’re part of an indigenous community, a local language community, and you’re kind of marginalized by the national government, or discriminated against in one way or another… It might be a big ask for your community say, “Hey, could we put up 100 hours of content with your voice, and maybe your likeness?” That’s potentially painting a target on yourself, right? When you’re associating yourself, and you’re the face of that community.

So I think it’s really interesting that there’s tools like this, where you could create high-quality voice and expressive voice that’s maybe synthesized, and not someone’s – maybe even styled transferred, where it’s not someone’s voice that can be tracked to a certain person.

But then if you think about then combining that with the video elements… So think about having a video that maybe is recorded with someone talking - you can use things like Stable Diffusion and other things now to actually shift that video and obfuscate the identity of the person in that. Now, the more nefarious use of that, of course, would be misinformation, and deepfakes, and that sort of thing. But there is a very positive use of this for these sorts of communities where it is important that they want to produce media content for their community, but if you’re marginalized or discriminated against… It’s interesting now that there’s these tools that are accessible, and have really nice user interfaces, and accessible to community members where they could actually produce some of that content themselves. So it’s an interesting dynamic.

Austin Parker: Yeah. I do think there’s a couple of important differences in how Otel kind of approaches this. One is we have this idea of telemetry being kind of sparse, right? And being highly contextual. So what I mean by that is we don’t want to necessarily emit a bunch of telemetry about the same thing that means that it’s completely identical to each other. We want telemetry to layer on top of itself, and when we turn it from these three pillars - people might have heard of the idea of traces, metrics, logs, the three pillars of observability. The Otel concept is more like “What if all this telemetry was just kind of this one interlinked braid, where it was all self-referential, where my metrics refer to my traces, which refer to my logs, which refer back to my metrics, and I can have this kind of shared context value at a technical level that helps me associate all these different things?”

[00:44:06.21] And then also, going a step further, of saying “Okay, there’s that core base level of telemetry that all my stuff should emit. But then I do need ways to process it. I do need ways to transform it for different use cases. Maybe some of it needs to go into different – a security team wants something different than an app dev wants, which is something different than a front end dev wants, which is something different than a DevOps team wants, whatever. That data needs to go to different places. That data might need to be transformed in certain ways as it goes to those different places.” But as long as we’re all working off the same base level, and if we’re able to separate those concerns, or we’ll say “The telemetry is down here, and then these kind of telemetry pipelines are one step above it”, then it becomes much easier to kind of satisfy everyone’s needs while not fundamentally changing the nature of what we’re doing.

An important part of OTEL is the idea that we support lossless transformation. So when you create a metric in OpenTelemetry, you can actually translate that metric into other forms. So you can change from a delta, you can change how it’s aggregated, or how it’s displayed, using views, using other kind of techniques… And that is not a lossy transformation. You can actually work that backwards, depending on how you have your stuff set up.

Break: [00:45:25.10]

Daniel Whitenack: Well, as always on these Fully Connected shows, I think that our conversation has led us exactly to where we often like to go in these types of episodes, and that’s to learning resources for people; maybe new ones that have come out, or things that have been updated… And we are talking about two different scenarios. One where the challenge is having an organization be transformed into one that uses AI or data science type techniques, and the other where maybe you’re an individual and you’re trying to break into an organization that’s already using AI. Maybe on that latter point - I ran across something that I haven’t gone through totally, but I very much want to, and I’ve seen a bit of buzz around… This is this new book that is called Meta Learning. Have you seen this?

Kitson Kelly: Yeah. So we currently are using that at the moment to go figure out the analysis of what files that we need, and then we send that whole bit to Rust, and then Rust will go and fetch any modules, cache those modules that are needed and hand that back. So before we invoke the TypeScript compiler, we asynchronously resolve all of our dependencies, and then we just create all of the source files and hand that over there.

Now, the problem is there’s some long-standing issues with that, because it’s not a popular API, so it’s broken in a couple ways that causes grief and havoc. For example, string literals that look like the are importing the module get caught by this preprocessor… As well as - there’s another weird, strange… Well, it’s not really a big; the preprocessor does a really quick pass on things, but the TypeScript compiler does obviously a full AST parse, and it gets down into the depths. So with dynamic import, for example, it will find some static strings… Like if you set a string to a variable, and then you press that variable into the dynamic import, and say “import this module variable”, the compiler will detect that and say “Oh, I can pull in the types for that and do the analysis of that.” So it’ll identify a dependency where the preprocess imports doesn’t see that node, so you end up getting cache miss hit. We’ve got a defect with Deno at the moment where we don’t find all of the dependencies.

Anyways, the whole thing is that’s a lot of work, so we’re in the process of moving that out to Rust… So we’re gonna do all of that dependency analysis, and then what we’re gonna do is just say “Here are all the files the TypeScript compiler is gonna ask for.” Now, Ryan and Bartek and I have looked at how to even begin to start saying “Well, what if we did the AST processing?” Because that’s computationally heavy, it’s really inefficient in JavaScript… What if we were able to do all of that and then hand that to the TypeScript compiler?

[16:14] The actual language enforcement and the type enforcement - we would probably never want to not use the TypeScript compiler for that, but we wanna chip away at everything else around it that is computationally heavy, that is just wasted. [unintelligible 00:16:30.12] if we’re gonna do any transforms, and we only do really – because we support very modern ECMAScript; you know, whatever V8 supports. And V8 supports things in advance that it takes a long time to work into browsers. We had top-level await before anybody else did, simply because it was in V8, and we could get it in there.

So we don’t do a lot of transforms with the TypeScript compiler, but obviously things like experimental decorators will always be transformed forever. So if you need that requirement, there’s a few things that we have to transform.

But there’s this other project which we got on to that is called SWC, and it effectively is a JavaScript and TypeScript AST compiler and transformer, written fully in Rust… And it is super, super-fast. It is orders of magnitude faster than the Babel and the TypeScript AST compiler. The target for SWC is really to kind of replace WebPack/Babel in Rust. So it’s kind of doing that heavy lifting; it can strip types out of TypeScript. So it doesn’t enforce types, but it does the type-stripping, which Babel can do as well, and then it emits stuff, and they’re beginning to implement the WebPack type of things… And it is super, super-fast.

We have that integrated into Deno at the moment. The main purpose that it serves today - it does two things. If you use Deno Dock and the module, it’ll read the JSDoc and emit a text thing to your console of the documentation for the module. If you do it with –json, you’ll get a JSON blob emitted out from it, giving you a documentation structure for that. That’s all powered by SWC. And then the other thing in Deno that heavily uses it at the moment is that we had Prettier incorporated into Deno… So you do deno fmt format and it’ll reformat your modules for you. The problem is Prettier not being Rust, it’s pretty darn slow, and it was a lot of stuff to load into Deno. We were lazily-loading it, so we’d go and fetch it and run it, and it was just becoming a bit of a pain…

David Sherret, who has written a couple really awesome tools, wrote dprint, based on SWC. So 90% of dprint is in Rust, which he usually moves to WASM, but we’ve integrated it directly into Deno… So deno fmt is pretty darn fast in reformatting files, and you can reformat your whole project code with it.

Nadia Eghbal: I think there are a couple of interesting things that might be happening in tandem around that now. We haven’t talked about this at all, but just the definition of a software developer has changed radically in the past five years, where a lot more people are learning how to code. Maybe they’re not at a very high technical level, but just enough that they are able to modify small things around them and see that power. I think learning how to code has just become so much more accessible, so you have so many people that are interested in modifying the world around them in much more casual ways. That is blurring the line between consumer and producer. Look at any child today, everybody is learning how to code, and just imagine when they grow up and they just expect that everything around them can be transformed. It’s almost like people are coming at it from a different direction, but then at the same time you see all these very proprietary platforms that are basically exploiting network effects to centralize where people congregate on the internet, and those things are still total black boxes.

I don’t know what happens when the youngest generation now grows up… Will they say, “This is bullshit!”? “This is not how we were raised to see the internet.”

Kris Brandow: I know there have been attempts at that… But yeah, I feel like Neovim is – if you do want to be tinkering, or you really want a very specific setup… Like, I have over time transformed my setup to be very, very particular… But I don’t spend a lot of time messing with things, unless I’m trying to write a new language, or trying to do something else. So it’s like, I kind of set it to what I want, and then kind of forget about it… But I also feel like that’s bad. I don’t like that, because then I kind of get stuck with things that are slightly more efficient. I don’t try to change things up enough. So I’m trying to find a good middleground there, and I think Neovim kind of pushes me more toward doing that than other editors might…

But yeah, I can see that opinion, where it’s kind of like – it’s definitely not your “I just need a car to get me from point A to point B.” It’s like “I want to maximize my time. I like to go out and do adventuring, I like to be in that type of environment.” I feel it’s like a Jeep owner, versus a pickup truck owner, where it’s like “You’re not using your pickup truck for the things you’re gonna be using it for.” But I feel like people that buy Jeeps are generally people that like to go adventuring, and stuff like that.

Guillermo Rauch: And also, sites of a certain size can not just pre-render every page… And together with all of that, when we’re talking about mobile web versus native, a tremendous advantage that I see for the web is that when you tap on a link, it can render just in time the page just for you, without you having to have downloaded a massive application with all the possible code paths, not just for you, but for every known language, every known variant, every experiment, and every other cohort of users.

The web has this magical – and it goes back to my concept of the hyperlink… You go to this hyperlink, and it’s almost like a black box; you don’t know what’s gonna come out on the other side. And that has this beauty of “I can give you the specific set of code, I can give you the specific HTML, and the specific CSS for what you need, just in time, at that very time.” And that’s only possible if you combine ideas of JAMstack with more traditional ideas, like server rendering.

So this concept that we pioneered with Next.js, this is what frankly drove and got us the attention from Google Chrome’s team, that started contributing directly to Next.js, because it matched very well what they’d seen internally at Google as being the best ways to scale really large sites and really large applications.

[32:14] It’s always with a combination of CDN caching for the parts that are known to be static, but also leveraging what I call the power asymmetry of the world, which is - the average mobile phone in India, which is the Xiaomi Redmi phone, is not gonna be as powerful as a server’s pre-warmed with all your application code, sitting on a data center. And not only will it not be as powerful, but by transferring some of the rendering load to that powerful server, you reduce the battery life, you reduce the download amount they are putting on that device.

So I think one of the ways forward for the web will be to acknowledge that there is no silver bullet where you say “Oh, if I make everything static, and if I JAMstack this or that, I’m gonna get great performance”, because you might also limit the ability to customize, personalize and create very dynamic content just for a certain user, context, country, language at a certain moment… But not only that. As I mentioned with Web Vitals, one thing that we’ve learned as well in 2020 - and before that, but it became very clear in 2020 - is just because a CDN can give you HTML really quick doesn’t mean that you’re gonna have a fast website.

The example I like to give here is when Facebook launched the new Facebook.com built entirely on top of React, which was supposed to be this engineering marvel, because the open web is so introspectible and debuggable, a developer went ahead and inspected it and realized that they weren’t even using CSS that much. They weren’t using box shadow, and they were using spacer divs.

And the box shadow one was particularly puzzling to me, because I remember the old days of the web, where we didn’t have the box shadow property, and we had to use background images… Or not even background images, we had these image tags. And Facebook was doing that because they realized that for that particular element that was floating on the screen, box shadow had really bad performance.

So performance is not just a matter of like you downloaded that box shadow property in-line into the HTML really quickly from the edge. That obviously is gonna help. But then begins this other complete new box of surprises, or Pandora box of performance. And that’s where the Web Vitals metrics that Google has created are extremely helpful… Because if you are experienced with backend engineering, you know that one of the easy things to measure in this pipeline is “Okay, I’m gonna measure the p99 percentile distribution of response times and render times.” In fact, github.com used to include this in their status page. Status.github.com would show you the mean page render time, the p95 page render time etc. And you would say “Okay, it’s 300 milliseconds.” And you’re like “Okay, that’s really good.” It’s like two blinks of an eye, or whatever. The threshold of attention, or whatever.

But then you realize “Okay, I gave that to the web browser, and now the web browser is struggling with rendering box shadows, or Intercom widgets, and GDPR pop-ups, and all the layout is shifting, and there’s three cookie banners, and another one to accept the terms of service… Actually, that one is fresh from today.

[35:54] Today I went to a website… This is a marketing website. And it was presenting two blocking modals. One to accept Terms of Service, and I was like “Terms of Service of what? I’m just visiting this website.” And the other one is a massive GDPR modal. And both of those things kind of came in a little later…

Daniel Stenberg: Yeah, they’re really similar. But what I really like, because I think it’s a really fun position, is to be on the bleeding edge, so that we can be early with development on new protocols. Like, we support HTTP/2 really early, actually, an HTTP/3 also very early… Because it becomes a good sort of spiraling – we get it in early so that people who actually implement and deploy servers and proxies can use curl to exercise their code and try their servers, and they can then help us make our code better. Because I think that’s fun, and it makes us polish and streamline our implementation early on, and sort of make better decisions earlier in the process on how to do protocols, really. And also then, of course, it’s just fun to have that position so that we can help others make sure that they get their implementations done better.

[01:04:10.10] So that’s what I mean with bleeding edge, because I think it’s a fun place to be at, and listening in what the internet wants is more – that’s what I mean there, is I want to make sure that we offer the ways of doing internet transfers that the internet seems to be wanting. The protocols curl supports, I want them to be done in a secure way, in a modern way, and the way we want to do. So if you want to do internet transfers for your application, you want to implement a car infotainment system, you want to transfer data for that, you want it done in a modern way, which means this current protocol’s current security, the current mindset of how to do things. So that’s what I mean. That’s what I mean with keeping up with how to do things; that means the cipher suites, and TLS, and versions with HTTP, and how to do authentication, things like that. Because I think it’s important, because I think the internet is moving quite fast, almost all the time… So if we wouldn’t keep up, we could easily just be left behind.

I think we have a role, and that role should be providing service and how to do things in the proper way. And also because a lot of users are just using curl already, so we make sure that a lot of things are doing correct in security internet transfers by using curl. So that’s also a reason… To make sure that curl does things properly, we help securing the internet, really.

Break: [01:05:44.11]

Autumn Nash: Okay, so “How the tech industry broke its promise to women.” The title is actually better than the link title. But I think it’s really interesting, because we dumped so much money into diversity. There are so many diversity programs. And it’s crazy, this is not something that is just warm and fuzzy. This is something that Harvard Business has done studies on over and over again. It says that you will get a better return on investment, you’ll make a better software, you’ll have better teams, and they’ll innovate faster. And this is all stuff that Forbes, Harvard Business, all these huge names in finance and business and how to build a business that are very well known have told us, “This is going to add more value to your business.” And just the amount of wealth that is going to happen in the wealth transfer to women over the next couple of years, economics are saying that it might be one of the biggest wealth transfers in history. And we’re “But we don’t need 51% of the population. We don’t need their inputs.”

And it’s crazy, because women actually – I think we dictate 71% to 80% of spending, and we influence that much spending, and I think 70-something percent, 72% or 76% of discretionary spending. So think about it, when these companies are selling products to people, you have one person who does a lot of the family spending. Or just in general is your customer. And you are not taking in their perspective to what kind of products are they looking for, what kind of testing should you be doing. And we not only sold women “Hey, come, be in tech, we need you”, and then we were “Just kidding, because money is an issue, and we’re in a tech recession, so we’re no longer interested in diversity.”

So what does that mean for tech? What does that mean for how we build software and our return on investment, if we have this thing that statistically shows we are not going to build as good of software, and have as good of a return on investment without having people from different backgrounds, especially women, who is going to be a majority of your audience, and the purchasing power, and then we’ve decided to just abandon that? Where do we go forward?

[00:06:11.08] Also, retention is horrible. It’s 56% by mid career, by 35. So not only is only 25% or 27% of the industry women, and that’s not even mostly a technical… I think black women are 1.6%, Latino women are 2%, and those are not widely, like, engineering, right?

Braden Hancock: I’d say from the very beginning one of the ideas that drew me in was there should be this different interface for getting things done, for transferring information from an expert into a program that can now do work for you. I think historically, that’s very imperative code. Very much like describe exactly what you want done, step by step. And that was less interesting to me. It felt a little bit more like “Your job’s just to translate from one language to another.” But the cool thing about machine learning or AI in general I think is that you get more a sense, in the right setup, of “If you can tell me what’s good, then I can find it.” There’s better synergy between the human and the computer, where now I can show you what I want, even if I don’t know how to get there, and you can get there, where “you” here is the computer, of course.

So I think that’s the broader idea, that it was really appealing to me all along the way, that had me coming to machine learning. Then throughout my Ph.D. I kind of dove into that problem much more deeply, of “What really is the best interface for getting domain knowledge from an expert into a model.” Those are themes that I explored for multiple years, that along with my co-founders I ended up being with led us to Snorkel, and then Snorkel AI now.

Kevin Ball: Yeah, so right now – to think about this in a DSL way… So to deal with a set of byte sequences in vanilla JavaScript, you’d basically define those as const somewhere, and then you’d set up a bunch of methods for transferring in and out of those, and you’ll probably be passing those constants around, and things like that. But if those actually represent functionality, you could essentially wrap those up as primitives in your language, and just say “This is something that’s always there.” We don’t have to import the constants and have a central file, or whatever. This is just – it’s part of the language.

Theo Browne: I’ve been an anti-JavaScript person for almost my entire career. When I joined Twitch, I was a Ruby and Python guy, and I joined an Elixir team and got really functional programming pills really early. I fell in love with FP. Then Twitch moved off of the… I don’t even know how to describe the Ember codebase. It’s not even that Ember was bad, it’s that we used it at the wrong time, in the wrong way, and the result was a disaster, with this Ember codebase on top of a Rails codebase. We decided – we didn’t even decide. We purchased a company named [unintelligible 00:50:22.15] that was all in on React and TypeScript. They were told they had to get the Twitch site working in their Electron app. They said it would be faster to rewrite the site. And before Twitch could say no, they had finished a third of the rewrite in React and TypeScript, which is to this day the codebase that is hosting all of Twitch. I was so excited seeing that movement happen, seeing this team that was so motivated by the DX that they knew we could have, but didn’t, that before anyone could say no, they’d already built half of it. And that got me interested.

Then I got team transferred to a Go team, and I hated working in Go. It’s not a bad language, it’s a badly designed language. And it allows you to do really powerful things. And the fact that you can write code that is so simple and readable and performant is cool, but it makes me feel like a monkey, not an engineer. And I was miserable. So my manager, who was a frontend guy, said “Theo, you’re clearly miserable. You’re not getting anything done. I know you’re not a frontend person, but this React TypeScript thing is pretty cool. By the way, I heard a rumor they’re going to go in the functional programming direction soon. You might want to get in now.”

So I started learning React, I got decent at it, I was starting to [unintelligible 00:51:38.18] I was about three weeks in when the Hooks announcement happened, and then I was all-in. I got to be the person who brought functional programming to the 500-plus person codebase at Twitch for the whole site, and that was a magical moment. A huge part of my career trajectory skyrocketed, because I got to go from just a junior engineer, barely understanding things, to making real impact. And TypeScript was just the obvious way to do it.

[00:52:04.19] I guess for me JavaScript was this necessary evil for the web, and then I avoided the web for like five years, and then when I came back - this was late 2017, early 2018 - the thought of not using TypeScript just didn’t cross any of our minds. Like, you used TypeScript, and this is a codebase that needs to keep working. You don’t have another choice.

And my only previous experience with TypeScript was I was working on a Chrome extension that my roommate was helping out with. He was a big TypeScript guy in 2017, and it broke all of my experience trying to work on this extension, because I had never used Npm really before even, so it was miserable for me at that point. I avoided it for a year, and then when I came back in, things were in a good enough state.

I think the key for me was having a codebase that had it already all working and set up, and then once you started working in that codebase, if I didn’t have TypeScript, I would be miserable now.

Andy Klein: We grab a particular – I think they go… We call them pods, okay? The original storage pod, 45 or 60 drives, and then we go pod by pod. That’s the way we think about it. So we go to that pod and we run smartctl on the drives in there, we pull out the data, we store that data, and then we add a few things to it. We keep the date, we keep the serial number, the model number, and the size of the drive, and some other basic information that we record from the device itself. So we know what storage pod it’s in, we know which location it’s in, and so on and so forth.

At that point, we have the data, and then we store it into - I’ll say a database [unintelligible 00:26:15.19] actually stores locally, and then it gets transferred up overnight. That’s part of – the boot drives get to do that fun stuff for us. So we take the snapshot all of that data, we store it locally, then it gets brought up overnight… Then there’s a processing system which goes through and determines the notion of failure. So if a driver ported something to us, it didn’t fail yet. The next day, we go in and we go back to that same pod, for example, and we notice that one of the drives is missing, right? We look for 60, we only got 59. What happened? So that gets reported. And then that becomes part of what the software on our side processes over the next few days. “Tell me about that missing drive. What happened to it?” And at that point, we interact with some of our other systems or maintenance in our inventory systems to see what actions might have been taken on that drive. We also have some smarts built into the software itself, to help identify those things. And if all of those things make sense, then we go, “It failed”, or didn’t. “It didn’t because it was a temp drive, that was in for a few days, and then it got taken out, and replaced by the clone…” So it didn’t really fail, it just didn’t get a chance to finish. So we shouldn’t fail it, right?

Or we migrated that system. We took all of the drives out, and we went looking for them, and they weren’t there. But we don’t want to fail 60 drives, and so that’s not what happened. So the system figures all of that kind of stuff out. It looks, like I said, it can connect up to the inventory of maintenance systems to help validate that, because we have to track all of those things, for obvious reasons, by serial number.

[28:00] So it’s fairly complex software that goes in and does that analysis, and it takes sometimes a few days for it to kind of figure out whether a missing drive is really a failed drive, or whether a missing drive is a good drive, and it just got removed for a valid reason. And then once that happens, then we record it. Once a quarter, I go in and I pull the data out and look at it, and I’m looking at it for the most recent quarter. I actually go back in and invalidate all of the failures as well by hand, against the maintenance records in particular, just because we want that data to be as accurate as possible. And then we confirm all of that, and almost always we get a really solid confirmation. If we find anything funny, we keep looking. And that’s the data we publish, and that’s the data we base the reports on.

Gerhard Lazu: Okay. Now, you mentioned something very important. GDPR. I know that many people think about that. Many people think about data locality, the laws and rules that govern where data is stored, where data is transferred to… I think that in that world, shroms too as well, that’s coming, which are even more strict rules around data and where it should reside, for European users… I see a future where data locality is going to be very important. The data will not be able to leave certain countries, not even in transit, encrypted, it doesn’t matter. No, no, no. It will always stay in that country. In that world, I think cell-based architectures – sharding, but even cell-based architectures, local based architectures will become very important. Do you see Litestream playing a role in that world?