Justin Searls: Well, I’ve been a HomeKit person from day one, and the thing about HomeKit was it was also awful when it came out, and super-unreliable… And you’d tell Siri to do a thing, turn these lights on, it wouldn’t happen. And it was a coin flip whether or not it was because HomeKit didn’t work, or because Siri didn’t work. But then over the years, especially with this new V2 architecture a couple of years ago that came out for HomeKit, I’ve noticed everything around me in my house is rock-solid. And if there’s ever a failure, it’s that I said the perfect phrase to Siri and it failed to convey that same instruction to HomeKit that it did an hour ago. Because I’ll open up the home app and I’ll click the little button, and then the garage opens, right? That’s clearly, like, good job HomeKit. That’s actually a really good example of a very broken thing at Apple that doesn’t make any money iteratively getting better, because they could see probably from their logs the failure rates, the latency and stuff, they redid the architecture… Siri’s architecture, there was reports a few years ago that they did have this kind of fork in the road to determine whether or not to redo Siri as an LLM… And just because the politics of it, Tim Cook and the executive team decided to stick with what they had and try to fix it, try to improve it from within. And now here we are.

David Crawshaw: Well, yeah… So what would Tailscale do with LLMs is the question I was asking from a Tailscale perspective. I think Tailscale is extremely useful for running LLMs yourself, for a network backplane… In particular because of the sort of surprising nature of the network traffic associated with LLMs.

On the inference side – so you can kind of think about working with models from both a training and an inference… These are sort of two sides of the coin here. And training is very, very data-heavy, and is usually done on extremely, extremely high bandwidth, low latency networks, InfiniBand style setups, on clusters of machines in a single room. Or if they spread beyond the room, the next room is literally in the building next door. The inference side looks very different. There’s very little network traffic involved in doing inference on models, and in terms of bandwidth. And the layout of the network is surprisingly messy. This is because sort of the nature of finding GPUs is tricky, even still today, despite the fact that this has been a thing for years now.

Dan Moore: [01:02:14.16] I’m torn. I want that world. I want that world. I’m not sure we’re there, because LetsEncrypt, the big lever there was Chrome, and the scary warning messages in the URL bar, and things like that. And I don’t know if we have – I mean, maybe you have that with the operating system vendors… So maybe that’s the lever. But it feels like we’re not there yet. But yeah, I would love a place. I’d love a world – and honestly, it’s interesting to me, because the more we talk about this conversation… Password managers and passkeys are both kind of two sides of the same coin, or they’re two approaches to the same problem, that both believe that computers are better than people at keeping track of verifiers of identity. And passkeys do it in a way that’s a little bit more opaque, and not maybe as compatible, but is a little bit stronger, because it’s private/public key encryption. Whereas password managers are more designed to like fit in with the world we currently live in and have all these nice add-ons that you mentioned, Adam. But I just don’t [unintelligible 01:03:28.18]

Adam Stacoviak: Well, that’s a thing. I mean, I think it’s a thing, and he made it a lyric. He didn’t create it. He didn’t coin it.

Tobias Fried: “Insert coin” type stuff if you beat it. It’s pretty hard, actually. I made the ball move pretty fast, and it’s hard to beat, but…

Jerod Santo: Right. Because if they received a bunch of token for their package getting popular, and they went to go sell it and they were just dumping on the market, and the demand wasn’t there, then the price would crash, and you’d have your typical peaks and valleys of the crypto sphere. So you’re trying to like stabilize the coin, basically? Or what’s this tokenomics? You’re trying to stabilize the value of the token, or…?

Matt Rickard: Yeah, you’ve probably never heard of it, because I tried to come up with it myself… I tried to coin the term. So it’s a new thing. But it’s me trying to describe a pattern that I’ve seen in software configuration where a configuration seems to evolve through specific, increasing levels of flexibility and complexity before returning to either hardcoded values or Bash. So you go from hardcoded values, which are the easiest/simplest configuration, but provide very little flexibility.

And as the program surface starts to increase, and with that configuration you start to incorporate environment variables, flags, and eventually you wanna start to check that into version control, so you turn it into a configuration file, maybe Yaml, JSON, something like that… And then as you kind of turn on this heptagon of configuration - and I only call that heptagon just because a lot of the ideas came from Kubernetes, and the Kubernetes logo has got the seven points, and it just kind of worked out well…

David Berenstein: So data collaboration, at least from our point of view, is kind of the collaboration between both the domain level experts that really have high domain knowledge and actually know what they’re talking about in terms of the data, the inputs and the outputs that the models are supposed to give within their domain… And then you have the data scientists or the AI engineers and this side of the coin, that are more technical, they know from a technical point of view what the models expect and what the models should output. And then the collaboration between them is now even higher, because nowadays you can actually prompt LLMs with natural language, and you actually need to ensure that both the models actually perform well, and also the prompts, and these kinds of things. So the collaboration is even more important nowadays. And that’s also still the case for [unintelligible 00:07:11.15] models and these kinds of things, which we also support within Argilla.

David Beale: Yeah. I think my sort of neuro ADHD anxiety stuff is like two sides of the same coin. And if I’m excited about something enough, I’m going to get so anxious about it that I actually start to feel awesome. And that’s kind of where I’m always trying to get. I don’t know, there’s always going to be cool people in tech. There just is. So if you can ask for help, get out there, build a network, talk to the people you know… Sure, it’s hard right now. It’s a hard market. Salaries got really, really high, I think. And that’s a big thing, that a lot of companies are just dumping people and trying to hire people cheaper. That’s what’s happening. And that’s why it’s hard for people with 20-year careers to get a job; they want those big salaries. Whereas that 24-year-old kid straight out of bootcamp, or CS degree, or whatever - they’ll take half that. And I think that that’s why we’re seeing a big shift, and so many riffs, and all this sort of stuff. But you know, it’s just always reinventing itself, and taking shape… And VCs are starting to really invest again. I’m seeing that. I think the market’s picking up.

Ryan Dahl: I mean, it’s worth pointing out, first of all, that Deno is MIT-licensed, so it is essentially public domain software. It’s completely free. It’s not a commercial product. Our commercial product is in kind of cloud hosting services. But yeah, when I was working on Node, there was a time where it was very clear that Node was taking off, and I could not hire any people to work on this. And that was a ridiculous situation, because the company I worked for saw this as some sort of marketing effort, when in fact this was the most important thing that that company was undertaking… Back in 2010 we didn’t have the same kind of funding situation that we do in these days. And building software like Node or Deno is an expensive undertaking. Software engineers are expensive, it requires a lot of time and energy to work on stuff.

And it’s good to have a clear direction, and not just herding cats. You move very slowly when you rely on external contributions, and it goes in different directions, and… Yeah, I have certain software that I want to build, and I want to push it in a certain direction, and this is software for the masses too, that is not so far away from all sorts of business concerns. So I think it’s totally reasonable to have a company around this, and I think it’s kind of the right incentive structure as well for building open source software.

There’s two sides of the same coin… There’s open source developers out there that will complain about how they have to work for free, and they have to deal with issues, and random people writing in… And then there’s other people out there complaining about VC-funded open source projects that “Oh, God, what’s kind of the motivation behind this? Are they going to do something sneaky here?” I think these are the same problem. Like, we need to get paid to write software, we’re building open source public domain software… MIT license is very, very free. People can fork it, they can sell it, they can do anything with it.

I think the important thing is to be honest and not change licenses out from under people. I think that is unacceptable. But beyond that, being able to build some software, release it for free, and then use that software and other situations for commercial purposes perfectly aligns the incentives… And I am fortunate enough to be in a position where we were able to raise money to work on this.

Ryan Dahl: I mean, it’s worth pointing out, first of all, that Deno is MIT-licensed, so it is essentially public domain software. It’s completely free. It’s not a commercial product. Our commercial product is in kind of cloud hosting services. But yeah, when I was working on Node, there was a time where it was very clear that Node was taking off, and I could not hire any people to work on this. And that was a ridiculous situation, because the company I worked for saw this as some sort of marketing effort, when in fact this was the most important thing that that company was undertaking… Back in 2010 we didn’t have the same kind of funding situation that we do in these days. And building software like Node or Deno is an expensive undertaking. Software engineers are expensive, it requires a lot of time and energy to work on stuff. And it’s good to have clear direction, and not just herding cats. You move very slowly when you rely on external contributions, and it goes in different directions, and… Yeah, I have certain software that I want to build, and I want to push it in a certain direction, and this is software for the masses too, that is not so far away from all sorts of business concerns. So I think it’s totally reasonable to have a company around this, and I think it’s kind of the right incentive structure as well for building open source software.

There’s two sides of the same coin… There’s open source developers out there that will complain about how they have to work for free, and they have to deal with issues, and random people writing in… And then there’s other people out there complaining about VC-funded open source projects that “Oh, God, what’s kind of the motivation behind this? Are they going to do something sneaky here?”

[20:01] I think these are the same problem. Like, we need to get paid to write software, we’re building open source public domain software… MIT license is very, very free. People can fork it, they can sell it, they can do anything with it. I think the important thing is to be honest and not change licenses out from under people. I think that is unacceptable. But beyond that, being able to build some software, release it for free, and then use that software and other situations for commercial purposes perfectly aligns the incentives… And I am fortunate enough to be in a position where we were able to raise money to work on this.

Jorge Castro: …the humanity back into it. And bring some of the little humor. I have this coin that [unintelligible 00:45:21.17] gives to people… It says “The strongest steel is forged in the fire of a dumpster.” I just go “Hey, welcome to the project. It’s a hot mess. Do you want to hang out and talk about Linux?” And that’s really the value I think people are looking for. They want to see someone, and they want access to someone.

Daniel Whitenack: So Dinis, you said something that was pretty intriguing to me, which I think is maybe a kind of distinction that is maybe interesting to draw out and get your thoughts on, which is this idea that when you think of kind of the intersection of cybersecurity and AI, you could come at it from two perspectives. So you could come at it from “How can we use AI to help us in our cybersecurity tasks, or to create new tools for cybersecurity”, and then the other side would be “Well, how do we operate AI systems in a secure way?”

So there’s probably some interaction between these two things, but could you give us a sense, from your perspective as an expert in this field, and also seeing a lot of things so far, how do you see the kind of maturity of these two sides of that coin? Anything you’d want to highlight on either side of that in terms of how both things are progressing, at least at the state of where we are now?

Justin Garrison: There’s two sides of that coin, of just like “Well, you’re in the door, but we want you to fix this box.”

Taylor Troesh: I want to take this time to coin a phrase. I think we should be calling self-driving cars - we should be calling them autos.

Ryan Worl: So the reason why people raise money is – let me put it a different way. The right reason to raise money is that you want to go faster. That’s basically why someone should raise venture capital; it’s they have something that’s working, and they want it to go faster.

My co-founder and I had so much conviction in what we were doing in terms of it being commercially successful that we knew on day one we would be able to go much faster if we raise money. So that’s why we did it. There was never a period of time where we were guessing, like “Oh, do people need this?” It was very obvious to us from day one that we wanted to go as fast as possible. And raising money is the way to do that, because we were able to hire a lot – you know, relative to the two of us, many more people, and pay them very well, and make them happy, and… You know, hiring people that are good at distributed systems stuff is very expensive. And those type of people also really appreciate job security. So being able to have a bunch of cash in the bank, even if we’re not spending it, is very important to those folks. So our internal stakeholders, as employees and founders and stuff, it makes it very comfortable to have that cushion, and it allows us to hire people that will make things go faster.

And then on the complete other side of the coin, if you want to sell products to enterprise buyers, as two people, without having raised any money, it’s going to raise a lot of eyebrows if they want to put that in production as the backbone of their multi-billion dollar business.

Jerod Santo: Dance monkey. [laughs] “I want you to do more. I’m gonna put some money in the coin slot.”

Adam Lisagor: it could go really, really well if it’s like Square or Robinhood, but it can go really, really poorly if it’s like the coin card that everybody paid money for, and then just kind of [unintelligible 00:05:58.05]

Tejas Kumar: Yeah, that’s a good question. I just follow a number of experts’ definitions of this thing. I tend not to try and coin terms myself, mainly because I’m just not very credentialed, if we’re being honest… So how do I see agents? I summarize it, I summate it - I’m trying to find the right word - I deduce it from definitions from industry experts who have done it before me. So people like Andrew Ng, the founder of Coursera, and now the founder of DeepLearning.ai - I think he’s got some great content about this, where he defines agentic workflows as workflows that have LLMs perform three tasks, either all three or a subset of them. And those are reflection, meaning generate some output and reflect on it, “Is it good, is it not?”, and then iteratively work on it until it cannot be improved further. So there’s reflection. There’s tool calling, as I mentioned, with RAG, where the large language model will, sort of like a human being, recognize – for example, if you ask me to do a complex calculation, like 324 divided by 9 times 7, I’ll just be like “It’s time to get a calculator.” I’ll recognize that this is the sort of boundary of my capabilities, and go use a tool. So number two is tool calling. And number three - I think it was agent collaboration, where you have – yeah, it’s LLM as judge. It’s this model where a capable model (pun intended) coordinates lesser capable models towards an outcome you want. So it’s like GPT 4.0 being the most capable of Open AI’s models would orchestrate like three or four different 3.5 turbo models that are doing various tasks or generations. And so those three, either one of them or all of them, make up according to Andrew Ng an gent workflow.

[05:47] According to David Khourshid, AI agents are an implementation of the actor model, which is just a programming model where you have an entity called an actor, that sort of acts in response to observing its environment. So the classic implementation of the actor model is PacMan. Actually, a great example of AI, but rule-based AI, where the rules are known ahead of time, is PacMan, where you have PacMan, the little yellow pizza thing, and it’s observing the environment: where are the ghosts, where are the cherries, where are the dots… And you as the player take on the role of the actor. But there’s also demo mode, where the actor model is in play. And according to David Khourshid, this implements agentic workflows. However, it’s rule-based, it’s not generative, but it’s still an agentic workflow, where PacMan is an agent.

So I just take a mishmash of those two - these are the preeminent leaders in the space in my mind - and marry them, and that’s the working definition that I have for an agent. So it’s not a sentence, it’s not a nutshell, but I’m trying to give you more sort of a broad framework of how I see agent workflows. I have seen this term abused, where people will build – maybe abused is too strong… But people will build a custom GPT; this is a feature you can use from Open AI’s GPT 4. They’ll just build a custom GPT, add a system prompt, add some knowledge that GPT 4 can do RAG on, and call this an agent. I disagree; I don’t think that’s an agent, that’s just a RAG application. It doesn’t really do any of the things like we talked about: reflection, tool calling, collaboration, or observing an environment and responding accordingly. So I’d say those four tenets make an agent an agent.

Paul Copplestone: I could be wrong. But because there’s another – this is a two-sided coin. The ex founders are the ones who appreciate our product-market fit more than the rest of the company, because they’ve been grinding on their own companies for the past however long; they know exactly how hard it is to get where we are. And this is not necessarily credit to [unintelligible 00:40:32.12] and to myself. A lot of it is just sometimes luck. We captured lightning in a bottle, and the timing was good… I could have launched this business five years ago, and it wouldn’t have done well. But there’s a lot of luck that played into it, and because of this, they see our traction, and they know that it’s a good thing. It’s not just something that they can reproduce themselves.

Mat Ryer: Yeah, it’s the opposite side of the same coin.

Justin Garrison: Thank you so much, Bailey and Taylor, for coming on and talking to us all about just WebAssembly in general. All the stuff in the ecosystem, what wasmCloud does, and how it’s kind of a difference… And in correcting me on my assumptions around like “Is Kubernetes and WASM kind of different sides of the same coin?” I like being wrong, but I also like having analogies that people can relate to, so…

Adam Stacoviak: Yeah, thinking here on the fly on the podcast, it’s like podcasts as negotiation; like, we just – we need to revisit the Fly conversation, basically, and see if this is interesting to them. Because I think there’s two sides of this coin from my perspective. I guess three. One, it’s super-cool. BA. This is awesome. That’s my abbreviated version of bad-something-else, to keep our show clean. Then two, I think, should we build it, and should we maintain it - what is the maintenance burden we’re taking on to build and deploy it? And is it literally just enough in maybe a couple iterations to be exactly what we want?

So should we build it, should we maintain it? That’s one question. Then the other question is can we have a partner love what we do enough to sponsor it or to bring it into the fold that we’ve already got? It seems like the first one’s a yes, because it is awesome. And the third one’s a possibility, because it’s within the range… And I think the middle one is probably one to maybe spitball right here, which is - you’ve built most of it. Is this 80% of it? Is there 20 more percent, and is that the hard part? Like, how close to a version of done is this?

Jerod Santo: So two more of your BDFL principles seem like two sides of the same coin. One is to stay on the bleeding edge, and the other one is to keep up with the world. And they seem to have different angles, but they sound very similar. Can you explain those two?

Jerod Santo: Right. But the person whose name gets guessed gets a point. Don’t think about it too hard… Flip a coin and pick a name.