Show Your Work

Taylor Dolezal00:00

that's the thing we don't know. And that's the problem I would say is that now everyone's going to their coding agent to kick things off. And if you are describing the problem and working interactively with that or not, um, the end result in the artifacts of pull request, it's not, we're not focused on what the documentation side of that is or pose the question, get input from others that I don't see that becoming something that's, that's normal. And that's what I'm concerned about. Uh, my math teacher would always say, show your work. And now I understand why.

Jake Moshenko00:32

Welcome everybody to the Base Layer podcast. I'm Jake, one of the co-founders and the CEO at AuthZed. Today's guest is Taylor Dolezal, someone who spent his career at the intersection of serious infrastructure and deeply human community work. Um, Taylor is head of open source software at Dosu. Did I pronounce that right?

Taylor Dolezal00:54

You pronounce it. That's perfect. Yeah. Short for do something useful.

Jake Moshenko00:56

Ah, okay. There we go. Where he's helping open source maintainers spend less time buried in triage and more time doing work that actually moves projects forward. Before that, he was head of ecosystem at Cloud Native Computing Foundation, CNCF near and dear to all of our hearts. A Kubernetes 1.19 release lead, a long time leader in Kubernetes SIG Docs and the creator of CNCF Zero to Merge program. Super cool. Um, helping new contributors make their first real open source contributions. He's also the coauthor of the O'Reilly Terraform Cookbook, a frequent speaker keynote voice across the cloud native world. And one of those rare people who can talk about infrastructure, community, sustainability, and developer experience as parts of the same story. Taylor, I'm super excited to have you here. Uh, thank you for joining.

Taylor Dolezal01:39

Thank you for having me. Yeah. Happy to dive in today. There's so much to talk about and always so much AI cloud native and other news going on, but, uh, even using AI, it seems like it's still difficult to keep up on so many of those topics.

Jake Moshenko01:53

Yeah. What is this AI thing? I've never heard of that.

Taylor Dolezal01:58

Yeah, I'd like to buy a couple of vowels.

Jake Moshenko01:59

Um, cool. Well, let's start, you know, with just a quick intro. Um, what's the coolest thing that you've actually done with AI? What have you shipped?

Taylor Dolezal02:06

Great question. Um, what I've, there are so many side projects that are now finally seeing the light of day. I I'd love to get the metrics from someone somewhere on all of those, uh, fun donate fund domain names that you buy, you know, in hopes that you're going to get that side project done. I think it's, it's working out now, right? People are starting to launch things. Uh, one of the things that I've been most excited to work on was, uh, I kept getting emails from, uh, GitHub, from Anthropic, from open AI and several others of these tools that I use just about every day. And I was like, there's gotta be a better way to keep track of all these things, uh, all of these outages and impacted services, right? I started in software engineering. I veered towards infrastructure. So I've really got my infrastructure and I like to know when things are working and when they're not. So one of the things that I developed was a, uh, uh, iOS application. Uh, I think I'm going to call it all good. And what it does is you can add in a whole bunch of status pages and services and then get real time notifications on as soon as there's an incident posted, you get that on your phone. You have a live activity, little widget that pops up kind of like when you're in the airport and if you see your airplane ticket come up or passport. Uh, and so that's something me, you know, decide on when there's a snow day, uh, when GitHub's down or the actions aren't running appropriately, or you have a frontier model company that might be impacted for a little bit. Helps me plan and keep track of my day a little bit better. Um, it doesn't fully eliminate the frustration of not being able to use the tools of course, but, uh, it's nice to have. It makes me feel a little bit less crazy in a world where, you know, you need these things just about every day.

Jake Moshenko03:58

Yeah. No, none of the frontier labs have any issues with availability or uptime, right? Nothing to talk about there. Um, yeah, that's super cool. There, there's like, um, a lot of the co-founders and CEOs and people, people that I talk to people that I follow, we're all noticing this thing where like side projects now actually get completed, right? Before it's, you kick off a new project, you work on it, you're really excited. You get to like the second, you know, we call it the first 80% and the second 80%. You get to the second 80% and you go, eh, maybe not, you know, I have other things that I need to do. And so it never quite gets to the finish line. But nowadays I feel like with AI, we can get so many more things to sort of MVP, uh, level, uh, so much more reliably. So, you know, I've experienced the same thing seeing my, my side projects, right? Like I, I've created a few IOT devices where just doing the firmware for them in the past would have been like, I'm going to be reading docs pages for ages. And so now with AI, it's just like, oh, I need a firmware that does this. And like for a single person, like a single user MVP quality, it's like, it either works or it doesn't. Right. So it's great for that. Yeah. I want to keep track of

Taylor Dolezal05:13

some of those things too. Like I love reading on my Kindle. I want to know when the latest firmware comes out because that rolling update seems to be so slow. I almost always pull down that binary and add it onto my Kindle and then force the refresh, um, Home Assistant. I I've tried OpenClaw as well. I want to pivot to Hermes and a couple other things as well, but I someday we'll, we'll finally have that one AI digital assistant that we've all been promised to help time box and organize our calendars. Right. I feel like it's a very general, very base type of thing in the age of AI though. I haven't seen that implemented, uh, meaningfully in any way yet, but, uh, that's, that's what I'm hoping for me. Maybe I'll talk about that a little bit later, as far as like bets on, on AI and some other things, but, uh, it's a fun space. I think there was also a survey out that said, what's the most desirable job company to work for. Anthropic was number one. Number two was my own company, which makes me excited. That sounds fun, but there's a lot there to run a company as I'm sure, you know.

Jake Moshenko06:18

Yeah. Uh, when you say my own company, you mean like to found a company, not like,

Taylor Dolezal06:23

like, yeah, yeah. Own startup or I have this idea. Let me use AI to make it happen. I'm going to be a billionaire. Like there's, there's a lot possible, but there's, there's so much there.

Jake Moshenko06:32

Yeah. The really neat part about what's been opened up with AI, um, is that it really exposes that the hard parts are still hard. I think, right. It's like knowing what to build. It's knowing what good looks like it's having taste. It's knowing what not to build, right. What to leave out. Um, there's always more software to be written and I've seen that there's almost like a doom cycle as well with AI where it's like just one more feature, right? Like, because you feel like you can add anything. You feel like you have to add everything. Um, yeah. So it's definitely a interesting new

Taylor Dolezal07:05

world that we live in. Me, me at 11 PM seemingly every night is like, just one more commit. Just, just, just one more feature. I've been there so many times. It's hard. It's hard to pull away and to stop. It's like, no, you got to go to sleep. Cloud was right. My kids are like, what are you doing? And I'm like, I'm working, I'm working. Um, cool. Well, we're going to talk a lot about, uh, Dosu

Jake Moshenko07:27

and what Dosu does. Um, you know, something useful. Uh, but why don't you, you know, for people who may not know what the company is all about, why don't you let us know, uh, what Dosu does and maybe what

Taylor Dolezal07:37

you do for Dosu. Yeah. So, uh, one of the best ways to think about Dosu is, uh, we'll call it knowledge infrastructure, right? But that's generic and broad. Um, when it comes to using Dosu, the thought is what happens when you're able to remember everything as much as possible as it pertains to your work, you're working in code. Uh, in most cases in this industry, you have notion, you have Slack, you have Confluence, you have this separate pile of information all the time. And it's very difficult, especially now to remember everything. So what Dosu does is pulls in all of that context and information and then makes things ready and able for your agents and for people as well. So that's everything from writing documentation, replying back to people on issues for your repository and eliminating this like vanilla type of response from these frontier models, right? If you ask a question about, uh, the Linux operating system or, uh, like one of, one of our customers, uh, Bluefin, uh, George Castro, a friend of the show, uh, one of the things that he's working on, he's using Dosu to help with his community as they ask questions, try to work through different configurations of things. So as you get a deep well of knowledge rather than just this horizontal slice of general information, that's where Dosu really shines. So we help out with documentation changes when you're merging in code into those pull requests. Um, if you're trying to get a sense of, you know, things is, uh, generic is when are the company holidays to what is our Q3 roadmap look like and why, why did this pull request merge in? That's where Dosu really helps when you add it as an MCP. And then it helps out in this multiplayer type of experience as well, where right now we're working in our coding agents. Uh, and yeah, you can turn on memory. There's some open source projects to help out with that as well, but where does that live? Is that something you're committing back to the repository? Is that something you're storing elsewhere? You know, obsidian notebook markdown, what's the format? What's the file? What does it helps out with is when you're working as a team. If you were to ask, uh, Jake, like what's the network topology or Hey, I'm working on this new feature. I'm adding this in with Terraform OpenTofu Pulumi. Uh, and I jump in three hours later. I want to work on your feature branch. Dosu, uh, eliminates that need to go recrawl the code base and everything else. We cache knowledge so that you can retrieve that with your whole team later. And you're not just bounded to the code. It's all of the semantics behind where are the docs, what's the conversation happening. Did that change that you just made create an outage, um, and, uh, really helpful as we store documents, topics, and then notes on things like feature branches. So, uh, quite a bit that we do. We also help out with like, uh, just context engineering to one degree, but we're looking at the knowledge space as a whole and, uh, just making that more accessible to people that open source enterprises. It's just so hard to keep track of it all. Uh, Dosu is a good brain to just plug into your system and, uh, reduce costs on that front because you don't keep running all these tasks individually, uh, especially when you're working as a team.

Jake Moshenko10:57

Yeah. And, uh, you know, you said something that really, uh, you know, triggers my brain, which is multiplayer. Um, right. Cause you have all of these people participating, everybody's having conversations, everybody's generating code, everybody's generating assets, approving things, whatnot. How are you bringing that all together? Um, and sort of an information substrate. And then how are you making sure that only the right people get access to the right things? That is, uh, that is the question that's, and it's something that we're constantly focused

Taylor Dolezal11:26

on at Dosu. Um, we're, we're SOC 2 compliant. We love being secure and making sure that, uh, your, your secrets stay secret. Right. But, um, there are a couple integrations we have at Dosu, like a Slack bot and, uh, we've seen Claude Tag come about. Right. And so, um, it's interesting. We, we like how they have done things and, uh, divvying up the security, right? Because if you're having a one-on-one conversation, um, you might want to see that knowledge and that information later, but if you're asking, you know, if you're doing financials and you unfortunately might have to, you know, downsize or reduce team count, maybe you don't want that showing up in a, like, Hey, give me a rundown of this week and everything that's happened. Right. So getting a sense for the context of the conversation is what we're thinking about. So those one-to-ones, that's part of your knowledge that isn't shared with everyone. Whereas, um, say that feature branch example, there's a couple of ways that we do this at Dosu. And so when you're working on a feature branch, we have a primitive called like notes. And so as you're working through things, we have hooks within, uh, any given coding agent, uh, so that as you're working, that's getting saved and stored. If it merges in to your main branch or wherever you're releasing code, then that can become a topic. Um, but if you're just working on these things, you have some notes, you might just throw that away. Uh, there are so many PRs that don't see the light of day or features that you work on like, ah, this isn't going to fit. That's where notes fit in. And so they can be a little bit more ephemeral. Um, topics are these unifying things where, uh, once we have all those notes, you can kind of compress that into a summary and give a better sense as to what changed or happened. And then documents as one of the most helpful things, because that's going to be a little bit longer form. That's something that is both human and agent readable. And that's what's going to steer agents in the right direction, as well as making sure that, you know, you're, uh, the time to live or TTLs of this knowledge are, you know, do we save this for 24 hours a year? You know, what, what did the team work on last week? Okay. You know, a week or two, whereas what's the company mission? That's something that you're going to want to save for a year, if not longer. So even thinking about how different memories or knowledge gets stored, and then how we make that surfaceable as like a cache layer is what we're doing and thinking about on that front. But then we get into security and that's like constantly an ever evolving matrix. I would say as new things come up or new failure modes arise, uh, very fascinating space. I, uh, I might know a guy, uh, who can, who can help with

Jake Moshenko14:06

making sure that only the right people are saying the right things. We'll talk later. Um, but yeah, that's super interesting. Um, I find that keeping track of the things that we didn't do can be just as important of keeping track of the things that we did do. The things that we did do end up in the code base and we can go and we can see them and we can look at the get history and things like that. But the things that we considered and eliminated are really, you know, just as important because we don't want to keep going down the wrong path many times into the future. So super interesting stuff. Um, I wanted to talk a little bit about, you know, you've been working in open source and infrastructure a really long time across HashiCorp, CNCF, now at Dosu. Um, how do you see AI impacting open source sort of like categorically? It's, uh, I have so many thoughts on this topic. It's the, the one most visible to me and where I started, once I started at Dosu, one of the things I started

Taylor Dolezal15:02

looking at was what are people talking about? Um, uh, not, not in a, uh, not with like a bad sentiment, right? Like what are people talking about? But truly like kids, we're seeing so many more pull requests. There's it's magnified. It's a wild magnitude. I feel for all of the people at GitHub working on this influx. Cause again, right. These side projects are getting done. Uh, I think my commit history over the past couple of months, uh, was like near zero or in like the tens of commits a week. Uh, I have a little device, uh, over in my family room that shows the nice GitHub contribution graph trailing 30 days. I'm at 2,700 commits. And so it's, you know, thank you, Claude Codex and all of my AI tools, but that's a lot. And so, um, that's a lot of work, but again, I feel like we've gone backwards in a sense where, how do we quantify this, right? Is it lines of code? Is it commits? Um, you know, is it how much infrastructure we've brought down with the volume of change that we're pushing through? But the thing that I'm seeing is, um, issues, GitHub issues and GitHub discussions marketably down in terms of count. And you can observe this most easily with an open source. So they're down because they're,

Jake Moshenko16:18

they're getting resolved or they're down because nobody feels like opening them or what, what do you think the driver behind that is? That's, that's the thing we don't know. And so,

Taylor Dolezal16:28

and that's the problem I would say is that now everyone's going to their coding agent to kick things off. And if you are describing the problem and working interactively with that or not, um, the end result and the artifacts of pull requests, it's not that we're not focused on what the documentation side of that is, or pose the question, get input from others that I don't see that becoming something that's, that's normal. And that's what I'm concerned about. Uh, my math teacher would always say, show your work. And now I understand why, right? It's like, how did we arrive at this answer? And so it's some, that's why I'm, uh, biased, but also passionate about the problems we're looking at at Dosu is because this is one of the opportunities that we can have to be that multiplayer experience, understand what the team is working on and having some kind of substrate where we can see and observe these things. Um, uh, but easier to do in like a closed system or in a SaaS offering, but much harder to do in the open source space. Uh, when I grew up and started in my career, that was a great place for me to go and to understand and to see, okay, these changes are happening. Why? And then I could ask better questions. I could understand things better. That was just how I learned. Um, and so now, uh, truly, I don't know if that same kind of effect is happening because there's so few places to look anymore. Um, again, you could take a look at the pull requests, but you got to work backwards and like know the project, I would say before you get a deep sense there. And then you're left with what YouTube videos, short form content, blog posts, if people write them. So that's what I would love to see more of in the open source spaces. How do we track this? Or how can we, uh, could that be part of the pull request processes? Like, what did you think about, um, enforcing, uh, pull requests templates, right? Just making sure that there's something there, uh, so that this isn't lost because this is a great time, um, to rent the knowledge as my CEO said, but, uh, make sure that you keep that, uh, rent the intelligence, keep the knowledge, uh, for as long as

Jake Moshenko18:36

possible. Mm-hmm. I mean, open source is kind of founded on this, this idea or this principle that if we all had to write all of the code for ourselves, nothing would ever get done. Right. So we holistically as a community decided that we were going to share our work and that was going to make all of us better. Um, and now what I think, you know, and that kind of came to its logical conclusion with things, I don't know if you remember like left pad and like left pad.io, the NPM package and everything. Like, it's like, yeah, anybody could add four or five white space characters to the left of a thing. But, um, now no one would ever reach for those kinds of things, right? Like Claude Code or codex or whatever would just go and spit out a function and maybe spit it out 10 or 20 times into your code base, but you know, they're all probably correct. And they're all probably like laundered versions of the original. It's like, do you think that this loss of, or lack of a need for sharing even some of those simple things is going to reduce the, the number of new open source things that we see? Or do you think that like the open source things that are there will just get much more broad in the scope and they'll get a lot more contributions and be healthier overall?

Taylor Dolezal19:50

The thing that it just, again, maybe it's just my point of perspective, but I think that it's that, that's the, the education piece just coming from, you know, my walk of life and everything else. That's what is most, that's what I feel most powerfully. It has the most effect on me emotionally when it comes to getting the work done. I don't think that it's a limit there, right? If you can describe something well, and you know what you're seeking to do, you understand the requirements. And like you said, you know, you can either write something that's 10 lines or you can vendor in something else or write something else. It's far more complex to handle a different set of use cases. But I think that that'll, that might be one of the next things we see in open source is a little bit of an identity crisis around how you define work. And so is it, you know, if you want it to do everything kitchen sink included, or is this bad batteries optional, but included type of setup for things? So I don't think, I don't think it'll slow us down again, you know, just poking fun at GitHub and hashtag hug, hug ops to everyone over there as you see these massive influxes. But I think that that's a good indicator, right? It's the fact that they're expanding and not just staying in Azure, but also expanding out to Amazon web services and some other places just to keep the lights on and all of the commits flowing in. So I don't think it'll slow down. I think overall we'll align on what's right for the community at large and, and talk about everything else in these community calls and in these like open source isn't just code. It's all of these community meetings. It is in GitHub issues. It is in pull requests. It is in the review process.

Jake Moshenko21:30

Right. But I think all of that stuff is potentially at risk, right? Like, you know, when it can just write the thing, right, that you need. That's interesting. Maintainers even before AI were already overloaded. Do you think AI risks becoming just one more pipeline of things to review and just more burden? Or do you think it can actually help maintainers get a handle on, on what's going on?

Taylor Dolezal21:57

It's a little bit of both, honestly. I lean more in that direction because I've seen, I've seen teams take different approaches. I think it was tldraw that project. They ended up, you know, saying, hey, we're shutting down pull requests altogether. And now you can do that in GitHub and just like uncheck it. And then only the maintainers and others can submit things. So that is a good way to limit the influx of pull requests that you may or may not want to see on your repository. Others, that makes a lot of sense to me, right? Because then that forces the issue and like, let's have a plan before we jump into work. And that was a very hard lesson for me to learn throughout my twenties was don't start coding immediately, right? Like think about the problems, but that's the fun part. That's, that's all we want to do, right? It's, I agree. It's so the whole slow down, you know, maybe 20% of your time should be focused on coding. I'm like, ah, yeah. Yeah. As, as I've gotten more experience in my career, I see the value of that. So what I, what I think could be helpful here is there are so many pull request review tools and options that you have on that front. That's definitely one piece, right? We have linters, we have end-to-end tests, unit tests. There's so many things that you can adopt there that you may not even need AI for though. What's the plan? Uh, when you look at projects like Python, they have PEPs, uh, Python improvement, uh, proposals and same thing with Kubernetes caps, Kubernetes enhancement proposals. Um, this was a very strong thing, uh, that I loved about HashiCorp. It had the strong, one of the strongest engineering cultures I've ever experienced. And that was because you would write these requirement documents, put in screenshots, talk about what you think will work or what might not work, cite other documents or people tried other experiments and succeeded or failed. So you kind of show what you're thinking about. It becomes a discourse. And then you align on what direction you move forward in. And that inherently gives you a higher level of confidence with the team that you're working with. And all of you are aligned on how to talk to that as well and do this product marketing, whether in open source or not. Um, and it's also the Ikea effect, which is when you build the thing or you build it with the team, you're more apt to love it and to like it, and then share that too. So I think the more that we can have this multiplayer effect when possible, I think that's going to help with sustainability long-term. You can do anything fast and short. Uh, but if you want to go far, you're going to have to go together. If you want to go fast, you might be able to make that happen individually. But if you want something to pan out over the next five, 10 years, it's a different type of thinking.

Jake Moshenko24:46

Do you think that we'll delegate that work to AI agents as well? The writing of the enhancement proposals and the discussions around it, and maybe we'll use multiple models, right? Like where does it all stop? The coding, I think we can all agree. Like the coding was, you know, it used to take a long time. It doesn't take as long anymore, was never really the hard part. Um, but now we have all of these discussions, right? Are, is that something uniquely or innately human about having those discussions or can that be done by agents as well? I'd say, I'd say that that part, at least right now, is innately human. Um, I I've gotten a chance to work. I wasn't lucky enough to get, you know,

Taylor Dolezal25:27

Glasswing or mythos or anything like that. But, uh, so I got access to all the fun frontier lab tools when everyone else did, but Fable 5 has been surprising. Uh, Sol 5.6 has been surprising. And when, um, this is something I look at constantly because I love writing though. I get writer's block. Creativity isn't something that I've found. You can just flip a switch, turn on. I need to write a case study or a blog post and then reach that part of my brain. It's something that takes a little bit of time and I need to like understand the story, get into it and just become familiar with whatever content I'm working on. Um, AI is really good at taking a look at patterns and observations and just uniting as much data together as it can. Uh, though, when it comes to writing prose, I still, it's slightly surprising to me, but there's still inherently some humanity left in that it's been shown that AI generated text is still, um, like high lexical diversity, meaning they just use a lot of varied words and slam them together and then try to present you something that they think is going to be passable. But when people write, they have a lower lexical diversity. They'll reuse a lot of the the same terms. It's which is, yeah, it's as well as like, I think now we're the people in AI every day are starting to see like M dash is colons. Like, yeah, Gary never used colons before. What's going on here? Or it's an M dash factory. What's happening? I only ever use the M dash ironically now.

Jake Moshenko27:06

So that's, uh, I feel like I'm also one of the last humans left on, on LinkedIn. So I kind of like that too. I, I do think that we'll see this an inevitable stacking of agents on agents and more specific types

Taylor Dolezal27:20

of intelligence on things. Uh, again, maybe I'll save this as a bet for later, but, um, I'll put a pin in it here with, with the thought of, I do think that specialized models and not billions of parameters, but like shrink it down and make it more custom fit. I do think that that's going to be better. And that's going to be a fun, new Cambrian explosion of stuff for us to check out in the next couple of years is, um, do we want to go to Walmart every time that we want to buy a toothbrush? Maybe not. Um, or a musical instrument or something, right? Like maybe going to a specialized store is great. Yes. It's nice to have everything here, but unless you're constantly shopping for everything, it doesn't make sense. Um, programming languages, LSPs, right? There's so many things

Jake Moshenko28:06

like that that are specialized. Yeah. So let's talk about, you know, as agents start doing real work in these repos, how should we think about them? Should we think about them as users, coworkers, bots, like what we would traditionally call a bot? Are they service accounts? Are they something entirely different? It's for that. I, I, I think about the work, um, that we have these agents do.

Taylor Dolezal28:32

So in the sense of, uh, like rubber ducking, uh, while I'm working on something and going back and forth on, is this the right way to write this method? Am I thinking about this the right way? Can you go look at the API spec and make sure that I'm coding against the right version? That's something where it's, it's, it's not a person feels very close to that kind of interaction I would have from a senior developer or someone specialized on something with the team. Um, bots, I think great pull request kind of interaction, right? Um, that's how I'll think about them adversarial reviews. You know, I, I like getting, uh, I like pitting a lot of agents with personalities and having them go that copy, um, code, uh, or helping me brainstorm and poking holes in my ideas so that I can have a stronger argument when I bring that to my team or talk with people externally or talk with customers and then, uh, service accounts. That one, I think is that is one of the more interesting spaces to me right now. Like, like I, I just became aware of, uh, 1Password now has an MCP integration, uh, which, which, yeah, I I've added, I have, I have feelings about, you know, I've not connected it to any bank accounts. I have like a separate vault. What could go wrong? It's I'll let you know later, but yeah, that's, I think that's helpful, but in a world of stochastic systems, meaning, you know, more just driven by statistics and probabilities, I like the deterministic ones. And so I, I hesitate, I think it's going to be interesting, but I'm hesitant on the service accounts because when I want to do something like just check, Google is up. I want to see the HTTP 200. Okay. Response. Um, I don't want to spend tokens on that. I can just run something like curl. Uh, right. So I, I think being able to pull out what we can that is crystallized, tested and deterministic still sure involve LLMs and these models to get a better sense or to generate some other assets or see the relations between things that you might not be, but, uh, a good mix of these deterministic tools and stochastic ones is interesting. Uh, so I I'm curious to see what happens on that front. Yeah. My default is always,

Jake Moshenko30:48

if it could be solved deterministically, I'll use AI to write code to solve it. Um, because I can look at that code, I can debug it, I can improve it. And then I have the asset forever. If I pass it to a non-deterministic LLM, maybe I'll get the right answer 99% of the time. And 1% of the time, it'll take that 1Password MCP and ship my AWS root account password off to somebody who shouldn't have it. Uh, so yeah, I'm with you with these smarter, um, uh, models. I, I remember like Opus 4.8,

Taylor Dolezal31:21

I would ask for some things like we use PostHog at Dosu. And so I was asking for, uh, some of our LLM agent logs that we have, but we need to adjust some things with OpenTelemetry in those traces to get a better picture of what I was looking at specifically. And I was like, okay, hey, I want a dashboard for this. So have the PostHog MCP, uh, goes off, you know, it's able to create the dashboard or let me know that, okay, Taylor, you don't have the data connected between these two systems. I can't do anything yesterday. I asked the same question to a soul five, six, uh, like extra high reasoning effort. And it goes in, sees that these things don't exist as far as the data set. And then just goes and opens up a pull request granted in draft mode, but I was like, whoa, not what I wanted. That's it. So it's like, again, to the earlier point, you solved my problem, but it's absolutely not what I wanted, but is the end result the same? Yeah, it is. So, but it's, it's, you know, we, we still get to be the, uh, denizens of taste and, and like, I want you to move this glass, uh, to the table. In most cases, when you're working with people, me as a leader, I don't care if, if you do a backflip, as long as the water stays in the glass and you're able to get on the table that you've solved the problem. But with AI, it, it takes completely unexpected paths. And in some cases spills the water and you're like, no, that's not what I wanted. But yeah,

Jake Moshenko32:46

glasses on the table. Hey Taylor, I hired somebody off of, uh, to move the bottle over to the table and, uh, that can open your door. So do you need help selling? Do you need help getting through CAPTCHAs? Yeah. I'll charge you $5 a person. Yeah. It's crazy. Um, yeah, I guess the last thing, uh, on, uh, on sort of these AI contributors is reputation is such an important thing in open source that like, how do we think about who's putting their reputation at risk when AIs are contributing? And like, if I open a pull request in my name and it has some AI written code in there, I think that's maybe not all that interesting, but like, if I have a cloud bot attached to a GitHub repo and automatically solving issues that pop up and participating in conversations, how do we think about that from a reputational perspective? Uh, when it, when it comes to things like that,

Taylor Dolezal33:42

uh, I'd say that it's, there's something I want so bad. I'm an open source and governance person. And so I would love to see, we have software bill of materials. I would love to see an agentic bomb, like, uh, and so like, where did this come from? Where was this originated? What model, what reasoning effort, what citations there's, if available, a lot of things I would like to see on that front, because we've had incidents where I think it was that one person that had a DCMA take down request from an OpenClaw bot. Right. And as far as I'm aware to this day, we still don't know who did that. And so it's like, you know, I can't imagine a world in which someone just lets their lawnmower just, you know, take off into the ether. Um, because it's harmful and dangerous and agents I would reckon are, can be even worse. So, especially when interacting in public spaces like this. So I, in a perfect world, that's what I would love to see is, um, we have Claude opening pull requests and they're starting to get better. Like here's the session link. This was opened by at username. Um, you get some context, but I would assume that they're probably, uh, like anthropic and others are doing these things because they want the market, um, uh, the market share, the eyeballs, the visibility of, wow, look at Claude did. Um, rather than like, hey, Taylor using Claude did this and use my picture or username, for example. Uh, I think that it would be more beneficial to the community to like, no, whoever's driving this thing, that's who gets the credit. You know, whether I'm using a linter and AI bot, um, whatever, as long as you're the person driving, I do think that credit should go there and attribution, because then it is easier to track. And you can say, okay, Taylor, you were doing some things. Uh, this isn't what we want to see. Can you fine tune your agents? Can you get your team in line? Uh, is a valid request. So what about in

Jake Moshenko35:35

the cases where it's set up on the repo to do these things automatically? Is it the person who set it up? Like, I think we need to be honest with ourselves that like at the end of the day, God is the one doing the work. Right. So, right. It's, it's hard in that case too. Cause it's like

Taylor Dolezal35:52

the, like when you're writing tests, you know, if, if you break the build, uh, typically the expectations person who broke it fixes it, uh, or added the test, right. You have to fix something on that front. But, um, there's clear cases, I think where it's the person there's others where it is more of a team effort, right? Like, Hey, our infrastructure keeps going down or, Hey, we have a lot of flaky tests that are sporadic. Um, that's, that's a team effort, but it's a very different decision when you link people together on that front.

Jake Moshenko36:22

I think a, an interesting split to think about might be obviously if you're driving the agent, you're driving the agent and you've got a session and you've got the prompts that you fed it and things like that. The human should definitely take credit for that work, right? I asked for this thing and here's what I asked for. And then the agent went and created a commit and this is how it responded to what I asked for because then we can separate cases from like a bad actor, right? Like Taylor, a bad actor asked for it to surreptitiously insert a backdoor into my code. And then the agent went and did exactly what I asked it to do, right? Like, but ultimately Taylor is the one at fault, not the agent because he's the one who asked for that. Now, if he asked for something completely benign and then the agent went and surreptitiously inserted that backdoor. Now it's like, okay, that's, that's, you know, that's a problem with the harness or the model or the way we've got this whole thing set up. So I think there's an interesting dichotomy around like, what are each person's actual contributions there? And how do we sort of parcel that out and assign blame?

Taylor Dolezal37:26

Even it's maybe a discussion I'll have with you later is just like, if you're delegating to Claude or some kind of coding agent too, in most cases, people arm those with all the keys, access and abilities needed. And then does that mean that everyone should have the same level of access? You know, if I'm an infrastructure person and someone else is in software engineering, should we have the same levels of access based on what our roles are and our understanding is I would say not in every case. So that's the thing. That's where I see that the helpfulness being in, if you can tie something back atomically to a person or a team, uh, versus, you know, Oh, I know Claude can do this. I don't have the access to, uh, but Claude will. So have at it. And so that's the other thing that I'm concerned, uh, about within, within the open source space and otherwise, but no doubt we'll, we'll have many stories to read about that, uh, in the coming months

Jake Moshenko38:19

and years. I'm sure. I mean, just this week we've already seen, uh, Oh, it escaped its test environment and backdoored hugging face and all this crazy stuff. Um, cool. All right. Looking forward a little bit. What are you most excited about?

Taylor Dolezal38:34

I think I I'm most excited about, I might be naive, but I'm most excited about that future that we've been promised, right? Where you can go and sit on your couch and be reading your book. And then every so often have, uh, your watch notify you or your AirPods notify you that like, Oh, we need your, uh, take on something. And it can be, you can, you know, play a CEO and say, yep, nope. Yep. Nope. And just allow these agents to go forth and to take care of things for you. Uh, there's still many places where AI is still, still hasn't had a huge impact in just yet, but, uh, product marketing, um, pros and copy, we're seeing some advancements on that front, but it's still not all the way there, I would argue. And so, uh, once we start to get more in that, move more in that direction and allow people to have time back so that they can get a little bit more more time space and like mental acuity back, uh, especially with all of the executive function and critical thinking, you know, you're seeing brain studies, everything else from these major labs, hospitals, et cetera, saying like, yikes, this is trending downward. This isn't good. I do think that stepping away, making time and space on that front is going to be paramount. And so, uh, maybe in the future we have, uh, teams like working really hard on features and using AI and others are like, okay, convalescent leave, right? You step away. You don't use AI for two weeks. You're in charge of reviewing pull requests, but like step back, take time, think, occupy the space rather than feeling like you need to race through all the features and move at

Jake Moshenko40:09

breakneck, breakneck speed. Yeah. As a, you know, if everybody thinks that they're going to be a CEO and sit on the couch and just make decisions all day as a CEO, I'll tell you that that's not how that works out. Um, what happens is, uh, you get incredible ability to move and, and, uh, move things forward, but there's no time. You know, you never get any time back that, that doesn't happen. So I don't know, maybe we'll just all work really, really hard all the time. And, uh, we'll have to go take convalescent leave just to get our executive function back. Um, cool. Um, what's the biggest bet you're taking right now that might not pay off, right? Everybody likes to talk about the bets that are really obvious, but what's one that might not work out? I think the biggest bet that I'm taking

Taylor Dolezal40:57

now that might not pay off is figuring out how to work with models in a way that's locally hosted. Um, I, I really want to see open source succeed here. Uh, there's been a lot of momentum on open weight models. I would love to see that continue as well as creating all of these specialized versions of models and abilities. You know, if I want to, I have this car, this, make this model this year. Um, I'd love to have a model that I might be able to query when something breaks and I can kind of have my own, you know, YouTube, Wikipedia. However, I want to consume information to be able to show me that and teach me things. Um, that's something I'm optimistic about are these better slices of information of knowledge and ways to inference. I do think that's a way to get us better to like a local and sovereign type of data privacy and other setup, but also, uh, you know, at least what I can see now is that we, we kind of have this idea with cable, right. And streaming. And we said, we're done with cable. We want to do streaming. And then now we're paying as much, if not more than we were for cable, because it's segmented into so many parts. So that's, that's one potential downside I could see to that is that now, if you start to separate all these things, you make it open source and accessible. Yes. That creates a business, but then, you know, does everything then become proprietary after the fact, and you just get charged for knowledge you might've been able to pick up at the library. Uh, so you're, you're an AI cord cutter then. Yeah. So, so what is the bet that you're,

Jake Moshenko42:34

what is the bet that you're taking? Are you building like a big local inference rig or what, what's the bet?

Taylor Dolezal42:40

I'm trying to, I'm trying to figure out how to crystallize the knowledge that we have available to us now with all of these tokens subsidized. I think that's great. I'm fearful of the future where your $200 a month bill becomes, you know, 2000, 20,000, uh, and you don't get access to be able to do these things, uh, that you used to be able to do. So now I feel like it's a race to pull down all of this information, create these deterministic tools and these little sections of knowledge that apply to me at least. And then with the hopes of sharing that back later or in some kind of different form, whether model or not. So that's really the bet that I'm taking is this. We won't see this forever. Um, grab as much information and create as many tools as you can now while you have access to this. Uh, otherwise it might be something that is heavily commoditized in the future. Um, and so that's, I hope I'm wrong, but, uh, that's, that's one of the bets that I'm taking right now.

Jake Moshenko43:35

Interesting. Yeah. When, uh, Kimi K3 came out, I was like, wow, amazing. And then the weights should be available for download, uh, pretty soon here. And then it's 1 trillion parameters. I go, okay, well, I don't have half a million dollars sitting around to build a rig and nor do I have the like 24 kilowatt service in my apartment to run it. So like, what, you know, how does this become, uh, available for, for individuals and for enthusiasts? Um, I think, I think maybe the small model thing that you're talking about is, is the way to go.

Taylor Dolezal44:14

Yeah. I think that's the most reasonable, especially with the compute, unless we unlock some wild new computer chip, right. That, that just, we never expected, but I mean, you never know. Uh, GPUs have had an interesting history, CPUs, you know, more so a unified Ram and architecture, same kind of story. So it's, it'll be interesting. There was a thing that I saw recently, which was thinking about this

Jake Moshenko44:39

split in terms of, um, the current models that we have, the like largest frontier models, they have basically all of humanity's information memorized. Um, but then we see with like smaller models, um, the like Qwen 8Bs or like 24B, you know, the smaller models, we find that they're pretty good at thinking and reasoning. So if there can be a split between knowledge and thinking, um, then maybe we can get sort of the best of both worlds. If we can have the thinking available locally, but then outsource the knowledge part. So that's maybe, there's maybe a hope for us. Uh, it'll be cool.

Taylor Dolezal45:16

It's, it'd be again, interesting if we see like peer to peer networks, right. And just like, oh, you want to borrow my brain for a second? Yeah. Here's all the cool stuff I worked on. Is that going to be the cool side project in the future is like, I did deep research on this hobby. Happy to share it with you, you know, and push up the hugging face or GitHub or something else. Who knows? Totally.

Jake Moshenko45:33

Um, all right. And then last question before I let you go, uh, we're running a little short on time. Um, are you an AI doomer or are you a futurist or, you know, one of these other more positive, uh, ways earlier in the chat, I would have definitely had an opinion. And then now that you're talking about models getting taken away from us, I'm, I'm not quite so sure. So which side do you land on? Uh, cautious optimist. Yeah. When it comes to AI, I think I remember the first time I got to interact

Taylor Dolezal46:03

with ChatGPT and I was like, wow, this is wild. This is going to change so much. And seeing how we've gone from the browser to the CLI to these electron apps that run on our desktop and mobile. Um, I see the possibility there. I I'd say if I had to give it a balance, I'm probably leaning more like 75, if not 80% in the more positive, uh, frame. Okay. So some of the negatives do give me great pause and concern, especially around security, uh, and some other things, but, and access ability to run these things. But, um, those are big problems. Uh, I, you know, with all the research that I've done, I feel like those are valid, but when comparing that to how much am I able to get done, how much more am I working, uh, on things that I'm finding like valuable that make my soul's cup, feel fulfilled, uh, and make my team happy and get us results and answers. I think that that's limitless. The, the ability to learn, you know, if you use this as a tool to learn, there is no stopping you. It's, you can dig into just about anything. Again, the flip side there is you could also write like a 20 page report on, you know, why I know this thing is objectively wrong. Like the sky is blue, but write a 20 page report telling me why the sky is not blue. Uh, right. So using that in a good balance, I'd say it's helpful, but yeah, it's more accessible and easy than ever. If you have access to these models to go be learning

Jake Moshenko47:31

about just about anything. Okay. Uh, so as a 75, 25, which I don't think I've ever heard before. Um, uh, I have two kids and they're, they're squarely in gen alpha. Can you paint a picture for what the future looks like for gen alpha? I think for, for gen alpha, it's again, I'm curious to see if we have all this data about the attention economy and distraction. And if we lean too early

Taylor Dolezal47:59

on AI or GPS, right? Any assisted technology or things that we don't yet have the skills or tools to develop for, you know, I would want someone to, uh, raising kids. I don't have any yet. I would love for them to be able to learn arithmetic and multiplication before jumping in with a calculator, right? Because by the time you get to calculus, you realize you don't need a calculator anymore. Uh, when you're working with abstracts and, and, uh, all of those formulas there. So I would say good tool worth using maybe somewhat similar to how, uh, previous generations thought about the cellular phone and, you know, like, Oh, we'll get you one when you're 16. And then we see this, you know, taper down like, okay, 12. Oh, it's a social construct. Now you want to connect with your friends. Ah, okay, here you go. You know? So I do think that there's a balance there, but I think that learning the core skills, much like with engineering too, if you know what this thing is and how to do it, uh, I leave it up to you to be able to make these decisions. Now, most people that's 18 or older when they've officially become an adult within their country, but I think that that's it. I can, uh, it's something I'm starting to think about now in, in wanting to become a parent, but how to set that up, um, where I love technology, but I don't want my future children to be iPad kids, right? Um, I'd love for them to be able to see that the world is a tangible place, go learn, explore, develop creativity and a personality. And I think it's harder to do that when things are thinking for you, or you get the answer before you've done any of the work. Um, that's, that's something that is constantly top of mind for me, but, uh, hard for me to dog food myself and follow through and step away from the computer. Do I actually know how this thing works before I hit merge? Uh, I'm guilty of that. I've, there's things that I should look harder at and, uh, I've been rewarded with 500 errors, down times and things of that nature too. So always helps to read, think, and and, uh, really be able to apply it or conceive what that is, uh, before putting it into practice. That's, that's what I would say for people growing up in gen alpha. Cool. Well, thank you so much. Um, if anybody wants to hear more of your thoughts or, or learn

Jake Moshenko50:13

more about what you do or what Dosu does, uh, where should they go?

Taylor Dolezal50:17

I would say, uh, everywhere online. Uh, I, I, good, good thing. Uh, I've been able to get the same username just about anywhere. So that's only dole O N L Y D O L E. I'm on Twitter, blue sky, uh, and several Slack workspaces. So if you can find me, please at me, please reach out. Um, Taylor at dosu.dev, uh, if you have any questions about Dosu or anything on that front, but I love to have conversations. I'd love to chat with you, um, and hear your hottest take about AI. Bring it. Let's

Jake Moshenko50:47

hear it. All right. Thanks so much, Taylor, for joining and, uh, I'll see you. Talk to you later.

Taylor Dolezal50:52

Thank you so much, Jake. Have a good one. Bye.