Taylor Hello, welcome back to Blood Sweat and Tokens. Sean, what's up?
Sean ⁓ living a dream over here. How are you doing, Taylor?
Taylor I'm doing well. I'm I'm kind of I can't believe this is episode nine. I feel like we've only been doing this for a little bit and I'm really excited about today's conversation.
Sean ⁓ yeah. Yeah, me too. Gonna give it a little bit more structure this time, yeah.
Taylor Yes. Received some feedback from some listeners that were a little bit unstructured. ⁓ that sounds about right. That's pretty much my brain and personality. I have a s I have a hard time like staying on path. ⁓ that said, ⁓ we're gonna give it a shot today. So we're gonna go with a ⁓ we said kind of like a three-segment thing. We're gonna start out with a demo or at least a discussion of where we are today on the existing project. Before we get into that, let me just kind of recap what we're doing here. This is Blood Sweat and Tokens. It is a weekly podcast that is focused on product engineering in the age of AI. We're trying to figure out how to take advantage of tools like natural language, ⁓ you know, agentic workflows and things like that to not only ⁓ increase the level of productivity that we're seeing, but also have an honest conversation about the impact to humanity, the more junior engineers that haven't yet kind of amassed their, you know, professional bona fides, if you will. So with that. ⁓ Sean, how are we doing on the product? Could you give us a quick recap of where we are, what we're build or maybe start with what we're building and then talk about where we are and then ⁓ it sounds like maybe you've got some things that you want to talk about today.
Sean For sure, yes. So we we've been yeah, we've been building alongside these episodes here and have kind of taken a few cycles where we go down a rabbit hole, kind of turn it around, learn some things. ⁓ you know, this is this is not unusual for a product that you want to build out seriously in the early days. And what we did at first was try to take that approach. And we talked about this a little bit last week where we're gonna we're gonna support everything up front, you know, kind of like bring whatever you want and and it will work. And the thing that we're building is we can consider it new age content management. We want to the the goal here is we want to kill the CMS. Everybody hates their CMS. And we want to make working with content for marketing and other business teams to be smooth sailing and To the point where they can just fly, they can use their voice, they can speak like they would speak to an agency, get in and and and then immediately have some feedback and have a preview of what what their change request would look like before it goes to production. So it will take a long time to get to that perfect state. Over the last couple of weeks, I've been working toward basically rebuilding it from the ground up and with a very narrow focused, the so-called ⁓ opt into complexity model, where we're gonna take this very strict, narrow, happy path. And if folks want to opt out of that for some reason, that they can do that. And I've made really good progress. There are still it's unavoidable in this particular case to have. A lot of different moving pieces, and that's totally fine. We're narrowing in. And last week, if you go back to last week's ep episode, we talked about the shift in that architecture. So I won't go back over that. What I want to do today is pull out one concept that I think would be a little bit more interesting than kind of a half-baked demo. And that is in I want to talk about this from two angles. And it's it's just this concept of abstraction. And so for folks who are newer to creating products or creating anything for the web, you find yourself, I mean you can probably relate to v just life in general. You find yourself doing the same thing over and over again and you think, what can I do to make this easier in the future? And when it comes to building products, that can look like something like, hey Every time I build a new product, I have the need to upload files. So do I rebuild that every time and kind of iterate on that? Or do I build a tool that is a file upload tool and then plug that into all of these different products? Now, if we go back to the architecture conversation, one thing I want to pull out is that we want basically to have this core, right? So somebody wants to start using this product, even if they have an existing site, we're gonna say, cool, give us your site. We're gonna, we're gonna import it all into this new system. And you will still have control over that. But we want to start you from a base that is really quick and snappy. There's a set of components, there is an ⁓ there's a mechanism for being able to quickly edit, to pull up a widget, to just speak into the site. And all those things we want to work the same for any one of these websites. So what I was thinking was we don't just want to hand people the template as it exists today, right? We want to hand them that base level template. And then they're going to that template will consume a couple of packages that we create. And so one of those might be for example like the the mechanism for how you can edit a component. But that's that would take us down a rabbit hole. What I wanted to ask you about specifically here, Taylor, is finding this this balance in terms of both flexibility and configuration. So let's say that we're going to develop a shared component library that any one of these these websites can pull from and then they can bring their own components to it as well. If so here's the scenario for you. You create a new website, you pull in these components, this component library, and then you need you need a new component that seems fine, right? You just build it into that particular website. But maybe you just need to tweak an existing component a little bit. So how do you draw the line between you know not over-engineering the component library so it's It's too complex. ⁓ but you know, giving somebody the flexibility to not as soon as they want to change one little thing, they don't need to rebuild the whole thing. You know what I mean?
Taylor Yeah, I do. I mean you kinda talked about and two different ends of the spectrum. We talked really broad concepts around the philosophy of balancing flexibility and configuration, first configuration. And then on the other end of the spectrum, like the actual minute, you know, here's how you would actually do it with say a React component or something. So ⁓ I'll speak in broad general out generalities, traversing that whole spectrum, I guess.
Sean Great, great.
Taylor The first thought came to mind is d is DHH, d d David Heinemeier Hansen. So this guy is the creator of Rails and ⁓ 37 Signals and ⁓ I know he's got some newer products that are equally as awesome. He is a extremely opinionated person and he has brought a lot of those opinions directly into his framework and the way that he thinks about it. As long as you like follow the configura the configurable path. I say this with irony because Rails is all about like no configuration. But ⁓ what I mean is like if you are using a framework that comes with batteries included, like Rails, if you just do it the Rails way, it's gonna work. It's like opinion and action, right? And I I use that as the ⁓ kind of corollary to this concept of flexibility versus you know, because I see flexibility as you can put any of those components together and build it however you want, completely custom, versus the alternative where it's like, you know, ⁓ I'm using I'm standing on the opinions of the framework. I'm doing it the way that the framework would expect me to do it. And therefore there's benefit in that. Like the general philosophy is pretty powerful because as an engineer, we and i it's worse now today than it ever has been. We get stuck in the
Sean Mm-hmm.
Taylor ⁓ what the analysis paralysis stage where it's like, hey, I've got 12 different email providers that I could hook this application up to. Which one should I do? I don't know. You know, and the stakes are high, right? Like this one costs more, but maybe has higher deliverability. This one, you know, costs less, but is sort of clunky and the documentation's outdated, so it's kind of a pain. You're gonna end up spending more on engineering time to get that working. So my general posture is yeah, just make it work. Like give me something that works and then give me an exit hatch, and that's gets to the more micro implementation end of the spectrum. Is any of this making sense? Are you following me? Okay, all right. So okay, so on the other end of the spectrum, when you're implementing stuff, this reminds me of what was the ⁓ there was like a single page app framework back in the day that was very Rails-like. In fact, I think some of the Rails core members peeled off and created it. Ember? Do you remember Ember?
Sean Yeah, absolutely. I'm following you. ⁓ I didn't work with it much. I I remember yeah, I remember it being a a thing, yes.
Taylor Okay. It it was early in the hay it was like Backbone JS and Ember were going head to head back, you know, ten years ago.
Sean Yes, yeah. I was a backbone guy. Yep.
Taylor the reason I bring up Ember is because Ember had this ⁓ of ⁓ like this inheritance pattern where you could override things at a lower level to get ⁓ control. Meaning like, and I don't remember all the system architecture, but like let's say you had a web component ⁓ ⁓ it did a lot of like boilerplate scaffolding things for you and it felt magical. It would just work. But then if you needed to tweak one thing, you could just override that one file. I think Angular did kind of followed a similar pattern where you could just implement just the things you need, which you know, then scaffold up to the higher order stuff, you know, the the other system primitives that would come out. And it it was great if you understood it. I think it it it became extremely complicated learning curve because at least I remember trying to get my head around that concept, it being painful. But that too lends itself well to like Ruby. So Ruby is crazy. You can like in Ruby, you can basically recreate any file and inject any file. It has all these wild kind of meta programming faculties that are equal parts terrifying and equal parts powerful. I think most of the industry has gotten to the world where it's like, and maybe this is the Rails philosophy. It was like, yeah, you can do all that crazy stuff, or you could just do it the easy way because 90% percent of the use cases is going to be a brochure website or you know, you know, a very simple, you know, CRUD application or something like that
Sean I I love that concept. So it's you know, and in the in the age of AI and and with this product we're building, we're going to really lean on yes, we're gonna get these voice instructions or written instructions from the user, and then an agent's gonna respond to that. It sounds to me a lot like, okay. You're gonna need the primitive level stuff anyways. You know, it's not like, hey, here's a hero component. It's more like the hero component is a configuration of all of these smaller primitive parts. And if those things exist and are are coupled with really strong documentation that an agent can discover, then in theory, spinning up new components is fairly trivial.
Taylor Yeah. I would think so.
Sean Okay. Okay. ⁓ no, I mean, yeah, that was ⁓ maybe both. ⁓ I think I think what's what's interesting about this approach too is ⁓ that I'm I I I just I love the parallels to Rails because the y there's that path which is I can be up and running in five minutes. And then there's the other path which is I can do something.
Taylor Was that was that a question or a conclusion?
Sean crazy complex and bring all of my own tooling and I can still make use of this system if I am getting some benefit out of it.
Taylor Well, you know, we we make this joke all the time about every day there's like a new JavaScript framework, or at least there was a period of time, maybe five or seven years ago, where I think there was literally a website. It was like one of those single purpose like websites or or single function where you just hit it and it's like, was there a new JavaScript framework released today? And it was like, Yes, there was. and I think the frameworks actually were trying to s in a lot of ways, and and you know, the one that's risen to the top of the pack.
Sean Yeah. Yeah.
Taylor In twenty twenty six is obviously Nix.js. It has the biggest market share, I think. I think that's a accurate statement. You would know better than me. ⁓ I know there's others nipping at its heels or at least generating a lot of interest. But at the end of the day, the framework is the opinions. the reason we even have frameworks. What is a framework? A framework is simply just a a set of patterns that is built on top of very primitive technology. So, ⁓
Sean I think so, yeah. Right.
Taylor When Node.js came out, and I remember because we were a Rails shop and but we we had made some poor choices in the engineering of some software and it was causing a lot of churn and an excessive amount of expense. And Rails developers were becoming less and less accessible, ⁓ you know, from a recruitment standpoint. And then JavaScript developers were coming way, way, way more accessible because that was the hotness. Like I'm thinking w when we were working together, right? It was like twenty fifteen, twenty sixteen, twenty seventeen, twenty eighteen, like
Sean Mm.
Taylor that whole period of time, we were making this transition away from Rails because I couldn't find enough talent to to keep all the systems like up and running. Granted, I I I was probably doing a lot of things wrong too, for what it was worth at the time. Like in retrospect, I'm sure I was making poor choices in very many different capacity. But what I will say is that when NextJS came along, Or I'm sorry, when we were transitioning to that NPM ecosystem, it felt very disjointed. It felt very like I mean, it was cool 'cause you could like have twelve different libraries come together and you're stitching them all together. It it felt like the lamp stack days of my career, you know, when I could just like cobble together whatever I needed and it was the Wild West. That's the way it felt when NPM kind of like, you know, really blossomed in our industry. you know, Next.js comes along. And closes the gap on a lot of those weird dependency connections, right? It says, no, you you don't need to use any of these email providers. You should just use this one thing. You don't need a special ORM. You should just use Active Record. You don't need, you know, there were many different rungs on that ladder. And and it became kind of a breeze. It was awesome. Like I remember just doing a couple things, a couple lines of code, and it was like boom, you've got, you know, this unbelievable experience.
Sean Mm.
Taylor But it ca i I think it encouraged this sort of dependency on the corporation that was managing or maintaining that software or the dependencies that were inherent to that. I mean, I guess that's a fairly common I mean, when you think about open source software, you're you're bargaining that the value you get from implementing that dependency or that library or whatever is gonna pay off. You know, it'd be cheaper than rolling your own or coming up with a a more bespoke solution. Where do you land where do you land on that?
Sean Yes. Yes. In terms of ⁓ picking up open source software or what?
Taylor No, I guess just dependent. I tell you what, let's broaden the let's broaden the scope of this a little bit. ⁓ few months ago, Jensen Twang of NVIDIA stated something to the effect of the next great employee perk might be tokens. But what happens okay, so but what happens when the thing that makes you productive is also the thing that your employer, vendor, or platform is controlling? So
Sean Mm, okay.
Taylor In a way, so this is kind of crazy. So what he said, he basically this is a paraphrase, right? He said that ⁓ for a five hundred thousand dollar a year employee that they should be consuming up to two hundred and fifty thousand dollars a year in tokens. Obviously he he's expecting max productivity, right? And if tokens equate to productivity, then he is suggesting and and he goes so far as to say that like companies may start incentivizing employees with token budgets. And companies like Facebook or Meta or whatever, they have leaderboards in their office, I think, of who's using the most tokens amongst the engineering staff. I don't know if you guys have one of these at NLFI if you heard about any of this. But basically there's like they're gamifying the and encouraging the mass consumption of tokens, right? They are tying a productivity outcome potentially to compensation, meaning like I'm going to. compensate you with this bucket of tokens, you must then use that bucket of tokens to be productive at your job. Also, there's like competition going on amongst your peers. It almost harkens back to like ⁓ like the old company store analogy. Or the analogy is the company store that like, you know, I sold my soul to the company store. Meaning like back in the day, during the Industrial Revolution, A lot of companies, particularly here in Kentucky, like I live in Kentucky and Appalachia is is basically been decimated by this kind of stuff back in the day. But the idea would be that the company gives you a wage and then sells you ⁓ or basically will sell your wares. It controls your housing, your health care, you know, your recreation, your church often, you know, and so you kind of get wrapped up in this ⁓ this for profit model that is just keep dumping the profits back into the company, which is weird, right? So if NVIDIA is gonna start NVIDIA is gonna start compensating its employees with tokens that benefit it or its vendors to spend those tokens and then use that as the productivity measure for perks and hiring status and all that crap. It's almost like a race to the bottom.
Sean Uh-huh.
Taylor So let me a let me ask you this. With that premise, one, let me know if I can clarify anything, but two, would you say that this token budget as compensation represents well, does it represent compensation? Does it represent like infrastructure like a laptop, you know, is this just basically another company asset? Or is this like a means for the corporation to kind of surveil the way that you're, you know, operating on their behalf?
Sean Yeah, it's really bizarre because it's my personal view on it is that it's if it's a if it's a thing that you need to use for your job, then it's an asset. It's like a it's like a laptop or a piece of software or something like that. And I think it's I think it's really interesting because ⁓ I'm gonna I'm gonna answer this in a roundabout way just through my experience because we ⁓ what am I thinking of? Okay, so my brain's going in like three three different directions as it always does, as it always does. ⁓ all right, so I think there's there's kind of the like philosophical approach. I'll come back to that piece of it, but from a productivity standpoint.
Taylor I know. It's crazy.
Sean It's a tool y you need to use. And and yeah, there's the oversight piece, which I think is a little bit goofy because it's really easy to it's I should ⁓ maybe say that differently. It's really hard, I think, to know where that usage is going. And so if you give people a bunch of if you give people unlimited tokens, well, you can spin up a hundred different side projects, right? You can spin up your own factories and build your own stuff. And is anybody actually really gonna know about that today?
Taylor Well, I mean, no, if using the company store analogy, you're spending those tokens at the company store. You know, if the budgets are controlled by the employer, your productivity is attached to that, then and you're and it's a monitorable asset, right? Like I know how much you spent at the company store, you know.
Sean Yeah.
Taylor ⁓ so where where's that? I mean, what did Huang said something about like if I had a five hundred thousand dollar a year employee and he only spent five thousand dollars that year on tokens, I would lose it. I would just go completely bananas because that's completely unacceptable. I'm trying to maximize that person's productivity. So it's kind of weird. Like so the i i the company okay, let's see. Coal company script is what they call it. I never heard that term. Script represents a non-cash value. Like it's not money, but it is a form of compensation. Okay. So company script was often redeemable only through company controlled stores, right? Because it's not a dollar bill. It is credit. You know what I mean? It would come in the form of like payday advances and obviously like food and materials and clothing, that kind of stuff.
Sean Mm.
Taylor The workers given this non cash value that only helps them produce more profit for the company. And this is inside an ecosystem that's all priced and controlled by somebody else. Meaning I don't have any to your point, I can't just take those credits and go spend them somewhere else. I can only spend them on the company on the company's outcome. So positioning this as compensation is disingenuous at best. And at worst, it's horrifying.
Sean Right, because you're well, I it sounds like what they're doing is they're mixing compensation and productivity in a way that's like it can't really be untangled from one another, which I think you know there needs to be a corallel or a correlation in I s I just made up a word that was correlation and parallel, I think, between the two. but
Taylor Exactly.
Sean It's it's also it's why I'm thinking of like, okay, well there's like the the ethics or philosophy or whatever behind and psychology behind the compensation piece of it, but I also think the productivity side of it is totally missing the mark. I mean, I think that he's trying to get to a point which is yes, if you've only spent five thousand dollars and ⁓ on tokens and I'm paying you $500,000, then my $500,000 is not being well spent. And I think that's probably a fair point. But it's also like you're it's it's not measuring, it's not measuring impact, right? It's like it's it's like saying, here are two engineers, and one one of them opened up 600 pull requests last year, and one of them opened up five.
Taylor I agree.
Sean But if those five had more impact than the six hundred, then this other employee is far more valuable. And I don't I think it's it's probably oversimplification to be like there's a one to one correlation between token consumption and impact on on the outcomes the company wants to see.
Taylor Yeah. That's a great that's a great point. yeah, ⁓ but it does create this weird dependency, right? If you think about it, like the at the using the coal company analogy, the miners needed the store to live, right? AI workers, or or knowledge workers, we'll call them, need AI to be productive and stay competitive. Right? So it creates this commodity that is owned, controlled, and maintained and
Sean Mm-hmm.
Taylor The employer or platform gets to define what counts as productive use of that currency. ⁓ and the worker can't shop around. They're locked in.
Sean Mm, mm, mm.
Taylor Which i it's it's really ⁓ it's really so okay, so let me ask ⁓ understand because I I as a business owner, I have the same quandary, right? It's like if I could give this employee a tool that makes them infinitely more productive, I kinda have a responsibility to do that anyway, right? Like 'cause growth, you know, and and we need to continue to scale and make more and more, right? And and do more and more. So I I sometimes generally wonder, like or or I'm I'm conflicted, right? Because we're sitting in between these two totally logical. I mean, I can understand why Jensen is saying this and he's a smart dude. ⁓ but I can also see like the impact to humanity on the other side of the fence. So let me ask you this. Is this analogy morally overblown or is it useful precisely because it shows how infrastructure can become control? The the analogy of the coal company.
Sean I think the the thing that I'm the the piece of the analogy I'm not that's not grokking with me is the so if you the at the cool company you got those you get the scripts and then you can buy real things with them. Even though the prices are inflated, the system is corrupt, etc. With the tokens I think the part I'm not following is you're saying that they can then use those tokens on what I mean, they can't then go use that for food or whatever, right? Maybe.
Taylor well, yeah, maybe one d I mean that's an interesting thought. I actually when I was reading all this stuff, I was thinking about the correlation of Bitcoin and how Bitcoin has like represented like this digital currency and if tokens now become like this new digital currency is the way the world is moving. Yeah, that's interesting. Like, ⁓ what if we were buying and sharing, you know, lots like like what you know how gasoline is like commodity based pricing or what or it's based on, you know, all these different economic factors. So
Sean I see what you're saying. Yeah.
Taylor Or the rate of replacement, right? Like when you go up to the gas station, the gas that's in the tank has been there for a week, but the price changed yesterday and it went up a dollar. Like that doesn't make any sense. It's it's a physical good. You bought it for a price and now you're gonna sell it for a price, but we have a whole system that encourages this fluidity, you know, of that commodity's value. What if tokens become a commodity? Like in the same sense that you can buy and sell. It's like, you know.
Sean Yeah. Yeah. I see.
Taylor Anyway, that's but that's somewhat beside the point of the analogy around the coal thing. I I think credits at the general store, script or whatever you want to call it, it's like an IOU, right? I mean it's like it's only worth what the person accepting it is gonna take. So if if the company is willing to like it give you an advance on your payday with interest, and that's a piece of paper, that piece of paper is only valuable for that company that's you know that's holding it, right? And because they control your entire source of income, you have no leverage, they're gonna take it off the top. My concern with these tokens and the analogy therein is that what's to stop a corporation from using the tokens to box in your productivity and thus your competitive edge? We have become dependent upon this technology. I mean, we see this all the time. It's like ⁓ we get these awesome models. You build workflows around them and then they get end of life and now you're the price got jacked up. I mean this is a major problem is that ⁓ the economics of this is a major, major issue. And part of the reason that I think Jensen is saying this stuff is because we're moving from transactional chat interfaces is the predominant way we use these tools, to agentic workflows, which consume like a thousand X more tokens, like across the board.
Sean Yep. Yep. Mm.
Taylor And so Jensen's point is like, hey, dude, if you're not taking advantage of all the automation and all the opportunity and you're not doing half as much of your entire annual salary as you you know, in agenc outcomes and token consumption, then you're not working hard enough. Right? So he can throttle up and throttle down the expectations based on just how ⁓ I mean, why stop at two fifty? I mean the truth is managing agenc workflows are tough anyway. I mean I I y anybody who's actually done that, it is not a breeze. I you can Ralph Wigum loop your way to, you know, something, but if it's production, if it's a client, there's a lot of liability in doing that. And I don't think everybody's, you know, I don't think that's an appropriate way to conduct yourself. And I don't think that's what he's saying either. I think what he's saying is that, you know, we need to maximize productivity. I'm gonna incentivize the high performers to come to my company and I'm gonna trap ⁓ with this big bucket of tokens, which represents both their ingenuity, productivity, creative output. You know, it it's a it's it's a closed ecosystem, right? And so then I spend the tokens back and maybe they get a better deal, but it doesn't on the per token bulk rate. So profits go up and but I'm still you know, how do I go to another company if they can't like I cannot perform the same way at another company.
Sean Mm-hmm.
Taylor if I don't have the same infrastructure or tooling that was afforded to me at the last company, in a way, you get like locked into that ecosystem because you become dependent upon it. And that is exactly what happened in eastern Kentucky.
Sean Yeah. That's interesting because so now the the five hundred thousand dollar developer costs seven hundred and fifty thousand dollars to any other company, essentially.
Taylor Well, maybe. I mean, if you think like think about the weird, incestuous relationships of all these billionaires and like the few companies that or the the many companies they control by the few. Who knows? You know? ⁓ it turns out OpenAI strikes a deal with NVIDIA, so its engineering staff gets a lower rate on these tokens. I mean I I can st it not like that's a that seems like a benefit for the company. Profits go up. I just I would guess that the consequence to the end, ⁓ you know, the employee that's just trying to keep up, you know, and hit those productivity goals. I mean, 'cause at the end of the day, isn't your the amount of tokens you consume is also a measure of your efficiency of that ⁓ you ⁓ your efficient use of that technology, right?
Sean It is. So ⁓ so something I wanted to ask you on that subject is so you're you're a business owner and 'cause we're we're talking more about this. This I mean, not the like company store policy, but I I I see where you're going with it and I ⁓ you know and and
Taylor Sorry, to be clear, this is a a very cynical take and I reckon I'm just mostly exploring the concept. I don't I don't know that I don't think that Jensen has ill intentions, but I do think that the ⁓ the language he's using is a ideological hop away from a a lot of weird ethical and moral quandaries.
Sean Right, right, right. But then if okay from from the business owner slash CFO perspective, right, you so you ⁓ you own an agency, you have your you know, your fixed costs of how much your people cost, and you've accounted for that in in how you conduct your business and everything, and you're d for your people to do their job. ⁓ so that you can beat your competition or at least match your competition, they're now more expensive to you in particular, even though they would make even if they make the same exact amount of money. And you could say Yeah, everybody gets a hundred thousand dollars in tokens and I need you to use it to make the that money back for the company, but you've gotta that's a big risk for you because you don't you don't necessarily know if you're going to get that back. So where do you draw that line between like you mean you don't have unlimited funds that you can throw at your employees?
Taylor Right. I mean, nothing changes as far as the way you would you would budget for this. You basically give the team a budget and you tie ⁓ you tie those budgets to outcomes. Right. We don't think about like it is an interesting I mean, I'm sure for company like these companies like NVIDIA, I mean these massive fortune one hundred trillion dollar enterprises, like they can do whatever they want. I mean, money is not the limiting factor for them. I think for most businesses though AI is basically just like a software license. You know, it's like what do you need, you know, to be as productive as we expect? And in our world, we measure productivity and margin, you know? It's it's a common metric that represents your profit. ⁓ in which case, okay, I mean we all can kinda understand that. It's, you know, revenue minus ⁓ expenses equals, you know, what it's worth.
Sean Mm-hmm.
Taylor so I I think I don't think much changes, to be honest. I think that like you just continue to to meter the consumption of this as a traditional business expense, right? Not compensation. Then you can give the team what they need based on the requirements of the job. You measure them on how successful they were coming out of that job, and then it keeps everything nice and clean. Like there's no you know, compensation motive or or profit motive attached to that compensation.
Sean Yeah, and and we it at least for now you get to take advantage of the the agent providers subsidi subsidizing their their subscriptions anyways. For now, right. Yeah.
Taylor For now. Do ⁓ so I mean interesting, I mean costs are soaring. So in the last couple of weeks, ⁓ wasn't there there's a bunch of announcements where they're moving to like consumption based pricing. Is that right?
Sean Well, it's that the so there was all that hype around Fable, Fable Five, and
Taylor Mm-hmm. Ooh, I used it for the first time last night. It was it was good. It was real good. It was ex expensive.
Sean Yeah, yeah. And it's but they're getting this is what they were I I think we were going to talk about this and then it had gotten ripped off the shelf already. And ⁓ so so they're like they're dangling the carrot, so to speak. and th and you can use it for a couple of weeks and then it's going to be this that goes in this other category of
Taylor Yeah. Yeah.
Sean extra usage which is not API cost but it's not not API costs. It's really weird.
Taylor These are the weird can this is what I'm talking about, like the economics of what we're living through right now. It's like they give you something, they hook you on it, like yank it back. Right? What's to stop the business owner who is compensating you with two hundred fifty thousand dollars worth of tokens from saying, Nah, you I think you could do more. I think you should do more, right? Or they yank it back. They yank it back and all of a sudden You've lost your competitive edge. You are not keeping up with your peers. They have controlled the entire ecosystem, the task to the product productive output to the compensation. And I guess that's no different than measuring a team based on its outcome, right? If we're being honest. ⁓ but it does feel weird and kind of gross that to talk about it in terms of compensation, I guess. Anyway. Here's what I wrote here's what I wrote down ⁓ to kind of conclude this.
Sean Yeah, yeah, agree.
Taylor section, but ⁓ you know, he may be right. Right? ⁓ the best workers of the next decade are gonna be amplified by AI agent. But the question is not whether or not workers should have access to the tokens, because we know that they should. The question is who controls the meter, who sets the prices, who defines productivity, and whether workers are being empowered with tools or quietly pulled into a new company store. All right. Okay. All right. So I kinda did not
Sean Yeah. Yeah, that's a that's a good point. ⁓ good closing thought.
Taylor strictly adhere to the agenda. So we're a little bit over on time on that last section. So a little bit less time on this this last one. But ⁓ we had talked about maybe it would be fun, Sean, for us to kind of go on a a little bit of a deep dive on or or at least a overview on tools or individual experiments that you're working on or something, ⁓ maybe a new product that you came across or a new service you're excited about or something like that. So ⁓ With that, ⁓ what w you got anything to share in that regard? Anything you've been working on or or ⁓ have sparked your curiosity as of late?
Sean Yes, my ⁓ my latest in Should I Just Build That is ⁓ Okay, I've been I've been I've been hammering on on both Claude and Codex over the last week or so with it's just like lots of different things. And I won't go into all of the background on it because I'm I'm moving a lot of things around. It's as messy as my brain is. But suffice it to say, I was in a position a few days ago, maybe about a week ago, where I was like, I have 10 active well, t I have 10 open sessions with either Claude or Codex right now. And I I think there's there's like a I'll have one window that's got a bunch of chats going and then another window that's more like this is the the product work that's happening right now or the research or something like that. And so they're not all necessarily active, but I try to get them organized in a way where I can see what needs my attention at any given time. And there's a whole conversation we could have about context switching because it's an exhausting way to work. But I was like, okay, look, here's the thing. I want this view and I want to know when these things need me. but sometimes I need to I need to open a yet another tab because I need to just run a terminal command directly. And then sometimes I do actually want to look at the code for one reason or another. ⁓ usually it's more like the agent wrote a markdown file and it wants input on the file. And like, so we still have too many surfaces, too many notifications. It's really confusing. And ⁓ so the yeah, my latest thing is all right, does is there a tool that exists that can package all of that up for me? And the closest thing I came to was a tool called conductor, which is conductor.build. I've got a coworker who's using it and loves it, ⁓ but he's using it only for personal things because. It's you know, just kinda like ⁓ is this it it hasn't passed security.
Taylor They don't offer it they don't offer it at the company store, do they?
Sean They don't offer it at the company store. And we have to be very mindful of for especially the folks of us who have access. You know, we we are potential targets. so we
Taylor Well, well, well.
Sean I I did so I did a full security review of this thing and it came back and was like, it's not just sitting on the desktop, it's actually sending information to their servers. It doesn't need to be, it's kind of silly. ⁓ but I didn't I also didn't know a ton about it. And so I so that prompted the question, how hard would it be to build this thing? And over the last two days or so I've been working through it. And I have to tell you, I I another conversation I want to have with you at some point is I think there's a huge conflict to All of these layers getting collapsed by by companies like Anthropic, where it's not wrong, it's just confusing. There's the LLM, there's the agent, and then there's the agent harness. And all of those things people think are just the same thing. And they're not. And so you don't necessarily need to use, you know, Claude Code CLI to use Claude Code. You don't need to use Claude Desktop to use Claude. And so that's what I'm playing around with is like, well, how hard is it? How much have they locked down through the Claude Code CLI? And my my first forty eight hours lead to me to think, hey, it's it's not that hard to tap into it actually.
Taylor so the Claude Code harness got leaked ⁓ I don't know, a few months ago. Do you remember that? The source code, it's ⁓ I listened to somebody who walked through it all and they were like, it's brilliant. I mean, and it's very simple. It's all in plain language. You know, you can go read how it constructs my understanding is you can actually see how it does most of the stuff and and you're right, the harness is the value. It's nothing ⁓ all this other stuff is like
Sean ⁓ yeah, yeah. Yep.
Taylor Consider these the primitive technologies. The harness is the framework. The harness is the thing that sits on top, bringing it full circle to our conversation from earlier. Yeah, and it's all about the it's all about it's all about that ⁓ nuance and sophistication and the effort put there. How hard would it be to build that? I imagine they they put a lot of a whole lot of effort into that thing. so I don't know. It seems like an ambitious side project, but
Sean Yeah, yeah. So I'm not trying to I'm not trying to rebuild Claude Code CLI. I'm trying to bring the ability to manage multiple claud code CLIs into an interface that's easier to manage and ⁓ gets coupled with a a built-in terminal and code editor. So I think it's a lot of like pulling things that exist, rearranging them and orchestrating them more than anything else.
Taylor Yeah. do you use the Claude desktop app?
Sean I am moving away from it. I had used it pretty exclusively for a few months and and this is yeah, this has been my reaction lately. It's been like, yeah, it's good, it's good. I think I can I can sell I can increase my productivity by using those primitives in a way that ⁓ serves my workflows more effectively.
Taylor Yeah. ⁓ it's Claude feels so I don't know. Like there I think I think we had this conversation before. Like they're moving too fast and so they haven't I don't know, it feels like it's real squishy. Like some days it works great, other days like it's particularly that integrated environment. Like they'll have like a ⁓ what is it? Claude code on the desktop app will have like an integrated browser. Brilliant, it's great idea, right? I saw they recently released like a a terminal that you could just pop that open to if you're using work trees or whatever, you could like hop right to the work tree in a in a integrated terminal, which is awesome. But I can never get the it always like tries to start a bunch of processes and then it doesn't work and then try to hit it in the external browser and that doesn't work. It's just a big pain in the butt. Like so yeah. I figured they'll they'll sort that out soon. Okay, we're getting close on time. I'll tell you what I've been working on.
Sean Yes, yes, For sure. Yeah, what have you been working on?
Taylor Yeah. so comfy UI. Have you ever have you ever heard of Comfy UI?
Sean Mm I don't think so. I don't think so.
Taylor Okay. So Comfy UI is basically like a it's like a open source interf web based interface for working with stable diffusion models to create automated workflows around images or image or video generation, that kind of stuff. it's cool. It's like it's like one of those infinite canvas type user interfaces where you connect different nodes that do different things.
Sean Okay.
Taylor ⁓ and you can tweak ⁓ and play with ⁓ and of course it works really closely with Hugging Face. Like you can download if you're not familiar with Hugging Face, Hugging Face is basically GitHub for open source models. UI is pretty cool. ⁓ It I it ⁓ I I been surprised at deep ⁓ the different nuances of how you the models and configurate ⁓ you know, configure the different sampler directives and things like that. And they have really dramatic impacts on the quality output or lack of quality or whatever. It's it's ki it's pretty fascinating actually. So That's what I've been working on lately. suggest you install it and play around with it. It's really fun if you haven't. I'm real interested in like building kind of deterministic workflows or at least stylistic stylistically deter stylistically deterministic workflows for Ample, right? So if I do like a headshot of somebody, I want to be able to, you know, make that exact same person look different or in a different pose or whatever. And so the idea is you create
Sean Mm-hmm.
Taylor construct these workflows and then you can generate, you know, different nuanced versions of that imagery around, you know, the same individual each time, which I think is is really cool.
Sean Okay, okay. So it it's it's it's bas it's all about visual workflows.
Taylor Yes, correct. So well no, I mean there's a command line component to Comfy UI too. So you the workflows are just JSON files and so you can author them, you know, without ever even using the web interface. But I actually kinda like the web interface 'cause ⁓ you know, I can kinda see the progress meter as it's running, I can visually understand how the nodes connect to each other. ⁓ and just like anything, you know, it's highly nuanced and so I always enjoy kind of peeling the onion and learning.
Sean Mm-hmm.
Taylor Why does this do this? And and when I tweak this just one little bit, like what happens? You know? So it's it's I've been having a lot of fun tinkering.
Sean Yeah. That's really cool. I I've I've done research on running some of those models on the and it seems on on the Mac mini and it seems like it's just not powerful enough for it.
Taylor Yeah. ⁓ Sean, thank you so much. This is a a great conversation. I think we're we're right about at time, so it's probably probably time to wrap it up, get started on the day. ⁓ any parting thoughts or or anything that you would like to focus on next time? I wrote down a couple of things. You we talked about agent harnesses and we talked about context switching, both of which are pretty interesting topics to unpack. ⁓ what about you?
Sean Yeah, I think I I that's that's really where my brain was going as well. And ⁓ and yeah, I'm I'm excited to to to also dig into that ⁓ that Rails approach and see where where that can take us with the product.
Taylor Yeah, that's a great idea. I can't wait to see what you come up with. All right. Well, thanks so much, man. We will catch you next week. ⁓ peace.
Sean All right, see ya.