Latest / Gross To Net / Ep. 23 - Prove Success with Jack Atlasov | Gross To Net
Transcript
- 0:21SPEAKER_02: Hey everybody, welcome to Gross to Nut, the podcast that asks what you're really optimizing for and what it costs to get there.
- 0:28SPEAKER_02: At 15, today's guest crossed the country alone on a Greyhound bus.
- 0:32SPEAKER_02: At 23, he paid off his debt, stayed for two years, and spent five months driving around America on$40 a day.
- 0:39SPEAKER_02: Because why would you spend your best parts of your life slaving away?
- 0:42SPEAKER_02: Then he spent 15 years inside the machinery of American e-commerce, scaling more than 20 brands and driving nine figures of revenue for other people's companies.
- 0:51SPEAKER_02: Now he writes job descriptions for employees who aren't humans at all.
- 0:56SPEAKER_02: And he builds software whose entire purpose is to refuse to believe what the AI just told him.
- 1:01SPEAKER_02: Everyone in this space is selling certainty.
- 1:04SPEAKER_02: This guy just published a note on his own project that says the lift is unmeasured.
- 1:09SPEAKER_02: Please welcome a guy that I really like.
- 1:12SPEAKER_02: We're going to talk about AI.
- 1:14SPEAKER_02: We're going to talk about business.
- 1:15SPEAKER_02: It's Jack Atlasov.
- 1:18SPEAKER_00: Sub Jack.
- 1:19SPEAKER_00: Hey, everybody.
- 1:20SPEAKER_00: Hey, what's up, Joe?
- 1:21SPEAKER_02: Did I say your last name right?
- 1:23SPEAKER_00: You did.
- 1:24SPEAKER_00: Yeah, you did.
- 1:25SPEAKER_00: You're one of the like four people in this world who can pronounce it correctly.
- 1:29SPEAKER_02: Man, you know what?
- 1:29SPEAKER_02: That just brings me joy to hear.
- 1:32SPEAKER_02: It's the only thing.
- 1:33SPEAKER_02: I want that to be on my on my tombstone that I pronounce several people's names correctly.
- 1:45SPEAKER_00: Uh, and I guess has some kind of lineage to Atlas, the you know, map uh etymology there.
- 1:53SPEAKER_00: So well, that's where I come from.
- 1:55SPEAKER_02: I think it's a cool last name.
- 1:57SPEAKER_02: Um, I'll I'll say that like uh you're Austin based as well.
- 2:02SPEAKER_02: And we kind of uh we kind of met and got to know each other via some uh like a an AI e-commerce group.
- 2:12SPEAKER_02: Um and then I started stealing all your uh GitHub repositories.
- 2:17SPEAKER_02: And I went to one of your workshops on AI.
- 2:19SPEAKER_02: I think that like I th I think generally it's a it's a travesty to me that you haven't published more stuff about um about how to like how to not just how to do this stuff, but how to think about it.
- 2:33SPEAKER_02: So like I thought it'd be great to have you on and just riff on AI a little bit.
- 2:40SPEAKER_02: Uh we we taught we actually talked a lot before we started recording today.
- 2:45SPEAKER_02: So we won't try to repeat that, but um, I do think that like you're so you're you're you're doing uh a lot of uh Amazon commerce, right?
- 2:56SPEAKER_00: Yeah, primarily mostly Amazon right now.
- 3:00SPEAKER_02: And I know that a lot of your kind of skills have to do when we started talking about doing this podcast, this uh this recording, I was talking about company in a box stuff.
- 3:10SPEAKER_02: I think that's a it's it's a really kind of like prevalent theme that keeps coming up, right?
- 3:17SPEAKER_02: And uh there's a lot of new releases, right?
- 3:19SPEAKER_02: Like Claude and Chat GPT focusing on enterprise, uh Gemini Enterprise kind of like recently being like like Gemini Enterprise, like a month ago, basically handed us like, hey, here's a company in a box.
- 3:34SPEAKER_02: Um so I think it's it's coming in a big way.
- 3:37SPEAKER_02: I don't know if people know how to think about this yet.
- 3:39SPEAKER_02: So let's let's go back to the beginning.
- 3:41SPEAKER_02: Like, when did you when did you kind of start getting getting interested in building workflows with AI, with you know, thinking about like how can how can this really be useful beyond just like writing my LinkedIn posts for me?
- 3:56SPEAKER_00: Yeah, totally.
- 3:57SPEAKER_00: Well, um, broadly speaking, I've been into AI for you know over a decade.
- 4:02SPEAKER_00: Like some of my first uh Twitter ex posts, if you go on my feed, you can see like the Wait but Why article that um that uh Tim Urban wrote a while ago, I think in 2015, on kind of the exponential takeoff of intelligence due to AI.
- 4:18SPEAKER_00: Um I studied biotech in undergrad, so I was already kind of thinking about the intersection of biology and technology.
- 4:25SPEAKER_00: And then I got really into cognitive science at a certain point.
- 4:28SPEAKER_00: So I was kind of primed to be paying attention to the field.
- 4:31SPEAKER_00: And I noticed the Transformer paper when it came out, but it was very much from like a scientific lens and like from a research lens.
- 4:38SPEAKER_00: And then this company, a couple of years ago, when I was heading growth and e-commerce at Bird Rock Brands, which is uh a company that was based out of B Cave, and we we had an exit uh this year, earlier this year, um, we were running super lean at that company.
- 4:54SPEAKER_00: Our like um OpEx was like 3%, right?
- 4:58SPEAKER_00: Under 3%.
- 4:58SPEAKER_00: And we were a you know a sizable business, right?
- 5:02SPEAKER_00: Um, many brands, many marketplaces, lots of stuff going on.
- 5:05SPEAKER_00: And our whole e-commerce team was essentially three people.
- 5:09SPEAKER_00: So that kind of leanness um and my personal kind of uh enjoyment of that leanness led me to starting to look at some of these AI tools, some of these um applications of the models within harnesses like Codex and Claude Code.
- 5:28SPEAKER_00: And now you're seeing, you know, OpenClaw and Hermes Agent and like some of these other open source harnesses.
- 5:34SPEAKER_00: So I was looking at a lot of that stuff like when Claude Code was first coming out early last year, because I was basically just trying to see like how can I staff up my team without necessarily bringing on another agency or like seven different freelancers, or you know, how can we just be more effective and do more?
- 5:55SPEAKER_00: So that's what that's what kind of initially led me to leaning into all of this stuff pretty early was just real-world business cases that I was trying to figure out as we were, you know, scaling this company.
- 6:09SPEAKER_02: Yeah.
- 6:09SPEAKER_02: That's awesome.
- 6:10SPEAKER_02: I I think that there is um there's a lot of that like necessity is the mother invention stuff.
- 6:16SPEAKER_02: Um that's really interesting to me too.
- 6:18SPEAKER_02: I I and I think probably a lot of uh early bootstrap, mid-stage entrepreneurs, people who are like cause because I know that I I went through, yeah, you know, I mean, my background is is pretty different than yours, but like at Yellowbird, I went through that thing where I was like, I've got a stack of like you write out the job descriptions of like, hey, here's here's all the stuff we're doing, and I've got this stack of job descriptions that I want to be able to hire for, and you look at the budget and you're like, okay, well, I can hire for two of these, and then the rest of the stack I have to keep doing.
- 6:57SPEAKER_02: And so there is, yeah, there is a real like, you know, that's a that's a real thing if you're scaling, especially if you haven't, you know, if you don't raise money like out the gate, which yeah, I mean I think it's I think in tech it got pretty common.
- 7:10SPEAKER_02: Even in CPG now, I think there are a lot of folks who were like have a pre-launch product who are raising a million dollars, and it's like, okay, well, if you have a million dollars, but then you don't have to think too hard about it.
- 7:23SPEAKER_02: You can just hire some folks, but like, what if you don't have a million dollars?
- 7:27SPEAKER_02: Like, not like not everybody is gonna have like a really obvious product market fit.
- 7:33SPEAKER_02: Maybe people just don't want to give up, you know, you have to give up 50% of your company for a million dollars.
- 7:38SPEAKER_02: Like it's it's gonna be significant for a physical product or a food product or something like that.
- 7:44SPEAKER_02: So why not just, you know, a lot of the work is gonna be on the computer anyway.
- 7:50SPEAKER_02: So why not just hire um a bot or an agent or team of agents to do it?
- 7:55SPEAKER_00: Yeah.
- 7:56SPEAKER_00: Yeah.
- 7:56SPEAKER_00: And and like for so many people, I think that approach that next stage of working with AI after just chatting with it, for me, the entry point into these uh coding tools and a little bit into the agent side was really through a very common use case, which is just data analysis.
- 8:17SPEAKER_00: You know, I'd hired a data analyst, but there was just so much to do that instead of you know hiring like six data analysts, I just decided to um use the tools, use Claud Code, and start to use Codex to start doing data analysis.
- 8:34SPEAKER_00: And I taught myself how to use Jupyter notebooks and Python and like what data, uh what Python libraries were for like different data analysis uh tasks and things like that.
- 8:45SPEAKER_00: So that was like a very kind of natural entry point into this whole world.
- 8:50SPEAKER_00: I wasn't like um right from the beginning, you know, trying to recreate like a company in a box.
- 8:55SPEAKER_00: I was just trying to figure out where I should spend the next dollar of advertising spend across all my different marketing channels.
- 9:02SPEAKER_00: So it really started like, you know, it feels like a long time ago, but really it was just like two years ago, or um a year and eight months ago, basically, where I was just doing, you know, uh LLM assisted data analysis, basically.
- 9:17SPEAKER_00: And then from there, it was just this very natural progression that had these punctuated moments of me really leaning in and going, holy shit, I can I can do this.
- 9:30SPEAKER_00: Like, oh, what else can I do?
- 9:31SPEAKER_00: And then just kind of evolving into building actual software, building workflows, creating agents, and then eventually, you know, setting up entire systems.
- 9:42SPEAKER_00: Um, but yeah, it had very like simple beginnings.
- 9:46SPEAKER_02: Yeah, I so I've had a similar thing.
- 9:50SPEAKER_02: I think that you're um I'll say that I say I'll say that you're you know ahead of like where I am, I think, as far as like implementing it, it sounds like you've gotten really, really deep on a lot of these things that probably a lot of people are casually casually looking at, right?
- 10:08SPEAKER_02: The the one that I had that I told you um before we started recording that I hear the most often is people being like, oh, it can summarize my emails, which is like now built into like I I guess Gmail's got like a new kind of launch.
- 10:23SPEAKER_02: It just has like a little I don't see it in all my Gmail inboxes, but I've got I've got one Gmail inbox that's like here's a beta feature, here's your AI inbox, and it just summarizes the stuff and summarizes your calendar and all the sort of stuff.
- 10:38SPEAKER_02: So so like all the little workflows that people were coming up with, like they're building into these enterprise grade, you know, yeah harnesses.
- 10:46SPEAKER_02: Um but that that one of like, hey, we've got a lot of business people, like business business has adopted AI, like it's no longer, I think there are security issues still that we could probably still talk about um on here, but by and large, it's pretty accepted to be using an AI workflow.
- 11:07SPEAKER_02: It's like weird in mid-2026, late 2026, it's weird to not be using AI at your business.
- 11:15SPEAKER_02: Like even even the older people, even people in their you know, 60s who are still doing who are still involved in business, who who are saying things like, let's use AI for that.
- 11:25SPEAKER_02: Well, can't why are why does it take us 20 days to close the books?
- 11:28SPEAKER_02: Don't we have don't we have an AI workflow or isn't there and the answer is yeah, the answer is kind of like, yeah, it should be able to help you, you know, if you've got a if your accounting team is two people or one person, and they used to be constrained by you know the time that it takes a human to gather all the data, a time it takes a human to like pro do all the spreadsheets, etc.
- 11:52SPEAKER_02: They used to be gated by that.
- 11:54SPEAKER_02: And I think now there's we're at this kind of like inflection point where business at large, not just like not just like quirky guys like you and me who were like, what's the next big thing?
- 12:06SPEAKER_02: Right, because there's always that wave of people who were like, like I'll say like I was trying to figure out when they first like in 2022, like late 22, they launched Chat GPT, and I was like, How can I build agents with this?
- 12:22SPEAKER_02: I I don't think I knew quite knew the term agents yet, but there was like baby AGI, and there were these uh there were these little like uh GitHub repositories that were like you ran it in terminal and it just was these recursive loops that it was just it was just these loops that would run, and so you'd run it on a task and give it kind of like a kill switch, and it was they were terrible, man.
- 12:45SPEAKER_02: I don't know if you messed around with the early agentic stuff.
- 12:48SPEAKER_02: It was absolutely terrible, but it was like it was like being there for it was interesting because now that the stuff is like good and it can do you know, build its own agent loop and go do recursive work, that it's it's helpful that you know I wasted all those weekends in 2022, you know, weekends I'll never get back.
- 13:10SPEAKER_00: But yeah.
- 13:12SPEAKER_00: Hey, if we're if you're having fun and you're learning and you're enjoying the time, then it I don't think of it as wasted.
- 13:18SPEAKER_00: But um but yeah, I think fundamentally you know, there's probably more buckets than this, but just like broadly speaking, I feel like many people that I meet will kind of like self-segment into two buckets, right?
- 13:33SPEAKER_00: There are, I call these uh buckets the snapshoters and the trend maxers, right?
- 13:41SPEAKER_00: And so many people, I'm not saying it's you know better or worse, but many people will look at something that's a process, well, at least that I believe is a process, like the improvement of some technology over time, right?
- 13:55SPEAKER_00: They'll look at it at a moment in time, they'll test it, and they'll go, they'll take a snapshot, right?
- 14:00SPEAKER_00: And they'll go, okay, this thing is like this, right?
- 14:03SPEAKER_00: And then they come up, they they they take away a belief about that thing or some kind of like assumptions, and then they move on with their life, you know, and they kind of like um maybe revisit it, maybe they don't, but that's their reality as it relates to that thing that they interacted with, right?
- 14:21SPEAKER_00: Like if it's AI, you know, and they look at it in 2022 and they kind of take that, you know, mental snapshot of like, okay, these LLMs hallucinate all the time, you know, they don't do arithmetic very well.
- 14:33SPEAKER_00: There are these weird little GitHub projects, and like, okay, that's the state of it, and like moving on, right?
- 14:39SPEAKER_00: And then there are people like me who really are just like interested in the trend.
- 14:44SPEAKER_00: Like, I just look at the graphs and the curves and the progress.
- 14:50SPEAKER_00: And that's from the reason I'm like that is because of my science background, right?
- 14:55SPEAKER_00: Because like all through school and through my like 20s, I was just consuming so much science, and there was so much exponential um like proof that was presented in front of us in in biotech and in, you know, a lot of these sciences where there's been a lot of progress over the last 50 years.
- 15:15SPEAKER_00: And the main kind of like thing that gets hammered into you when you're learning about this stuff is that your intuition is not a great um guide to like how these things evolve, right?
- 15:28SPEAKER_00: So I've just been trained, I guess, to like not be too judgmental about any one particular moment in time with technology and instead to ask myself like what is the trend?
- 15:42SPEAKER_00: What is the curve?
- 15:43SPEAKER_00: And then it makes it a little bit easier for me to like justify investing time when things look kind of not fully baked, right?
- 15:52SPEAKER_00: Because I know that, you know, maybe not everything's gonna end up improving on an exponential.
- 15:57SPEAKER_00: But if I try five different things and, you know, um invest some time in five different things, like one or two of them will, and then it'll be really interesting in six months or 12 months.
- 16:07SPEAKER_00: So that's kind of how I think about it.
- 16:09SPEAKER_02: I do, I do agree.
- 16:10SPEAKER_02: I think that there needs to be like we just need more scientific method.
- 16:15SPEAKER_02: I know not everybody is gonna go get a degree in a science field.
- 16:21SPEAKER_02: Um I, you know, my degree is not in a science field, but I've always been super interested in it.
- 16:28SPEAKER_02: So I've always been really interested in you know, in physics and whatever, like computer sciences and things like that.
- 16:36SPEAKER_02: And I and so like I I agree, I think that there's this thing of like people snapshotters is a good word.
- 16:42SPEAKER_02: I'm gonna steal it.
- 16:43SPEAKER_02: Um it's mine, it's mine now.
- 16:45SPEAKER_02: I've adopted it now.
- 16:46SPEAKER_02: Yeah, but I I I do think that there are people who are like taking a snapshot of AI.
- 16:51SPEAKER_02: Like I've had that thing where I like have a conversation with somebody who their snapshot of AI is like 2022, where it's like, oh, it's gonna give you wrong information.
- 17:01SPEAKER_02: It's like, yeah, it'll still give you wrong information, but it's like take it's like if I if it's like if you didn't update your idea of the internet since 1996, and you're like, oh well, it's a it's a noisy dial-in modem, and you gotta use prodigy, and you know, AOL still sends you a little C D, whatever.
- 17:20SPEAKER_02: It's like it's like if you haven't updated your model, then it's kind of you're that's there's nothing useful you can do with like a 1996 mental model of the internet.
- 17:29SPEAKER_02: So uh I appreciate that that you know, keeping like because the scientific method will say, like, hey, I'm always probably wrong about part of it.
- 17:38SPEAKER_02: Like, that's the reason we call these things theories, because we're like, it's not we're gonna learn something about it, we're gonna learn something about this, and we realize that like our current view is wrong.
- 17:47SPEAKER_02: So there is a scientific view of like we're better than we once were.
- 17:53SPEAKER_02: Our understanding, our application is better than it used to be, but it's not as good as it's gonna be 10 years from now or 100 years from now.
- 18:01SPEAKER_00: Exactly.
- 18:02SPEAKER_02: Um, and I and I agree, I see a lot of people, and I think that I do think too in business, there's this weird thing in business too, where people get rewarded within their careers for being really confident, for snapshot shotting and saying, like, I I mean, I've read way too many business books than I even want to admit, like just being curious about business and like, hey, what is it that I'm yeah that I'm missing here as far as like and and most of the people who are like in who are making bold statements about business, they're like, this is my framework, and it's like this is the framework for running a business, this is the framework for marketing, and it's like it's a framework, right?
- 18:42SPEAKER_02: This is a framework for running a business, but like but like that there are a lot of people who get rewarded for that level of like snapshot confidence.
- 18:51SPEAKER_02: Like, this is how we ran my business successfully in 1970, and that's the way to do it.
- 18:58SPEAKER_02: And you're like, okay, well, that's a way to do it for sure.
- 19:02SPEAKER_00: Yeah.
- 19:03SPEAKER_00: It's great to have strong opinions as long as they're loosely held.
- 19:08SPEAKER_00: You know?
- 19:09SPEAKER_00: That sounds familiar.
- 19:10SPEAKER_02: That sounds familiar.
- 19:10SPEAKER_00: I'm gonna attribute that to you, but that yeah.
- 19:13SPEAKER_00: No, it's not mine.
- 19:14SPEAKER_00: It's not mine.
- 19:15SPEAKER_00: It's um, I want to say it's one of the uh uh founders of Intel.
- 19:20SPEAKER_02: Yeah, strong opinion is loosely held.
- 19:21SPEAKER_02: It's uh I know that I've read that in a business book too.
- 19:25SPEAKER_00: Yeah, yeah, yeah.
- 19:26SPEAKER_00: Yeah, I mean, like I I appreciate whenever I interact with people that have like a strong viewpoint on something, right?
- 19:33SPEAKER_00: It's it's definitely more interesting to have conversations and debates and and like explore topics when somebody is like, hey, you know, this is how I see the world and this is what I think and this is why I think this way.
- 19:45SPEAKER_00: But I also am very sensitive to like, okay, but is this a dogmatic belief, right?
- 19:52SPEAKER_00: Like, are you anchoring your personal identity to this belief?
- 19:55SPEAKER_00: Are you, you know, um still looking at the world?
- 19:59SPEAKER_00: Like, are you paying attention, right?
- 20:02SPEAKER_00: And like, you know, we're starting to talk about some of this esoteric stuff, maybe, but this honestly is like why I love AI is because it facilitates and brings up and kind of encourages these types of conversations, both for me internally with myself and with people like you.
- 20:19SPEAKER_00: Like the best, some of the best parts of working with AI for me have nothing to do with like the things that I build and have everything to do with the kinds of thought processes and the kinds of reflection and the way of like showing up in the world that I get to um you know embody and inhabit more when I use the technology.
- 20:42SPEAKER_00: So, like, for example, right, um building with AI is a different experience than consuming products and consuming things, you know, in the world, right?
- 20:54SPEAKER_00: And um most of us in most areas of our lives are consumers.
- 20:59SPEAKER_00: That's just the nature of life, I think.
- 21:01SPEAKER_00: Like, we're not creating all the time across everything.
- 21:05SPEAKER_00: But with AI, I am able to do some amount of creation that I wasn't able to do before.
- 21:11SPEAKER_00: So on balance, right?
- 21:13SPEAKER_00: Like overall, my life has shifted more into more creation than it was before.
- 21:20SPEAKER_00: And what does that do for me?
- 21:21SPEAKER_00: Well, it creates more opportunity for me to pay attention.
- 21:25SPEAKER_00: And like my whole life, like my entire experience of being alive, can essentially be summarized as how much attention am I paying to my life, to the world, right?
- 21:37SPEAKER_00: Like we've all experienced, I think, this feeling of like driving home some from somewhere, and we like don't even know how we got home, right?
- 21:45SPEAKER_00: Like we just kind of are somewhere else, right?
- 21:48SPEAKER_00: Did we really even live those 30 minutes or hours?
- 21:50SPEAKER_00: Like, where were we?
- 21:51SPEAKER_00: Right.
- 21:52SPEAKER_00: Um, I just have this fundamental belief that like attention is the most precious thing in the universe that we as conscious creatures get to.
- 22:00SPEAKER_00: Experience.
- 22:00SPEAKER_00: So now we're really philosophical, but bear with me for like 30 seconds and I'll finish this thought.
- 22:05SPEAKER_00: Like creation, the act of creating creates more opportunity for attention.
- 22:10SPEAKER_00: And attention is the thing that makes me feel alive.
- 22:14SPEAKER_00: You know, different kinds of attention too.
- 22:17SPEAKER_00: Focused attention, diffuse attention, imaginative attention, you know, exploratory attention, like all the different flavors of attention.
- 22:25SPEAKER_00: But as a tapestry, that's like some good there's some good stuff there, right?
- 22:30SPEAKER_00: And if I can use AI to kind of support that within my work life, within my personal life, then and that's pretty sweet.
- 22:38SPEAKER_00: That's pretty sweet.
- 22:39SPEAKER_02: That was 35 seconds, by the way, but I'll I'll allow it.
- 22:43SPEAKER_02: I'm just kidding.
- 22:43SPEAKER_02: I wasn't, I wasn't counting.
- 22:44SPEAKER_02: I no, I love this, man.
- 22:46SPEAKER_02: I love this.
- 22:46SPEAKER_02: I think I do want to get to some uh practical stuff that people uh that people can do.
- 22:54SPEAKER_02: And I think that this is a good like talking about attention, I I think one of the ways that I think about that, because I I also feel this like it uh it allows you to create maybe where you didn't have the where like you you're you're creating versus shopping.
- 23:12SPEAKER_02: And my example is like is like you know, one of my one of my first kind of software builds that I did using AI was to build like a CMMS system, which is like, you know, we we run our our own manufacturing and like CMMS system is like a you know computerized maintenance management system.
- 23:33SPEAKER_02: And so it's like if you have a bunch of machinery, heavy machinery, right?
- 23:38SPEAKER_02: Like it's like you can have this giant, you know, automated bottle filler, which we do, and it's got all of these parts and pieces, and it's like you don't just turn it on every day, it's like there are microscopies, there are these tiny little like ball bearings, you know, and there's and there's 18 of them per piston, and they each have a number on them, and they have to be like managed.
- 24:02SPEAKER_02: And so it's like these compute these CMMS systems are basically just like um it's like I have to keep track of how many extra ball bearings I have, how many extra like gaskets do I have that fit on piston one?
- 24:16SPEAKER_02: Because guess what?
- 24:17SPEAKER_02: Piston one has five different sizes of gaskets, and I have to have you know three backups of each one of them, and I've got to know my vendor and my cost and all that sort of stuff.
- 24:26SPEAKER_02: So it's just tracking all of these things across the you know, across your whole system.
- 24:32SPEAKER_02: And these CMMS, you know, systems are you know$30,000,$50,000 a year, right?
- 24:38SPEAKER_02: You can get the one that Toyota uses that's you know like hundreds of thousands of dollars a year.
- 24:44SPEAKER_02: Um, but then like at the at the lower end, it was like you you really don't need a ton, right?
- 24:51SPEAKER_02: You're not managing you're not managing like a global network of manufacturing locations.
- 24:58SPEAKER_02: And so like my first like software build was like, let me think really deeply.
- 25:02SPEAKER_02: And that was the that was the thing that was interesting is that like if you're shopping for a CMMS or uh or a CRM or you're shopping for some system, you're just kind of shopping and you're relying on marketing materials or a salesperson to tell you how this is gonna change your business.
- 25:19SPEAKER_02: But like if you go to build something, if you say, hey, we're a small business, I think I could vibe code this or I could work with an engineer.
- 25:28SPEAKER_02: Like, because it even before this, I had worked with like engineers on smaller projects on like, you know, hey, it's 1500 bucks to like customize this thing within your you know, Google workspace.
- 25:41SPEAKER_02: You know, here's 1500 bucks, 1200 bucks to like build the skill that just repeats itself.
- 25:47SPEAKER_02: And when you do that, you have to think really deeply about what the skill is.
- 25:51SPEAKER_02: Like you have to think really deeply about what it is that software is for you because it's like if you're using the Toyota CMMS system, that's like a small company like me using Salesforce.
- 26:04SPEAKER_02: I'm gonna use 3% of Salesforce or 2% of Salesforce.
- 26:09SPEAKER_02: If I'm gonna go build it, then I'm forced to write down and spec out what that is, right?
- 26:15SPEAKER_02: I have to pay a lot of attention to what it is I'm doing.
- 26:19SPEAKER_02: Uh, and I have to be really, really specific about it.
- 26:22SPEAKER_02: And I think that that I think there's an element of it, like I also I also feel like we're still in this phase where people are like shitting on AI product, like it's really popular to shit on like an AI-made product.
- 26:38SPEAKER_02: Um I think we'll get over that.
- 26:42SPEAKER_02: And I think I think it's a result of like everybody just kind of being allowed to put there's still a lot of people getting away with putting out crap and slop and whatever.
- 26:51SPEAKER_00: It's like uh GeoCities websites and early internet, right?
- 26:55SPEAKER_00: Like we all look back on that now with some nostalgia, right?
- 26:58SPEAKER_00: We're like, oh, that was cute, you know.
- 26:59SPEAKER_02: But now you just aged yourself and me a lot.
- 27:02unknown: Yeah.
- 27:04SPEAKER_00: Hey man, one of my first uh coding classes, HTML, was GeoCities website.
- 27:12SPEAKER_00: Yeah, exactly.
- 27:12SPEAKER_00: My space is coming back.
- 27:13SPEAKER_01: Are you gonna get on MySpace?
- 27:14SPEAKER_01: It's coming back.
- 27:16SPEAKER_00: I I I have to.
- 27:17SPEAKER_00: I have to.
- 27:18SPEAKER_00: It's not me too.
- 27:19SPEAKER_00: It's a requirement.
- 27:20SPEAKER_00: Yeah.
- 27:20SPEAKER_00: People my age.
- 27:22SPEAKER_00: Um, but hey, actually, George, like this project that you just told me about could be a perfect example of how I think about building and creating workflows with AI.
- 27:32SPEAKER_00: If you want, we could go through this project and you could see how I would approach it as somebody who's a non-expert.
- 27:40SPEAKER_00: Or I could talk about something that I've built, but it might be kind of interesting because I'm not an expert in CMMS systems, for you to kind of see what the process looks like.
- 27:49SPEAKER_00: Okay.
- 27:50SPEAKER_00: Or hear what the process looks like.
- 27:51SPEAKER_02: Why don't why don't we do why don't we do that?
- 27:53SPEAKER_02: We can uh uh let let's let's decide on a workflow we want to build and then we'll kind of walk through it um even at a conceptual level because I because I do yeah, I do want to like give people some like a here's how to here's how to think about AI.
- 28:08SPEAKER_02: So maybe we could do CMMS or maybe we could do, I'll say there's a there's a um there's one that I've there's one that I've built like 60% of that I think might be interesting too.
- 28:20SPEAKER_02: And you tell me your reaction here.
- 28:22SPEAKER_02: Uh having like a CRM system for field reps that my specific my specific use case is like um is like uh food service sales.
- 28:36SPEAKER_02: So something that you might not have a lot of like personal experience with, but like this is a pretty broad thing that people in my network will be really familiar with, is that like food service sales specifically is a very in-person business, belly-to-belly business is you know kind of what they say about it.
- 28:54SPEAKER_02: Um, but the idea is like, hey, there's these really like broken up, like fractured kind of uh distribution networks.
- 29:02SPEAKER_02: And whereas like in retail sales, like retail sales are kind of like B B2B, right?
- 29:08SPEAKER_02: Like I go to you know Walmart once a year and I meet with like one or two people who make category decisions for thousands of stores.
- 29:17SPEAKER_02: That is a lot less that's a lot less common in food service sales.
- 29:22SPEAKER_02: Food service sales is mostly like, hey, I'm building out around you know, this Central Texas distribution center.
- 29:28SPEAKER_02: Like Cisco, for example, has has has one distribution center that covers like Austin down to or uh technically it's Waco down to San Antonio.
- 29:37SPEAKER_02: And so if you're building out that distribution center, you've got somebody, some some rep is going into the store, going into the physical locations.
- 29:47SPEAKER_02: There's not as many like calls.
- 29:49SPEAKER_02: And so like I had had this thought about like, hey, can I have like a voice interactive thing?
- 29:56SPEAKER_02: So these people are the their the time that they've got, because you have a lot of sales reps who are wasting half of their week updating their CRM.
- 30:04SPEAKER_02: So the thought was like, hey, what if they've got the list of like accounts?
- 30:09SPEAKER_02: What if they can interact with their voice, right?
- 30:12SPEAKER_02: What if the rest of the sales network had that information?
- 30:17SPEAKER_02: And then what if there was some what if there was some database that I could opt into that would say like, hey, if I am opting into this, I'm gonna share inform I'm gonna share meta information with it with a broader network about like, well, the buyer for, you know, the buyer for pluckers has changed.
- 30:35SPEAKER_02: You know, the the buyer for pluckers has changed, and this is the new person.
- 30:39SPEAKER_02: And that information can be part of that broader network as long as I'm if I'm also sh getting information from that broader network, but the opt-in is kind of like optional.
- 30:49SPEAKER_02: So we could build that or we could build CMMS or at least talk through them.
- 30:54SPEAKER_00: I'll I'll let you pick.
- 30:55SPEAKER_00: It really doesn't yeah.
- 30:56SPEAKER_00: Let's do the second one that you said.
- 30:58SPEAKER_00: Um the uh distributed kind of CRM for food um services reps.
- 31:04SPEAKER_02: Great.
- 31:05SPEAKER_02: Why don't we take a break?
- 31:06SPEAKER_02: Because I feel like that's we're right at 30 minutes.
- 31:08SPEAKER_02: I want to uh I want to get into that.
- 31:10SPEAKER_02: So when we come back, uh me and Jack are gonna build a uh CRM.
- 31:15SPEAKER_02: Why why don't we because we've got a lot of food people uh on this listening to this podcast, why why don't we uh why don't we build it and give people access to the repository?
- 31:25SPEAKER_02: I'm I'm making big I'm taking big swings, making big promises.
- 31:29SPEAKER_00: We totally could.
- 31:30SPEAKER_00: It might take me like to actually build, build, um it might take me like a month because I'm traveling a lot.
- 31:35SPEAKER_00: That's all right.
- 31:36SPEAKER_00: I'm you're hired, man.
- 31:36SPEAKER_00: I'm hiring you.
- 31:39SPEAKER_02: Um okay all right, we'll be we'll be right back.
- 31:41SPEAKER_02: We're gonna we're gonna walk through um how to build uh how to build a software for an actual company with AI, with Jack.
- 31:50SPEAKER_02: Stick around.
- 32:55SPEAKER_02: We're talking with Jack Atlasov, and we are gonna build or at least talk through how to build a uh a CRM for our out in the field uh mobile sales team.
- 33:08SPEAKER_02: So uh Jack, we're not gonna do, we talked about this in the break a little bit, but we're not gonna do like a you know hands-on keyboard uh we're building the theme thing in real time because that would take longer than the time that we've got.
- 33:21SPEAKER_02: Uh maybe if there's maybe if there's interest from the community, we'll uh you know spin this into something that people can uh toy with or you know, pull apart themselves.
- 33:33SPEAKER_02: Um but I think what I what I want to do is I I want to I'll kind of I'm gonna kind of because I think one of the ways that you build stuff, um, and we we went over this in the workshop, is like you have you have AI kind of query you, right?
- 33:48SPEAKER_02: Like do the interview, figure out what you need, what what the actual and so this is a really good skill to have.
- 33:56SPEAKER_02: Um let's let's kind of walk through how we would how we would build this.
- 34:00SPEAKER_02: So yeah.
- 34:01SPEAKER_02: So if I'm saying like, hey, what I want is a CRM for uh for like field sales, specifically for like my food service sales force who's gonna be out talking to restaurants, they're gonna be saying, hey, we've got yellow bird, whatever.
- 34:17SPEAKER_02: And then I want them to um that this is kind of like uh all sales teams, right?
- 34:23SPEAKER_02: All sales teams have this thing where they're trying to do the work, do the sales work, have the calls, have the meetings, and then update their CRM, um, which can which for a lot of salespeople is like a drag because updating the CRM doesn't close any business at all.
- 34:39SPEAKER_02: It just kind of lets management know that everything's going the way it needs to go, gives them their follow-ups, all this sort of stuff.
- 34:47SPEAKER_02: So, so where do you where would you start with this?
- 34:49SPEAKER_02: How do you want to start thinking about this?
- 34:51SPEAKER_00: Okay.
- 34:52SPEAKER_00: So I would actually start two steps before the actual like querying piece of it.
- 34:59SPEAKER_00: Um and this is where people who are not familiar with um some of like the research processes that are involved, or at least my processes, this this is where this this might be completely new to them, but I think this is like unbelievably important um and where a ton of the value is.
- 35:20SPEAKER_00: So what I do at the very beginning of and basically almost anything that I've never done before, which I've never built a CRM for distributed sales force for food services businesses, right?
- 35:34SPEAKER_00: So the very first thing that I do is I adopt a posture of epistemic humility.
- 35:40SPEAKER_00: Okay.
- 35:41SPEAKER_00: So what does that mean?
- 35:42SPEAKER_00: Yeah.
- 35:42SPEAKER_00: So what does that mean?
- 35:43SPEAKER_00: It means I'm humble and I I remind myself that I don't know what I don't know.
- 35:51SPEAKER_00: And that's where I start.
- 35:53SPEAKER_00: And so what does it mean to start from a place of not knowing what you don't know?
- 35:56SPEAKER_00: So, George, you have a lot of knowledge on like CRMs, and you probably have a lot of knowledge on food services, and you probably have a lot of knowledge on sales, distributed sales forces.
- 36:10SPEAKER_00: But um if I, if I was in the the position like I am now of building something like this, I would actually, even if I had that knowledge, which I don't, I would actually pretend that I didn't.
- 36:23SPEAKER_00: And I would start the whole thing off with a skill that I like to use called Scout, which I built, which is all about uncovering all of the research, all of the thinking, all of the frameworks, methodologies, concepts, and even the words themselves, like the actual words within this universe of CRMs, distributed sales forces, foods, food services, right?
- 36:53SPEAKER_00: I want to take in a ton of data because the the truth is I have some vast universe of knowledge about any given topic, or I have some universe of knowledge about any given topic, but I have no way that I can tell, right?
- 37:11SPEAKER_00: I have not been able to find a way in in this time that I've been on this planet of knowing where the boundary is between where my knowledge ends and where the true end of knowledge for a given topic is, right?
- 37:25SPEAKER_00: So there's some universe of understanding that I have, and it's buried or it's encapsulated within the broader universe.
- 37:34SPEAKER_00: And it's possible, it's it's probably likely even, if I don't know a lot about a subject, that the best solution or even just a more optimal solution is somewhere beyond the event horizon of my knowledge, like at this moment in time, right?
- 37:52SPEAKER_00: So, my first role in this whole process, my first job is to expand the event horizon, expand the uh boundary of the observable universe of my knowledge, right?
- 38:04SPEAKER_00: So I'm gonna spend some amount of time doing that.
- 38:06SPEAKER_00: I don't know how long, you know, it could be an hour, it could be a day.
- 38:09SPEAKER_00: Just depends on how interested I am and what I'm what I'm doing.
- 38:13SPEAKER_00: And that might look like a hundred different runs of Scout, which essentially you could think of as just like deep research that's focused on specifically these unknown unknowns, terminology, methodology, concepts, research frameworks.
- 38:29SPEAKER_00: You know, what research have scientists and researchers done on these topics?
- 38:33SPEAKER_00: Like, what have practitioners found to have worked really well?
- 38:37SPEAKER_00: Like maybe the key to the whole thing is like removing friction from a distributed network, identifying the nodes in the graph of the network that are hyperconnected to every other node and blah, blah, blah.
- 38:51SPEAKER_00: Right.
- 38:51SPEAKER_00: Maybe that's the key.
- 38:53SPEAKER_00: And if I don't somehow discover that, then whatever I end up building will probably not be as good as it could have been.
- 39:01SPEAKER_00: Right.
- 39:02SPEAKER_00: So I at least want to take a few moments to a few hours to maybe even a day, you know, of doing this work, right?
- 39:11SPEAKER_00: So that's where I start.
- 39:13SPEAKER_00: I start with expanding the boundary of, you know, my my observed knowledge.
- 39:17SPEAKER_00: Um, and then from there, I do this process that I call signal boosting or, you know, increasing signal reducing noise.
- 39:26SPEAKER_00: And I talked about this a little bit in my my workshop.
- 39:28SPEAKER_00: And the nice thing I'll say about the first part of what I said, which is like that research and then the signal boosting thing, is everything like like you mentioned before on this call or on in this podcast, everything is changing so quickly.
- 39:42SPEAKER_00: One of the issues that we have, I think, with any quickly changing technology is like where do we put our focus, right?
- 39:48SPEAKER_00: Where do we put our effort?
- 39:50SPEAKER_00: I'm very uh sensitive to monitoring what are the things that I'm doing that don't change that much over time.
- 39:58SPEAKER_00: Because a lot of I can then accrue a lot of value to that, right?
- 40:01SPEAKER_00: Like I as I repeat these workflows and pay attention to what doesn't change from project to project or from workload workflow to workflow, I ascribe more value to that, right?
- 40:14SPEAKER_00: And if things are changing and falling off and new things coming in, I ascribe less value to that.
- 40:20SPEAKER_00: You know, so I'm optimizing for stickiness over time in terms of methodology, right?
- 40:26SPEAKER_00: So this first thing, this research thing, and the second thing, which is this signal boosting piece, these have been with me now for basically almost a year, you know, in the shape, broadly speaking, like in the shape that they were when I first kind of like adopted them.
- 40:41SPEAKER_00: So I have a lot more conviction to say to your audience, you know, like this is worth it.
- 40:46SPEAKER_00: Like it's worth spending the time to do the research, go broad, go general, you know, take in a ton of information.
- 40:54SPEAKER_00: Let your AI pull out patterns and insights and unique takeaways and non-intuitive, you know, connections.
- 41:03SPEAKER_00: And like, and actually, there's some interesting research from the theorem-proving side of AI research, where one of the most fruitful um uh kind of like uh workflows that they've developed currently to solving frontier math problems with AI is they specifically instruct it to explore all possible paths, even paths that look like they wouldn't initially be promising.
- 41:31SPEAKER_00: So, because it's a latent space, right?
- 41:34SPEAKER_00: It's a vast inside these AIs, right?
- 41:36SPEAKER_00: Like there's it's a vast universe of knowledge.
- 41:39SPEAKER_00: And it's very easy for the AI to get pointed in some direction and miss all this other, you know, interesting stuff over here.
- 41:47SPEAKER_00: So I really just want to give it a ton of time and info to, you know, go go wild, you know, have fun.
- 41:54SPEAKER_00: And I'm gonna put what is I'm sorry, go ahead.
- 41:57SPEAKER_00: Yeah.
- 41:57SPEAKER_00: Well, I was just gonna say in practice, this can look like a whole bunch of research that gets stored on files in your computer or in files in your computer.
- 42:05SPEAKER_00: It can look like a whole bunch of research that gets stored in a GitHub repo, or you could even use Google's notebook LM product and just throw everything inside of you know Google's Google Cloud's notebook LM.
- 42:15SPEAKER_00: But somehow taking the time, capturing this research, developing terminology, developing concepts, identifying patterns, and then you know, going from there.
- 42:26SPEAKER_02: So so I I'm gonna say uh I love this.
- 42:29SPEAKER_02: I I think that I've got a version of this that I do as well.
- 42:35SPEAKER_02: And and there's a tur there's a term that I've found at least uh because I'm mostly uh building stuff with Claude Code.
- 42:42SPEAKER_02: I think you use maybe codex primarily or um codex more, yeah.
- 42:47SPEAKER_02: I know that Claude Code at least likes the term prior art.
- 42:52SPEAKER_02: So that's one way to get like prior art.
- 42:55SPEAKER_02: Um, you know, use prior art is a simple kind of like thing that at least if you're using Claude Code, and I can't speak to Codex because I haven't used it that much, but like in Claude Code, you have this idea, this concept of like use prior art, right?
- 43:10SPEAKER_02: Look for prior art, where where it it it does kind of open up that um it's not quite the same as like your scout skill.
- 43:19SPEAKER_02: It's not quite the same as like yeah, as that, but it is saying it is introducing the concept of like I'm not the first person to think about building a CRM, right?
- 43:29SPEAKER_00: And that's why I think oh sorry, go ahead.
- 43:33SPEAKER_02: No, I was just gonna say there so prior art will encapsulate like what is Salesforce and how does Salesforce work.
- 43:39SPEAKER_02: So it will encapsulate the enterprise, right?
- 43:42SPEAKER_02: And there's a lot that there's a lot that's been written about Salesforce.
- 43:46SPEAKER_02: Anything that's been written about Salesforce is available to your AI, right?
- 43:50SPEAKER_02: Like any code that's been shared, and you know, it can browse their websites and and learn.
- 43:55SPEAKER_02: Like the way that I built my first CMMS system was I see.
- 44:00SPEAKER_02: saved the screen recording like I and the transcript of the of of a bunch of pitches that I got on CMMS systems.
- 44:08SPEAKER_02: Yeah.
- 44:09SPEAKER_02: And I use some of that stuff as prior art.
- 44:10SPEAKER_02: I was like, look at their website.
- 44:12SPEAKER_02: Like you can go through the d your AI can go through the product demo.
- 44:16SPEAKER_02: If there's a YouTube product demo, if some if there's a video of somebody saying here's how this software works, here are the features, that's prior art and you can put it in there.
- 44:26SPEAKER_02: So I think that what you're the this first step I'm going to say is like get smarter, i.e, don't start from scratch.
- 44:34SPEAKER_02: Like you're the what you know about CRM is not what is known about CRM.
- 44:40SPEAKER_00: So that's right.
- 44:41SPEAKER_00: And that's where I that's why I say like start from a position of humility, right?
- 44:46SPEAKER_00: Like even when I'm an expert, if I care about what I'm building, right?
- 44:52SPEAKER_00: If I don't really care that much about what I'm building, maybe I'm just going to drive the whole thing from the beginning, right?
- 44:56SPEAKER_00: I'm going to be like, hey, my slice of this universe of knowledge, right?
- 45:00SPEAKER_00: My slice of this, I call it, I think of it sometimes as like a tree of knowledge, right?
- 45:04SPEAKER_00: Like a semantic tree.
- 45:05SPEAKER_00: Like, hey, maybe I'm good with this.
- 45:07SPEAKER_00: Like I don't really need this to be amazing, right?
- 45:10SPEAKER_00: I just need to solve my problem today or for the next couple months or whatever.
- 45:14SPEAKER_00: Fine.
- 45:15SPEAKER_00: No big deal.
- 45:16SPEAKER_00: But if I really care about what I'm building, I want to start from a position of humility because I'm probably not the first person to solve this problem.
- 45:25SPEAKER_00: You know?
- 45:27SPEAKER_00: And again, that comes a lot from my scientific kind of training background, like just looking at research papers and realizing that, oh, there's a whole you know sources cited section at the bottom of these papers.
- 45:39SPEAKER_00: And like these researchers are standing on the shoulders of other researchers all the way going back to the Greeks and to before even the Greeks, right?
- 45:48SPEAKER_00: So I think like you're building with this process, you're building a semantic tree of knowledge.
- 45:57SPEAKER_00: And then the other thing I like to use for this, like you mentioned, is you know, you're looking at like um interesting implementations of what you've done in the past already that exist.
- 46:07SPEAKER_00: I have another skill that I built called Genome, which asks the question of something in the world, right?
- 46:15SPEAKER_00: It sees some like you point it at something in the world, right?
- 46:18SPEAKER_00: And it asks the question what ideas, thoughts, patterns, beliefs, realizations, concepts, right, had to have been had in order for this thing to come into existence.
- 46:32SPEAKER_00: Yeah.
- 46:33SPEAKER_00: So it basically, you know, you could point it at Salesforce, right?
- 46:37SPEAKER_00: And it can say like, oh, okay, you would have had to have realized that there was like all this, you know, these problems and and these, this unmet demand and like this, that, and the other, right?
- 46:48SPEAKER_02: So I like to use those skills, genome and scout, to kind of start to populate this and we're gonna treat we're gonna link to those skills, those skills you have available on GitHub along with some other a few others.
- 47:00SPEAKER_02: Signal sweep being one that uh I've really liked and kind of customized on my own.
- 47:06SPEAKER_00: Yeah.
- 47:06SPEAKER_00: And Signal Sweep is another kind of research skill exactly.
- 47:08SPEAKER_00: Yeah.
- 47:09SPEAKER_00: So so anyway, that's the first step.
- 47:11SPEAKER_00: The second step you actually alluded to which is this signal to noise boosting you alluded to with your YouTube example.
- 47:17SPEAKER_00: So there are pockets of knowledge that have a much higher value to you as you're building this system or this tool than other pockets of knowledge.
- 47:28SPEAKER_00: And actually we know this to be true because when we look at LLM performance and we see the very big models, right?
- 47:37SPEAKER_00: The 10 trillion parameter models that are pre-trained on like the entire internet and they have some amount of performance and intelligence, right?
- 47:45SPEAKER_00: And then we see this process what's called distillation, right?
- 47:49SPEAKER_00: Where a much smaller model is able to achieve almost the same performance for much less cost or at much faster speeds.
- 47:59SPEAKER_00: So what the hell is happening there, right?
- 48:01SPEAKER_00: Well basically this thing that we're talking about is happening there where there are the scientists and the researchers that are and the engineers that are um making these models are essentially helping the model understand what data is most useful and what data is not as useful.
- 48:20SPEAKER_00: So they're boosting signal and decreasing noise, right?
- 48:23SPEAKER_00: So they're pruning it kind of like a bonsai tree, right?
- 48:25SPEAKER_00: Like they're they're pruning all the unnecessary bits and keeping the very high signal.
- 48:30SPEAKER_00: And what you'll find is oftentimes you can prune 90% of the data and just have the 10% that's really, really strong and you'll get 90 plus percent of the of the result.
- 48:44SPEAKER_00: So it's worth going through and doing this kind of pruning step.
- 48:48SPEAKER_00: And the way you do it is you use the AI to basically identify these patterns and go through your data and see where things are contradicting where they might not be as helpful you know where again where um there's uh you know superfluous information that's like duplicated, you know, unnecessary right and you just compress, compress that context, right?
- 49:08SPEAKER_00: Compression has an intelligence quality all of its own.
- 49:12SPEAKER_00: Right.
- 49:13SPEAKER_00: So you could have like two points on a map and there's lots of different ways to get from point A to point B.
- 49:18SPEAKER_00: And one of those ways is a straight line and the straight line will take less time right so that that will taking less time has a quality of intelligence all of its own basically so um so those are the first two steps.
- 49:35SPEAKER_00: Then the third step what are we going to call step number two?
- 49:37SPEAKER_02: So step step one is like get smarter step two is compressed.
- 49:42SPEAKER_00: I call it compress.
- 49:43SPEAKER_00: Yeah yeah compress or you could think of it as signal or compression get smarter boost signal get smarter compress.
- 49:52SPEAKER_02: So I think that the like I and I agree with these first two steps this is a good distillation of how I am building stuff today with AI and I think it's important to note that we've made it through the first two steps and we haven't started building a CRM yet we just are creating and these these steps are important right because like if you just if you just give a prompt right like a lot of the stuff that we're considering slop I would consider stuff that has been that has not really been considered it hasn't gone through these first steps of like let me think about the problem let me use the capabilities of AI to build a knowledge base and then and then compress out of that knowledge what do I really need to start making to start get my hands on the keyboard and like start building something so we got smarter step one we compressed step two what what are we doing at step three step three is we're gonna start the building process.
- 50:55SPEAKER_00: So um step three and building is really it's a process it's not you know an action that happens once you can think of it as like a loop um I guess fundamentally which is um the loop looks like this it it's context alignment building and and prototyping verification and then repeating back into context alignment.
- 51:23SPEAKER_02: So context alignment buildingslash prototyping validation verification and then more context alignment that's how I like to build um I'm not saying that's like the only way to build but once you've done step one and step two uh that's how I I do step three one of the things uh one of the things that come that seems to come up over and over in the stuff that you've built and written about and talked about is that verification part of the loop.
- 51:51SPEAKER_02: I feel like as much of a as bullish as you are about AI and building with AI, I feel like you're a recurring theme that I see from you and from your work and from your repositories and stuff is assume that this is wrong.
- 52:09SPEAKER_02: Like assume this is wrong.
- 52:10SPEAKER_02: Like and and I think that this is a really valuable thing I don't want to skirt over right because there is this thing of like um we could we can maybe talk a little bit about um like agent coordination um here where it's like um and this is this is an idea that I've kind of I I think that when I uh when we did that workshop there I know the workshop was kind of primarily about e-com but there were a lot of other things that came out of that and for me it kind of it kind of pushed me forward in my I in my thought about and my ideas about like agent coordination and so like the the idea that like you have you're having an agent that builds v1 of the software right like we're we're over in phase three which where we're building and testing and you know verifying and we're kind of doing this loop over and over again that kind of exists in phase three before we get to like a quote unquote final output or final draft.
- 53:14SPEAKER_02: And it's important and I'm I'm kind of like uh paraphrasing stuff that I've heard you say but I do think it's important to not have the same agent or the same the same thing, the same you know agent that builds the software test the software right that you've got that you you don't have the author self-editing you're sending like my that's my mental model for it is like hey I've got you know Stephen King over here working on a book I want him to send that book to somebody else to edit so that I'm getting another perspective on it and not just the same person who's who's been like tunnel vision on like I'm creating the best version of this stuff.
- 54:01SPEAKER_02: You want like a different agent or a different team to red team this stuff, right?
- 54:08SPEAKER_00: Absolutely.
- 54:09SPEAKER_00: Yeah and you can do this in different ways.
- 54:12SPEAKER_00: So you can do it uh and you you know one might want to do this in multiple ways all in the same project.
- 54:19SPEAKER_00: Like for example you can do it at the level of agent coordination but you could also do it at the level of the actual like um spec design or the actual implementation plan of the project.
- 54:32SPEAKER_00: So you may at the end of all this research right and the signal compression you may at some point then have like a plan that you put together for what you want say phase one of like the minimum viable product to be for this distributed CRM um or this CRM for distributed sales people for for um food or for for service like what was it for restaurants basically right yeah so anyway so you have some MVP right um but you could if you were so inclined to create for every feature that this project is going to have you could as you plan out these features you could add for every feature what does success represent like what does it look like what functionality has to happen for success to be achieved and if success is not achieved then that feature is not complete and that feature is literally to the LLM is told to be failing uh by default right so you assume failure from by default yeah you have to prove yeah you have to prove success you don't prove failure you prove success you assume failure and that sets up a much more robust uh testing framework or it sets up it's basically like you know you're you're you're making sure your soil is healthy so that the plants that grow out of that soil have all the you know um inputs that they need to be you know healthy plants right so you if you're a gardener right or you're a farmer you we I I mean I used to live on a horse ranch so I I know a lot about this like we really care you know about like the quality of the soil and the quality of the the inputs to the soil right so um so yeah you can do it at that level you can do it at the build verify different agents level as well and now the tools make that a lot easier to do um you can literally add that into your CloudMD or your project level CloudMD or your codexagents.md or you can put it in to your configs you can like you know hard code essentially this like build verify loop into your settings in your uh harnesses like Claude code or codecs or whatever your preferred harness is so um yeah what I like to do is I I like to do both I like to have it both assume failure for all the different you know granular pieces of the features that I'm building and then also have both an agent builder and a verifier as well.
- 57:17SPEAKER_00: And then that's just kind of set up that's just kind of like created and set up but then from from that point I do this whole process of context alignment.
- 57:27SPEAKER_00: So I when I have a plan, right?
- 57:30SPEAKER_00: When I have this initial plan, even though I may have like this infrastructure set up where there's a builder verifier and there's specs that are uh default failing I don't start building yet I go through one more step which is context alignment.
- 57:45SPEAKER_00: So I run a skill called Wayfinder or there's another skill which includes Wayfinder inside of it includes this skill called grill with docs.
- 57:54SPEAKER_00: And so there's grill me sometimes people call it grilling there's grill with docs.
- 57:59SPEAKER_02: These are all by this guy named Matt Pocock which he has an open source repository on GitHub which is called uh agent skills I think but you can find it we can link it in that I want to I want to maybe pause pause here and talk about because I think that we're in uh we've stepped off now into a pretty deep pool uh yeah where where I think that there's um there I mean we could do another two hours of this right but we don't have that amount of time so when we're talking about uh we're gonna link here to like your um GitHub but like let's talk just briefly about what a skill is and then and then I want to talk briefly about how how an app like this would get deployed because there are there's there are these connectors which are becoming more automatic but like these APIs where you used to have to go like wrestle all these connections together right like the the the platforms the you know uh frameworks etc like these they're making connection a little easier um skills are are kind of uh well tell tell me tell tell us what skills are and like how we go about building skills in our um in our workflows yeah so skill you can think of skills essentially as just a a prompt that has been captured in in time like you know a photo of a prompt you really like this prompt you want to kind of have it in your back pocket for future use and um that that could be all a skill is right is it's just a prompt that doesn't change over time that you've found hopefully through testing actually improves the outcome when you run that prompt with whatever you're trying to you know work on where that prompt is relevant.
- 1:00:00SPEAKER_00: And skills um also can include tool calls, you know, uh pieces of code so they can do certain functions.
- 1:00:10SPEAKER_00: They can perform certain functions for you.
- 1:00:12SPEAKER_00: They don't have to they could literally just be text that just is consistent where you know you reach for this skill because you like this prompt.
- 1:00:20SPEAKER_00: It generates good outputs.
- 1:00:22SPEAKER_00: The number one missing piece that I see with people with skills is they just will have a skill and they'll never actually investigate whether it improves the outcome for them or not over baseline, over not having the skill at all, right?
- 1:00:36SPEAKER_00: So that's why all basically all my skills come with evals, which are like tests that the LLM will run against the output of the skill versus not using the skill.
- 1:00:48SPEAKER_02: And if the skill is not additive then you know you don't really need to the reason I asked you specifically about skills is that you've got a repository on GitHub called Skillforge which is I think got like 884 stars on it.
- 1:01:04SPEAKER_02: Yeah.
- 1:01:05SPEAKER_02: So probably your most popular like on GitHub so far um which is about like validating skills creating and validating skills so check out skillforge um let's talk a little bit about like the so let's go back to the CRM example right because we're in build phase um we can talk about whatever phases are going to come after the build phase but like we've got a loop going here if I'm saying like okay because we haven't done this like research pass this full research pass like on this call there's probably yeah I'll just say a lot of stuff that we don't know right there's probably a lot of prior art that we're not bringing in a lot of like expertise it's just me and Jack talking through it right now.
- 1:01:53SPEAKER_02: With that being said there are some things that I would expect to come into play with a with an app like this and and one of those things is like I need a managed database right so I might be in Postgres I might be in something else but like um there are like I can be in Superbase for like my managed database.
- 1:02:15SPEAKER_02: So I might be connecting this project to something like Superbase.
- 1:02:19SPEAKER_02: I might be in some kind of a cloud I might be in some kind of SQL library.
- 1:02:25SPEAKER_02: I might just run it maybe I decide I want to just run it inside of like maybe I'm maybe the my workspace is Google Gemini maybe I'm just running it inside of like Google as part of like an app script or something there's a bunch of different ways but like at some point if I've got a CRM the backbone of a CRM is a database right so you'd expect there's some kind of a database a way to manage the database you're not having just the point I'm trying to make is that you're not having yeah you're not having the LLM re rely on its LLM memory to go look for this database.
- 1:03:02SPEAKER_00: You are going to connect and implement some kind of a database solution yeah and that's where the context alignment becomes really important because yes maybe you've done a bunch of research and boosted a bunch of signal but now like we really need to make sure that the vision that you have in your head and the vision that the LLM has are aligned.
- 1:03:24SPEAKER_00: And maybe some part of that vision you move towards what the LLM is suggesting and maybe some parts of that vision you request the LLM move closer to like what you need because you just know you need certain things.
- 1:03:36SPEAKER_00: So for example with the with the database example right let's say the LLM has identified you know these four different databases that people are using for things like this.
- 1:03:49SPEAKER_00: But you have some contract with like a database provider right that like you're just locked into this contract for the next year.
- 1:03:57SPEAKER_00: You might as well get the most out of it.
- 1:03:59SPEAKER_00: So like you're gonna tell the LLM no we're gonna use this database even if it says like this isn't maybe the best for this but it's okay.
- 1:04:07SPEAKER_00: You're gonna be like okay you know Mr.
- 1:04:09SPEAKER_00: LOM like we're gonna use this one right like that and there's a whole bunch of that kind of stuff that should happen.
- 1:04:16SPEAKER_00: Again, unless you don't really care that much about the project you're you're pressed very pressed for time.
- 1:04:21SPEAKER_00: But like this context alignment piece around like you know what does it mean to you to to reduce friction in a CRM?
- 1:04:29SPEAKER_00: You know like what what does that look like to you?
- 1:04:31SPEAKER_00: Like what what are the, you know um let's just say, because we didn't do the research let's just say for for example that like one of the key things to get something like this to work is like identifying like the key super connectors within the sales rep graph, you know, that absolutely need to have access to the latest data and they should also be empowered somehow to like distribute information to like other sales people in the network.
- 1:05:00SPEAKER_00: Let's just say that came out of this research process, right?
- 1:05:03SPEAKER_00: So like there may be some aspects of that particular piece of knowledge that are truly only available to you.
- 1:05:15SPEAKER_00: Like the LLM would never find that information out in the world, right?
- 1:05:19SPEAKER_00: No matter how much it did research or whatever.
- 1:05:22SPEAKER_00: So when it surfaces its suggestions to you around how should we design the dashboard to reduce friction, right?
- 1:05:29SPEAKER_00: You may actually have like a very clear viewpoint that you want to give to it and say, you know, um actually that's not how this industry works.
- 1:05:38SPEAKER_00: And like these people are not thinking about it this way.
- 1:05:41SPEAKER_00: They're not thinking about it from the standpoint of like networks and graphs and this and that and you know no like we're gonna go this way.
- 1:05:50SPEAKER_00: You know, we're gonna take this other path.
- 1:05:52SPEAKER_00: And like you own that decision.
- 1:05:55SPEAKER_00: So at the end of this process you're you're still very happy with the result ideally, right?
- 1:06:00SPEAKER_00: Because you were part of that decision process.
- 1:06:02SPEAKER_00: And that process is called domain-driven design.
- 1:06:06SPEAKER_00: So if people are curious about it, you know, this comes out of Stanford and a lot of like engineering uh best practices where you have somebody that has a deep understanding of the actual domain that the thing is being built for, whether it's a workflow, software, whatever.
- 1:06:22SPEAKER_00: And they're paired with, in this case, an LLM, but in the past it was like, you know, an engineer.
- 1:06:27SPEAKER_00: And every once in a while, like the domain expert should probably raise their hand and go, actually, we're gonna go in this particular direction, you know?
- 1:06:36SPEAKER_00: Yeah.
- 1:06:37SPEAKER_00: So that's where the human taste and judgment and deep understanding comes in.
- 1:06:41SPEAKER_02: Right, right.
- 1:06:41SPEAKER_02: Uh yes.
- 1:06:42SPEAKER_02: And it's like solving your, I think this is I think this domain expertise, because the domain expertise doesn't necessarily need to be like you can get hyper, hyper domain specific.
- 1:06:54SPEAKER_02: And it could be like, okay, I want it to be specific to how my company thinks about this thing, right?
- 1:07:00SPEAKER_02: Because that's kind of the power of like, I, you know, I've I've made a statement and uh that you know SaaS is dead, and I don't think that's true yet.
- 1:07:10SPEAKER_02: I don't think it's true yet, right?
- 1:07:12SPEAKER_02: It will it will be though, and it's like you can personalize something fairly small, right?
- 1:07:20SPEAKER_02: You could personalize something pretty big, right?
- 1:07:22SPEAKER_02: Like I think there's still there's still maintenance, like the bigger it gets, there's maintenance issues, there's drift issues, there's like you know, if I'm just making something for a team of five people, I can I can do it like this.
- 1:07:36SPEAKER_02: And my domain expertise can be so specific that it's not just here's how the food service sales industry works, because I'm not trying to sell the full, I'm not trying to sell the whole industry.
- 1:07:49SPEAKER_02: I'm just trying to give my team in some instances, like that there's a thing that I see happening right now where people are building these tools with AI and then they're trying to sell them to the industry.
- 1:08:01SPEAKER_02: And that my my opinion is stop it, stop it.
- 1:08:05SPEAKER_02: I I think I think stop it, right?
- 1:08:07SPEAKER_02: Because like the tool that we're paying for is the capability of like the AI system to build the tool.
- 1:08:16SPEAKER_02: Um, and so like if you're selling something for$300 a month at some custom vibe coded thing, I think you have the ability to go, like if you've got somebody in your company, you've got the ability to go uh a lot more like hyper-specific um with that domain expertise, the domain expertise being like the exact problem that you're solving, not just the industry problem that you're solving.
- 1:08:39SPEAKER_02: Um and then when you are um when you're deploying this, so like just to just to kind of go back to systems a little bit, you're gonna likely end up with some database connection.
- 1:08:51SPEAKER_02: So there's a you you're gonna have to make a connection here, you're gonna have to host this app somewhere.
- 1:08:57SPEAKER_02: Yeah.
- 1:08:58SPEAKER_02: So like there's gonna be the LLM can suggest stuff.
- 1:09:02SPEAKER_02: And this is this is an area where like that I'll say specifically, like if you've never built a little app or a little thing with AI yet, I would encourage you to do this as a weekend project, right?
- 1:09:14SPEAKER_02: Like the you know, the CRM.
- 1:09:16SPEAKER_02: If the CRM is kind of like your pain point for like a small distributed sales team, spend a weekend doing this stuff that me and Jack are talking about.
- 1:09:25SPEAKER_02: Um, and you're probably gonna end up with a lot more questions than answers the first time you do it, right?
- 1:09:32SPEAKER_02: Because there is gonna be, and I remember like doing these things for the first time as somebody who's not a software engineer and being like, okay, well, how do I deploy this?
- 1:09:40SPEAKER_02: Right?
- 1:09:40SPEAKER_02: Like, if I want somebody to use this app, well, where does the app live?
- 1:09:44SPEAKER_02: Like, does it live somewhere?
- 1:09:45SPEAKER_02: It does live somewhere, right?
- 1:09:47SPEAKER_02: And you can deploy it on like a lot of my apps live in Vercel.
- 1:09:50SPEAKER_02: I think that you use Vercell too for like, yep, and that and these these like frameworks are getting better by the day, right?
- 1:09:59SPEAKER_02: Like totally.
- 1:10:00SPEAKER_02: Um, yeah.
- 1:10:01SPEAKER_02: And like Replit, I think, has their own, they've got their own Git repository now, so you don't even have to connect GitHub.
- 1:10:08SPEAKER_02: So like if you're on Replit, it's just like I was on Replit, it was too expensive to like really dive in and build stuff a while ago.
- 1:10:15SPEAKER_02: I haven't used it.
- 1:10:16SPEAKER_02: Have you used Replit in a while?
- 1:10:18SPEAKER_00: I did a while ago, similar to you.
- 1:10:20SPEAKER_00: Yeah.
- 1:10:20SPEAKER_00: And and they are trying to be super vertically integrated, you know.
- 1:10:23SPEAKER_00: And I think if uh Vercel wasn't so friendly with like Cursor and all of that, they probably would build their own Git replacement too.
- 1:10:32SPEAKER_00: But yeah, we'll start to see this kind of like complete vertical integration across the whole stack of everything you need, you know, yeah, with the exception maybe of like the frontier models.
- 1:10:42SPEAKER_00: So there'll be like the frontier model and then like the whole infrastructure stack from hosting your code and version control, which is GitHub, to, you know, hosting your app and databases, all that is basically gonna become very plug-and-play.
- 1:10:58SPEAKER_00: And um, like to your point, you know, part of this research, especially if you've never done this before, is and should be where should I host this?
- 1:11:07SPEAKER_00: What's the best solution for this kind of app?
- 1:11:09SPEAKER_00: But like the more you understand what it is you're building and what's required and whatnot, like the better the LLM is gonna be able to help you make good choices there if you've never done it before.
- 1:11:19SPEAKER_00: But it's just like you said, George, like you really kind of do have to go through this process a few times.
- 1:11:25SPEAKER_00: Like it would be weird if you got on a bike and were just like an amazing bike rider, right?
- 1:11:29SPEAKER_00: Like bicycle rider or like whatever, you know.
- 1:11:32SPEAKER_00: Like you, if you're into jujitsu, like if you just like started rolling and like immediately were a black belt, right?
- 1:11:37SPEAKER_00: Like, that's not how it works, you know?
- 1:11:39SPEAKER_00: So like yeah, I guess I know we're at time.
- 1:11:43SPEAKER_00: Um I don't know how how much longer.
- 1:11:46SPEAKER_00: Yeah, let's uh I'm already I'm late for another call.
- 1:11:50SPEAKER_02: Yeah, right.
- 1:11:51SPEAKER_02: Yeah, yeah, yeah.
- 1:11:52SPEAKER_02: We I don't wanna um I don't wanna keep you.
- 1:11:55SPEAKER_02: So I would say that like good, like you could go build something, go build something this weekend.
- 1:12:01SPEAKER_02: Uh I think we probably need to do a follow-up to this in some format.
- 1:12:05SPEAKER_02: Uh, but all of these all these connectors, all these things are eventually going to be vertically integrated, but it doesn't mean that you shouldn't go learn some of this stuff now.
- 1:12:14SPEAKER_02: Like you can deploy small tools just for yourself or just for your team.
- 1:12:19SPEAKER_02: Um, use use this like uh the it's the Atlasov three-step process.
- 1:12:25SPEAKER_02: And step one is get smarter.
- 1:12:27SPEAKER_02: Step two is what was it?
- 1:12:29SPEAKER_02: Compress.
- 1:12:30SPEAKER_02: Compress.
- 1:12:31SPEAKER_02: And then step three is is this kind of like build and verify loop.
- 1:12:36SPEAKER_02: Um, and you can be like you can go at any point, realize you go back to another step, go back to like I didn't get enough knowledge, or I got the wrong sort of knowledge, or anything like that.
- 1:12:46SPEAKER_02: Jack, where can people uh where can people follow you?
- 1:12:49SPEAKER_02: Where can people find your work?
- 1:12:51SPEAKER_00: Yeah, you can follow me on Twitter, Jack underscore Atlasov or X as it's called.
- 1:12:57SPEAKER_00: You can find me on LinkedIn, Jack Atlasov.
- 1:13:00SPEAKER_00: Um, those are probably the two best best possibilities.
- 1:13:02SPEAKER_02: I'm gonna link to your uh GitHub as well because I feel like you've got some uh I I I want you I wish you had more stuff on GitHub.
- 1:13:09SPEAKER_00: I'm waiting for you to just so can I tell you can I tell you what the actual amount of repositories I have on Twitter?
- 1:13:14SPEAKER_00: I I would love because I think your public is like 20 something.
- 1:13:18SPEAKER_00: Yeah, most of them are pretty outdated now, but um yeah, public in terms of what I maintain is like six, but I have a hundred and thirty-four the private repositories?
- 1:13:28SPEAKER_00: Most yeah, most of it is private.
- 1:13:30SPEAKER_00: Yeah.
- 1:13:31SPEAKER_00: Yeah, nice.
- 1:13:33SPEAKER_02: I've got like 20.
- 1:13:34SPEAKER_02: I've got like 20 like total.
- 1:13:36SPEAKER_00: So hey man, you know, it's it's yeah, I went a little crazy for a while with building.
- 1:13:42SPEAKER_00: So yeah, I know.
- 1:13:43SPEAKER_02: That's why I wanted to have you on, man.
- 1:13:45SPEAKER_02: Uh, thanks for talking with us today.
- 1:13:46SPEAKER_02: I feel like I feel like I want to do uh eight more of these and and just get into the nitty-gritty.
- 1:13:52SPEAKER_02: Uh let us know if you want us to build a CRM for food service.
- 1:13:56SPEAKER_02: We're gonna let Jack go about his day.
- 1:13:59SPEAKER_02: Thanks so much for joining us, Jack.
- 1:14:01SPEAKER_02: Thanks, man.
- 1:14:02SPEAKER_00: Thanks for having me, George.
- 1:14:03SPEAKER_00: Really appreciate it.
- 1:14:03SPEAKER_02: Glad we finally got this together.
- 1:14:04SPEAKER_02: Hope you have a great rest of your day.
- 1:14:06SPEAKER_02: Thank you guys for joining us today.
- 1:14:08SPEAKER_02: You can tune in for more gross to nut, like, subscribe, all of that trash.
- 1:14:12SPEAKER_02: You know, do it.
- 1:14:13SPEAKER_02: But but uh I love you and come back soon.
- 1:14:16SPEAKER_02: Bye-bye.