Latest / The AI Ecosystem Podcast / The MCP: Where Ecosystems Meet AI
Transcript
- Bob Moore: I'm your co-host Bob Moore here with Lindsay DeFalco, Chief Product Officer at Crossbeam. ⁓ Man, we had great feedback and conversation around the episode from last week. ⁓ I feel like more AI tools are listening than humans. And that, mean like so many people asked me for the transcript. like it's, it's a, it's this weird thing where, I even know ⁓ that consumption of all this stuff is actually going into human ears as much as it's going into, people that are distilling it down into AI summaries or we have an internal agent now that has like, quote unquote, listened to every episode of the podcast and kind of like, answers questions or references episodes. And that gets weaved into the rest of our. kind of internal sales collateral and product marketing stuff. I don't know. ⁓ I guess I used to say like only my mom listens to this stuff. ⁓ and now it's like only the robot that my mom dispatches. ⁓ I don't ⁓ know been your experience. Lindsay, is anybody out there, ⁓ out there listening? Lindsey DeFalco: Yeah, I was going to say it's not just your mom, it's also people that are employed by you. I think our usage spiked. Bob Moore: Yeah, that's right. Great. We're going leave it at 10 out of 10. Lindsey DeFalco: Viewership. ⁓ Yeah, no, I do think it's fascinating. I also got a couple of requests for the transcripts. And while I would shame people for not listening, ⁓ it's interesting because there is just a different way to consume the information now. ⁓ And of them. Bob Moore: yeah. ⁓ Like audio information at one second per second. It just feels like a broken model at this point. Yeah. Lindsey DeFalco: Yeah. Yeah. And it's like people learn differently. Some people, I was actually just wondering right before this, like who actually watches this versus who listens. And, you know, who is feeding it into ⁓ sort of AI tool and getting back a summary or I wonder if anyone's taking a transcript, putting it in notebook LM and then making it make a new podcast about it. Bob Moore: ⁓ No, don't do it, Lindsey, don't finish that sentence. Yeah, the reason it makes me sad is that it could be better. They take the trans here from this and yeah, it's like the 11 Labs voices talking for five minutes. It's like, ⁓ I got the, I got a feel for it. Lindsey DeFalco: And I will say, I don't know if I trust it, because I asked Claude to pull out some good jokes from the last episode. And it said that there were no jokes. And then it said the closest thing to humor was, and then it listed a few quotes. So I don't know. Maybe that means it actually has better taste. There you go. Bob Moore: Yeah, it could have said the whole thing is a joke, which might have, like, would not have been inaccurate. Interesting. Lindsey DeFalco: Yeah. I Yeah, that does like what is interesting about this whole concept though is that and kind of pull us into a topic that I think would be interesting is like you now ⁓ have this We talked about the ⁓ my God, no one's going to have jobs in tech in two years. And then kind of came back to reality in the next episode and we're like, hey, ⁓ we all have jobs and that feels really far away. there's this like middle world where we're in right now where ⁓ have these coworkers that are actually agents or ⁓ or AI tools. so ⁓ just the way that you think about ⁓ and storing all of the information that other people need to access is totally different. Like I wouldn't have spent time ⁓ to look up how to automate dumping transcripts from Riverside into ⁓ sort of repository that Claude can access when people are. ⁓ using it to ask questions two years ago. Like that just wouldn't have been something that was necessary. And now it's like having a repository of this content is just as important as people having access to it live or having the recordings. Bob Moore: Yeah, and the cognitive load of building those pipelines is like approaching zero. So like the math always for me on that kind of stuff is like, well, I'm a big fan of just like being a workaholic and just ⁓ like more. If it takes me two minutes to do something, like, you know, I've always been skeptical of like, let's over engineer some. Pipeline to automatically download the transcript and move it over to some Google Drive and like this is gonna break at some point and whatever else and like I'm totally flipping on that where Before I would have said look it's gonna take two minutes every episode And it's gonna take me 200 minutes to build this if we think this podcast even gets to a hundred episodes I think we're like like over optimistic. I'm really down on the podcast when I shouldn't be this is It's great Yeah, I am Lindsey DeFalco: Bummer. Bob Moore: Self-deprecating until things break through. So now fit, use or you use something to create ⁓ like that auto downloading or? Lindsey DeFalco: So speaking of my confidence in the podcast, was actually like, well, we're not there yet. So I'm just manually dumping them into a Google. Bob Moore: ⁓ you're doing it. You're doing the thing. Okay, good. So yeah, you're right. We're in this hybrid state, right? But I used, you know, I've been using Zapier a lot more now, ⁓ which I've always been a fan of and used for various like, kind of shortcuts and like personal efficiency optimizations. But they have this AI copilot now that's just chat interface that sits inside of their UI for building workflows. And I just didn't play language. I needed a web hook for, for like a Chrome extension thing that I built. and, normally would have been like, Ooh, how do I do this in Zapier? there's probably like 30 different like web hook input things and like, ⁓ I needed to do all this data mapping and routing. And I was just able to have like a classic chat GPT style chat interaction, tell it what I wanted to do. It understood it and it launched me into their UI. ⁓ and created it, created like end to end the exact Zap that I needed and made all these little micro choices. like where you used to need platform specific knowledge, the smartest platforms are actually obviating the need for a lot of the UIs that they've historically built, which have gotten ⁓ kind clunky, which I think was kind of true. Like there was some cognitive load to loading up Zapier and building a new Zap ⁓ once every three. and now it's just instant. So chances are that transcript that we could probably do that with a Zap pretty quickly actually. Lindsey DeFalco: Yeah, interesting about that, I did something similar the other day with creating like an analytics dashboard. ⁓ what was funny about it is that at the end of it, ⁓ like, normally if I had just spent an hour or two hours working on creating an analytics dashboard, I would have learned something. But at the end of watching Claude Cowork create an analytics dashboard, I learned nothing. ⁓ so I found that fascinating because I was like, man, Bob Moore: Yeah. Lindsey DeFalco: I think I need to change my approach for some of these things to teach me as we're going and not just have it do the whole thing because then any small update I want to make, I got to go fire up the whole thing again because I didn't learn how to do it. ⁓ And it's this interesting, like, does it matter? Do I need to know how to do it? Probably not because that's only going to get better. ⁓ But the way I'm thinking, I'm still like, wow, this feels weird that I created this thing and I didn't ⁓ learn anything. And I also like, I don't know the inner workings of it in a way that I'm super comfortable with. And so I think that's like, ⁓ that's just where AI still is right now and it's not going to be there in a year. ⁓ it messes up enough now that I still ⁓ that way. Bob Moore: Yeah, know, a few episodes ago, I called your analytics dashboard slop. And after this description, I stand by. Yeah. ⁓ No, I was, I was on a call yesterday with some, like founders in the, one of our venture firms portfolios. ⁓ we had this exact discussion. was just like the, cost of maintenance versus the cost of just like net new generation. And it. Lindsey DeFalco: Yes, I haven't forgotten it. Bob Moore: It creates an extremely ephemeral amount of output. at the margins, ⁓ know, it's kind of costless to manufacture this stuff that ultimately is kind of disposed or disposable. And like, what's the line that you cross to actually build something into a repeatable, like multi-user platform versus just like, the point I made on that call was it actually writing a SQL into ⁓ a business intelligence platform that's, you know, capable of being maintainable and distributed across your company and has distribution and kind of all these things that, that make it persist is way more cognitive load and time consuming than just like going into Claude and being like, go figure this out. I want to know XYZ thing. and yeah, we probably end up in the Zapier world in the middle somewhere where like you've got these systems of record that have distribution. And then you have the AI just making it wildly easy to generate outputs in them. But maybe not. I don't know. Maybe it just all becomes disposable. Lindsey DeFalco: Yeah, there's this interesting trend I've been saying ⁓ where, and I can't, I don't know how much of it is LinkedIn influencers. It's, hard to tell. ⁓ But I, I have found a similar, had a similar experience in ⁓ for myself in cross main, which is like, there's this theme right now with marketing teams where they are starting to operate in a or ⁓ And you know, Bob Moore: ⁓ yeah. Lindsey DeFalco: they're starting to think about their content as a repository. And what's interesting about that is like, know, if you've, if you've done engineering, then you know all Git and you understand there's like a repository of content, there's branches, you have to get approval to merge that branch back into the main content or main code, whatever it is. ⁓ And really it's it's a concept that humans emulate when they work in teams. It's just not clear. But like if we're working on a project together, right? Like we're either working in a shared drive or we're like passing a file back and forth, right? And that's like a similar concept. And so we're in this world now where it's like, how you set up the infrastructure so that agents and humans ⁓ together? ⁓ that brings us back to this ⁓ like ⁓ repository framework. ⁓ I think that trend is really interesting. Bob Moore: ⁓ It reminds me actually, like, I feel like this insight is something that most certainly predates LLM era, but maybe is, I don't know, more ⁓ important ever now, which is like back in the, ⁓ let's get our daily RJ Metrics reference, back in the RJ Metrics days ⁓ when the original team that was creating DBT as an open source project, kind of was like putting their vision together and Tristan and Connor and Drew were kind of like getting the foundational rails together for what ⁓ was. ⁓ was this core belief that they had that analytics ⁓ basically ⁓ engineering and that the practice of defining, for example, like, you know, what ⁓ is a true customer in your customer's table, right? Like what are the conditions under which the row is actually valid or like you have these canonical business definitions that very often you need to contort the SQL in order to get out the other end that those things ought to be committed to code. And then you could actually do that ⁓ SQL code in a lot of cases. And that's like dbt, the data build tool ended up becoming that as an open source project. And it made it basically created this giant transformation. whereby analytics teams that used to operate in like, you know, either a built-in proprietary modeling layer inside of their BI tool, like LookML and Looker, or just like kind of wing it a bunch of copy pasted SQL that lived all over the place, could finally operate like engineering teams could. And the core innovation of dbt is that it lives in Git. And ⁓ your dbt model to live inside of a Git repo and your analytics team or analytics engineers, are basically making git commits and like for forking off of the main thing and you can have like staging environments and it's like this really beautiful thing that works so damn well that dbt became like this global standard for I kind of think there's an analytics team of note on the planet that doesn't have dbt and it's ⁓ in its stack somewhere so ⁓ why not marketing teams why not whatever and the honest answer is probably that same like cognitive load technical overhead which is like Git was invented by the same guy who invented Linux and is a big computer nerd and it's a bunch of like command line syntax stuff that is a little bit of like a head trip to figure out if you're not already highly technical. But with an LLM as middleware, I've written more code in the last six months than I did in the first 10 years of my career. I don't think I've written a single Git command directly. Lindsey DeFalco: Yep. Bob Moore: I've probably issued more Git commands vis-a-vis Claude Coe than I ever have, but like sitting at the command line, you know, doing pulls and pushes and merges and that stuff, it's just like, ⁓ couldn't rebase a branch ⁓ you paid me at this point. Like it's falling out of my head and it doesn't matter. So maybe it never needs to be in your head in the same, in first place. Lindsey DeFalco: Yeah. Yeah. I recently spun up a cloud code and I have so much baggage from my coding days about, you know, getting stuck with some branch that was out of sync or some like terrible merge conflict. And yeah, it just handles all of it for you now. You don't even need to need to think about it. ⁓ yeah, I think it's interesting. I think it's like ⁓ come up a lot as a, you know, a concept also, I think to approve things that agents are doing. ⁓ ⁓ because like merge requests in themselves are this like, Hey, approve this, before it goes into the main version of this thing. and I think that, you know, in itself ⁓ just, helps the agentic side of it. ⁓ And ⁓ I actually ran into this at work cause I created a skill and it was an ask product skill. and I was like, Hey, how do I get this ask product skill? published for the whole company to you is not just a private skill. And I literally sent the raw file over Slack. And I was like, why isn't this in a repo? And why can't I just ask a question? Because then I updated it, and I had to resend a new file. And then our ops person wrote back and was like, hey, can I combine this with our Ask Services skill? And I was like, yeah, sure. Go for it. But I want to maintain this, this, and this. And it's all just version control. And I think we're going to see more of those, the technical concepts of like, version control, know, branching approvals, merging, work its way into a lot of workflows because we're in this blended world of like agents do things and you're working on the same stuff and you want to approve it for now before it goes into the main version, but still have like visibility and control over who did what. Bob Moore: Yeah, it's funny. the same call I was on yesterday, of the people demoing showed that they are doing exactly what you're describing, which is they have all their their cloud skills in ⁓ GitHub repo. But like the behind that is like, what happens when someone who's not one of the engineers at your company wants to work on a skill? Because like not everybody has a GitHub seat, A, and then B, not everybody who knows how to use the skill builder skill necessarily knows to interact with Git. And maybe, you know, for the same reasons I stated earlier, they don't really need to, but I don't even know that they know that they need to know to like there. So it almost creates these, if you want to do version control, you're kind of locked into a user base that ⁓ is accustomed doing version control. which is limited to engineers, which like cuts off, you know, the majority of your org chart from being able to contribute. It's like I too, it's kind of a thing. Like I look at our skit, the skills we have in our Claude org internally at Crossbeam and it reminds me of Slack channels back in the day. It's like a bunch of stuff separated by dashes with no organizational nomenclature rules. I ⁓ my like type A brain wants to go in there and rename them all and organize them and give them prefixes by department and like version numbers and all kinds of other stuff. But it's kind of silly, A, you don't need to do any of that because the LLM knows what all the skills do and you don't even need to name the skill by name. The fact that it has a name is kind of like a weird artifact. And then additionally, it just feels inevitable. If Anthro-Aby doesn't do this directly, and they probably should, then there will most certainly be a startup that probably is one that exists already that is just for management and maintenance of skills with as a backbone, that every company that has their own skill library should be, should be using. And then ⁓ all you then need is just like a pipeline from that Git repo into your actual like Claude org-wide skill repo. I know you're smiling. Does that exist? Did you start this company last weekend? Lindsey DeFalco: No, no. I ⁓ thinking about how we were ⁓ about before of like, have no idea what ⁓ and companies are going to exist. And we also talked a lot about like bundling and unbundling, which the same thing happens when you ⁓ abstract cases and then you reapply use cases. ⁓ And it's funny because you're sitting here talking about that. And it's like, we went to we went to these LLMs where you can ask anything and It's like wide open and like, that's the appeal. It's like, whatever you want to do. Literally someone's like, can AI do blank? The answer is like, yes. You just have to figure out, ask it how to do it basically. ⁓ And then we get into this world. And I think I mentioned this last time where we're talking about skills and skills are, yeah, they're implying like they're applying instructions, but some of them are full use cases, right? And then it's like, you're talking about a product to manage skills. And it's like, at what point are we going to get skills grouped up into products that you install as like connectors you know, I don't know some other word that doesn't exist already for this. And then those are like your specific or persona specific products. And it's so funny to watch this, like, ⁓ maybe we don't, but like you said, it's, it's like untenable that you can have 50. We already have like 20 skills at the company level. and we've been on this for what, like a week. ⁓ so. Bob Moore: Yeah, oh, it's gonna get ugly. And I think like the idea of just give me a dev environment, give me a staging environment, give me like a way to promote these skills out to production and roll them out more widely and give me different user groups inside of our org that we can have certain skills be accessible to. It's all coming, but it's like ordinarily. Lindsey DeFalco: Mm-hmm. Bob Moore: years ago if ⁓ an idea or a need like this surfaced I would have been like let me go find who's working on this I'll try to like it's not our thing across beam But this is like an angel investment I want to make or like there's some like exposure to this space because I think it's so real and right now I'm standing here like should I just go build this? afternoon like instead of playing Nintendo switch tonight should I like ⁓ build this entire company in four hours? ⁓ I'm kind of demotivated by that because I know that if I don't somebody else will And like the speed with which things are moving, these innovations that I feel like need to exist are just spontaneously appearing. And in a lot of cases, it's faster than I can even realize the need. And then by the time I realize the need, it's just kind of there. And I do think that there's a lot of that going on. So like when I have a novel idea, speed to market, coming up with an idea, ⁓ that novel, speed to market, not that differentiated. Lindsey DeFalco: Mm-hmm. Bob Moore: distribution kind of funky because, you know, are people even the consumers anymore or like, you know, are, are people only discovering new stuff through asking questions of LLMs and then there's this giant black box of how to get the LLMs to know you exist and care about you. it's very weird. It's a very weird time for net new companies and net new entrepreneurs right now that aren't super, super deep in the stack. Lindsey DeFalco: Yeah. Yeah, actually, you just reminded me because, know, ⁓ GPT notoriously was getting roasted by Anthropic about the fact that they're rolling out ads. And I was doing this like research on buying something on chat GPT and it kept kind of like poo pooing the thing that I wanted and suggesting a different type and ⁓ like a different brand. And I finally stopped and I was like, hey, what's your problem? ⁓ Is this do really not think this one is ⁓ is good or is this an ad? Or is this some sort of like the I don't know is this getting surfaced higher because of like a EO And it said no, don't know if I believe it but it was like no This is just this is just what I think and I was like, yeah. Okay. I don't know if I buy it Bob Moore: Yeah, who knows what's in system prompt. ⁓ Yeah. I'm glad the ads aren't there yet. They do feel inevitable. And I do feel like I would immediately start buying those ads, which is also like kind of ⁓ wild. ⁓ I saw chart the other day. It kind of blew my mind. I didn't think it was this extreme, but it was a market share by percentage ⁓ of ⁓ Lindsey DeFalco: Mm-hmm. Yeah. Bob Moore: the LLMs in the consumer market. And chat GPT was like 95%. Lindsey DeFalco: Yes. Yeah. Bob Moore: which like in my world, kind of amazing still to realize just like how big that market is and how much consumer adoption for chat GPT is there and how dominant. mean, it's like, it's on the same order of magnitude as like Google's share of the search market. Yeah. Lindsey DeFalco: think they did the absolute best job of basically became like synonymous with AI chat, like chat GPT is like, you know, universal phrase. I even internally, was interesting because when we were talking about like what tool we were going to use, ⁓ the general sentiment like, ⁓ Claude. And then if it's like word association, it's like code, Claude code. ⁓ like they did. Bob Moore: ⁓ yeah, yeah. Lindsey DeFalco: Like Anthropic did such a good job in the code space ⁓ that, and you know, now in the enterprise space as well, which I know OpenAI has also made very strong strides in. But yeah, I think they just like did different things really well and that's ⁓ sticking. But I don't know that if I asked anyone in my family who does not work in tech, if they knew what Claude was that they would actually know what it is. Bob Moore: ⁓ yeah. Yeah, you're totally right. like my, ⁓ my dad's on, ⁓ chat, GPT all day. Maybe a little gem and I mix in there. This is funny. This, who knows what AI generated, ⁓ content farm this actually is, but, this chart here, I'm like, ⁓ yeah, look at that. That chat, GPT is like way up there. And then you realize that, that the w they made this Y axis, like quasi logarithmic. Like what even is this scale? Like zero to five is the same distance as zero to 10. And then, ⁓ like, is that logarithmic? I don't know, but they did whatever they could to make it so that these lines were actually visible. It's funny, if this was true zero to 100, every single one of these lines would be like at this spot or lower. So they're doing everything they can to tell a story there, but it's pretty manipulative Y-axis. Lindsey DeFalco: Yeah, yeah. Yeah.