Liam Martin v2: Distributed Intelligence, AI-Native Organizations | Work 20XX Ep58

**COLD OPEN**

*[in three, two, one.]*

---

**JEFF FRICK (Host):**
Hey, welcome back everybody. Jeff Frick here with Work 20XX, back at Running Remote in Austin. We're here for our second year. It's been a great collaboration with Liam and the team. And so we're excited to be back and Liam's just coming off a crazy, a crazy panel. So welcoming in officially Liam Martin, the co-founder of Time Doctor and the co-organizer of Running Remote. Liam, you are way out on the bleeding edge of this AI journey. We just had a really interesting panel. And I'm going to lead with your first question, which is: What the [____] is happening? What WTF is happening here, and do we know, and how do we find out, and where should we go for information?

**LIAM MARTIN (Guest):**
Honestly, I was hoping that you'd know, because you interview a lot of people in this space and you probably have a better idea than I do. But I think that we're in the transition year. I think this is bigger than 2020. I think this is bigger than cloud. I think this is bigger than mobile. This is the year where we really switch from AI being a fun little tool that you use to AI being part of your workflow, part of the way that you do work, and unfortunately replacing a certain part of your organization.

**JEFF FRICK (Host):**
I want to talk about kind of a fundamental piece of it, which is how do you try to stay up? I mean we can't stay up. There's new announcements every day. There's new companies rolling out. New tools every day. There's new versions of the tools that we know every day. And as we know, the tools will only get better tomorrow than where they are yesterday. In terms of your strategy for having an AI — I tell people you got to have an AI buddy, or a couple, in terms of people that you can lament with, share with, get ideas with — how do you kind of, what's your kind of ecosystem of people that you have?

**LIAM MARTIN (Guest):**
So I have a lot of private WhatsApp groups where we can really speak openly about those types of things. A lot of founder groups connected on WhatsApp, which is great. But outside of that, I try to connect with people in person. So I just ran a small event — it was a large house that had about 22 bedrooms — and I invited a whole bunch of my friends that understood this better than I did to that venue, and we hung out for two days and we talked about what the hell is going on with AI. And I have to tell you, some of the stuff that they were doing made me more anxious than when I walked in. Because I'll have to tell you, right now we're seeing software built faster than ever before. We're seeing dark code inside of applications — so applications that are built in part or completely just by AI agents that we don't really understand how they work, because no human has actually reviewed that code base. And then we're integrating that into incredibly important systems — that we were just talking about before — OpenCLAW, right. And yeah, an OpenCLAW can get access to your credit card because it has access to your computer. It can just look at your file system and figure out where it is. More importantly than that, it can get access to your Social Security number. It can get access to your Pornhub history. It can get access to everything that you know that makes you. So where does that leave us? Do you really want an AI agent that acts independently of itself, knowing all of your deepest, darkest secrets? I don't think I would.

**JEFF FRICK (Host):**
No, but I have to tell you, the thing that has surprised me the most here for just a couple of days is the number of people running OpenCLAW. I mean, I have an episode that I haven't dropped that I probably recorded with a bunch of experts when the whole OpenCLAW thing went down — I mean, it was like over a weekend. I think it changed names three times in a day. It was like the largest, the fastest download of history, in terms of — but it was scary as all get out. And yet I bet half the people I've talked to here are running some version of their own OpenCLAW, which is fascinating.

**LIAM MARTIN (Guest):**
Yeah, it's wild. It blows me away. And the other thing that is blowing me away is when you look at the classic SaaS model, right, and every time that Claude is coming out with a new update or OpenAI is coming out with a new update — I don't know if you saw when Claude came out with its Mythos application, this new build of Claude, and when they announced it, when it got leaked, all of the security firms, SaaS security firms, dropped by about 25% on the stock market. We are literally dismantling software piece by piece — this thing that took us two decades to be able to build. In terms of SaaS, I feel like either companies are building these applications for much cheaper, or in a much more destructive way — kind of leading to this — they are building open source GitHub repos and just giving it away for free. Higgsfield.ai, which is a $1 billion valuation company, just this week a guy released all of his version of Higgsfield. So you can basically plug in your own API key and you can run that application right off your own Claude Code deployment.

**JEFF FRICK (Host):**
It's funny as you're saying that, right? Because one of the big trends that's happening and has been for a while is hyper personalization on the consumer side and in our media consumption habits and whatever. And as you're talking about this code — and you think of back in the old days, right, MVPs — and you think of the old shrink-wrap software days, right? You had to spec that thing out and it had to last for a year. Now, suddenly — I mean what is a SaaS application? I mean, what is software? It's a database with some logic and a UI. So now I'm just thinking, as you're talking, I can hyper, I can hyper-manufacture my own exact version of that CRM app, or whatever I choose to call it, for the function and the task that I need done now.

**LIAM MARTIN (Guest):**
Absolutely.

**JEFF FRICK (Host):**
Now it's almost like I can create my own SaaS application and maybe throw it away tomorrow because I don't need it tomorrow.

**LIAM MARTIN (Guest):**
I think that we're probably entering the age of hyper personalization for applications. So I believe that you either own a piece of the back end — so you own a database, you own where that information is stored — or you sell tokens. So you're either in one of those two camps. User interfaces are going to change from static to dynamic. So right now you look at a Salesforce. Everyone kind of has the same Salesforce user interface. And if you don't, it costs you hundreds of thousands of dollars to get some consultant to come in to actually change that user interface. All of those days are done, because you're going to have an MCP server that's going to pull in all of that data — that is essentially what a CRM does — and then create a dynamic user interface designed just for you and you as an individual.

**JEFF FRICK (Host):**
Right, it's wild. So another kind of interesting thing — it came up in your panel — is that if anyone can basically say write me a marketing plan or write me a social media campaign and I just hit boom, where's the value, right? It's kind of your classic supply and demand situation. If everybody can do it and everybody can do it easily, where's the value? It's no longer in it. And in fact — I mean you know, it's a funny thing that happens — how much stuff do you generate that you never even read, because it's just so easy to generate? That you never get around to consuming this stuff.

**LIAM MARTIN (Guest):**
Yeah, it's one of the interesting things from an internal documentation perspective. Those documents have gotten bigger and bigger and bigger over the last three years. And it's not because people are writing longer documents, it's because AI is doing the writing for it. And what I now do, to be completely honest with you, if someone gives me a large report, I just throw it into my AI and then say, what do I need to know? And it tells me, right. So it's wild. We're getting AIs — we're getting a human to go to an AI to build these ridiculously long reports, that then I take and put into my AI to decode back to the actual two paragraphs that I needed to know about in the first place. I'd rather just go back to the two paragraphs. To be completely honest with you, it's much more efficient.

**JEFF FRICK (Host):**
Well, then begs the question, right? And then what's the decision that you're going to make from that? And would it have been more efficient just to take the human out of the loop in that process that you just described, because you're talking about a data set that somebody created something in, that gave it to you, that you then put in to get a summary. Kind of sounds like maybe there's a direct connection there that you could cut the middleman out.

**LIAM MARTIN (Guest):**
Yeah, yeah, yeah. I was having a conversation on our panel about us building an MCP server. An MCP server is an AI-enabled API that has specific tool calls for large language models. And so one of the product managers was talking to me about — hey, you know, we're thinking about putting out this MCP server. And I said, okay, you know, what are you talking about? Oh, I've spoken to seven people about it. And here's the comments about it. And we're thinking about rolling it out to the marketing team.

And I'm asking him questions on my Zoom call, right. But while that's happening, I am building the MCP server inside of Claude Code. And at the end of that 15-minute call I said, well I built one, so you don't have to worry about the second half. I've basically built your entire second half product roadmap. And he just looked at me with this shocked look and I said, if this is the way that you think that we're supposed to operate as an AI-native organization, you're very, very wrong. You need to show, never tell.

**JEFF FRICK (Host):**
Right, right. And the fact that you could do that so fast is just crazy. We were at Google Cloud last week, Google Cloud Next, and one of the phrases I thought was really powerful — objective-based computing versus instruction-based computing. Right. I tell it what I want. I don't know what skills it's going to take to get there. And I don't really care. I just want you to do the best you can to help me figure out this problem that I'm trying to solve right now. That's a different way to think of the world.

**LIAM MARTIN (Guest):**
Yeah, you hit the nail on the head. I mean, I think that every single iteration of these new LLMs that come out — our orb of autonomy is going to get smaller and smaller and smaller, and I don't really know where that ends. I hope that it does end at some point where there is a human that needs to be in that loop. But to be honest with you, I don't know whether or not that's actually going to be true.

**JEFF FRICK (Host):**
It is funny, because everyone talks about human in the loop, and I'm a pretty strong proponent that today — maybe yesterday, maybe not today. Yesterday's stuff was maybe 80% right. When it came out, you needed to take a look at it. But you know, another thing that came out at this Google show, right, is that a probabilistic model with enough data approaches deterministic accuracy. Or you put a little deterministic piece in at the end. And so if you suddenly have these giant MCP-based multi-agentic with all these APIs and all these different applications making decisions — where is the human even gonna — I mean, where do you even insert the human in the loop? And they're going to be operating at machine speeds. So I don't think that's going to be possible forever across all use cases.

**LIAM MARTIN (Guest):**
So one of the interesting perspectives on that is if you look at the way that token usage — so you're using Claude, if you're using Codex — you get about 100,000 tokens a day, right. In terms of if you're on the max plan. What percentage of that token usage do you think is just connected to the human user interface — literally created — the token usage is generated, is used up, to simply provide information back to the human being — it's 67%. So 67% of the token usage is not actually doing the thing, it is preparing the thing for the human to then make another decision.

**JEFF FRICK (Host):**
to consume

**LIAM MARTIN (Guest):**
How fast do you think it's going to be before you could just say, well, what if we just remove the human from the loop, because the AI gets it right 95% of the time, at 1,000 times faster than a human could decide. So, and the 5% that it gets it wrong, well, we'll just rerun it again and we'll fix it. And that's — I mean, at least that's the way that I'm thinking about it. And it's a very depressing thought. And I hope that we're able to figure out a place where humans can be in that loop and can't be dislodged — not just for me being able to retain my job, but I think for the safety of humanity as well.

**JEFF FRICK (Host):**
Yeah. Well, let's talk about the people that are managing humans today. In these kind of crazy times. And you had an interesting quote I pulled up someplace about burnout warning, and you made a really interesting observation, you said if their hours are up and their token usage is down, that should be a big red flag.

**LIAM MARTIN (Guest):**
Oh, absolutely. I mean, so token maxing is one of those things that a lot of people talk about. But there is — that is a little bit of a splitting-hairs situation, so there's a much more efficient way to be able to use your tokens. And I don't want people just burning out of their tokens every single day because those employees are literally using up compute. And if they're using it up improperly, that's wrong. But if I've given you these types of resources and you're not maximizing that every single week as an example, then it's something that I have to really take a hard look at to be able to say, well, maybe you don't use your Codex or your Claude Code subscription, and if you're not using that, then maybe you don't work in the company anymore.

**JEFF FRICK (Host):**
Yeah. So you're out on the cutting edge and I love it. And that's why you're so nervous all the time, which is awesome, because it shows you pay attention. Because the one thing I know is that the more you know about this stuff, the more you know that you don't know anything, the further behind you feel. You created an internal AI lab, Chainsaw —

**LIAM MARTIN (Guest):**
Yes.

**JEFF FRICK (Host):**
Tell me a little bit about — you know, why did you do that? Why did you create a formal thing, and what are some of the things that you're doing with it.

**LIAM MARTIN (Guest):**
So we recognized if we wanted to adapt into an AI-native organization, we couldn't do it all at one point. And probably a lot of people that are listening right now or are watching — they've got a 100-seat org, a 200-seat org, a 500-seat org. It's really difficult to become an AI-native organization that has three employees that can just do things so much faster than a 200-seat organization. You've probably been in companies like this before, where it's like — hey, I want to move a pixel on the website. You got to talk to eight people about that and maybe in the second half we can do it. So Chainsaw doesn't have any of those rules. It's an entrepreneurial organization inside of the larger organization. And so we have a mindset there, which is: Show. Never tell.

In the amount of time that it takes us to discuss whether or not we should do something, it's a lot easier for one of those guys to take a Claude Code deployment, who's an engineer, build a feature, present it to the group, and then just say, hey, should we test this with customers? And that's allowing us to be able to get that advantage, going, going faster without all the bureaucratic slowdown that you have in a large organization.

**JEFF FRICK (Host):**
So I don't know if you just answered the question, but how would you define an AI-native organization? What would you see, either aspirationally or kind of the core tenets? Because before, you know, with remote-first organizations, a lot of communication — async, documentation. If I had to pick three, right off the top of my head, for an AI-native organization, how do you see that? How would you define that? What do you hope to be?

**LIAM MARTIN (Guest):**
So one of the great advantages of remote organizations is because they've already created the documentation, they're already ready to become AI-native organizations. Because one of the biggest tenets is documentation. If we're all keeping this information undocumented — so if I'm talking at the water cooler with you about the business, well my AI can't actually absorb that information. So documentation is absolutely critical. But then once you take that, you actually have to open that information up. So it's not just humans, it's humans working with agents. And a lot of the times — I think I joked with you before — my opening talk here was a quick ten-minute talk and I had a whole bunch of Google Slides about two weeks beforehand, before Running Remote. Claude Design came out, the brand-new feature from Claude, and at this point it's a brand-new feature. So I took my slides, I uploaded it to Claude Design, and I said, what do you think about these? Do you think I should change it or do you think I can make it any better? And it said, well, the core premise that you're presenting here is absolutely wrong. What you talked about two months ago in this call, which was, we're moving from distributed people to distributed intelligence. Is such a better angle. And I agreed with it. And then it put together the best single slide deck I have ever seen. The AV department that has worked on hundreds and hundreds of shows, he said, this is the best slide deck I've ever seen in my entire life. So now I know I will never create a slide deck ever again without the use of Claude Design or Claude Code or some other type of LLM tool.

**JEFF FRICK (Host):**
Right.

**LIAM MARTIN (Guest):**
Because it's AI with humans, and it pulled out those insights — that insight that I had — which is we're really moving from distributed people to distributed intelligence. And it built, it pulled that insight out as that seed and built a presentation around it.

**JEFF FRICK (Host):**
So one of the themes last year was — and I think Brian Elliott was the first one that kind of said it in so many words — was that the same management characteristics, behaviors, attitudes that are successful in driving remote work happen to be the same behaviors, attitudes, and point of view for driving AI adoption. And right now, the number one management dictate is we need to get our AI adoption up within the team.

**LIAM MARTIN (Guest):**
Absolutely.

**JEFF FRICK (Host):**
So it's pretty interesting that it's very similar. It's risk-taking. It's curiosity, it's pushing boundaries. It's being — you know — documentation clearly a big piece of it. But it's really interesting how that has dovetailed into many of the same attributes from a leadership and organizational point of view.

**LIAM MARTIN (Guest):**
Yeah, I mean I think that, as I said, the vast majority of remote companies — if you have like a Venn diagram of remote companies, I would probably say 90% of remote companies can become AI-native organizations, but a very small percentage of in-office companies can really adopt an AI-native mindset because they just don't have the documentation in place. And also additionally, when you think about AI agents — well, AI agents are remote by definition. They're in the cloud, right? So they're working with you. They're not in the office. You can't talk to, you know, OpenCLAW about how your day was at the water cooler. It's much better to be able to do that when everyone's working remotely, at least in my opinion.

**JEFF FRICK (Host):**
Yeah. So tell me a little bit about your stack. I know you have — it's probably too long to list — but what are your top five tools that you're using every day?

**LIAM MARTIN (Guest):**
So as of two weeks ago — and I might actually change that — I have Claude 4.7 as my architect. So that's usually the thing that I log into at the very beginning of the day. And then if I want to do a large amount of coding, I'll use Codex for that. So we actually have a CLI integration between those two tools to create a bridge, because I can't — Codex, when you give it really, really good specifications, it will produce amazing results for you. But it's not as iterative as something like Claude. Claude is just much more human in the way that it understands you so much better than Codex. So I speak to Claude about what I want to build, and then I send that to Codex, and then I have a bunch of other kind of tools that connect to this. But fundamentally for me — as I said something else to a couple of other people today that kind of scared them, which was — I have not logged in to a browser in at least three months.

**JEFF FRICK (Host):**
Really.

**LIAM MARTIN (Guest):**
So everything that I'm doing is inside — inside of an LLM. If it doesn't have an MCP server or a CLI integration, I don't use it, because it doesn't provide me the amount of value and the speed that I need to go. A real great shout-out to a fantastic analytics product that I recently just started using called PostHog, and PostHog does app analytics. It's a pretty popular application. Their website is now just a single MCP terminal install code line. You grab it, you install it inside of Claude or Codex or whatever you're using, and now you're up and running with PostHog, because they recognized it's not about the user interface. The user interface is no longer important. It's about the data that you have and how you can present that to an AI.

**JEFF FRICK (Host):**
Right. I think that's been the biggest change for me probably over the last two years, because I remember when these things were coming out, I'm like, well how do I get my data in, you know, how do I apply this power to my data? And, you know, but sure enough, suddenly they have memories. Suddenly I don't need to learn all these silly tools. I can just point it to my Google Docs or upload my PDFs or do whatever. And that's even before talking about dynamic connections, which is very much more powerful.

**LIAM MARTIN (Guest):**
Right.

**JEFF FRICK (Host):**
I want to shift gears from a leadership point of view. One of the other unlocks here — when I was listening, Wade was talking about their executive council — is getting negative information, getting feedback, like the thing just told you your premise in that deck is absolutely wrong. It seems like a huge unlock for a manager and a leader to actually have somebody who's willing to look at the data and say, Liam, no, or maybe you should think about this — because getting negative feedback from employees worried about their job position has historically been a really tough thing to get to the top. And by the time you hear it —

**LIAM MARTIN (Guest):**
Sure.

**JEFF FRICK (Host):**
it's probably too late.

**LIAM MARTIN (Guest):**
One of the things that I've really done is — LLMs are historically really, really bad at providing you negative feedback and constructive criticism. Oh yeah, I'm thinking about building a cat Twitter for cats. It's called Catter. What do you think? That's a great idea, Liam. You should raise $20 million to be able to do that. I mean, that's ChatGPT fundamentally.

**JEFF FRICK (Host):**
Right.

**LIAM MARTIN (Guest):**
It was always this happy little puppy that always agreed with everything that you were doing. I have built a large system prompt inside of my soul.md file inside of my Claude deployment.

**JEFF FRICK (Host):**
Your soul file.

**LIAM MARTIN (Guest):**
My — yeah, so the way that these memory architectures work is there is a core identity of how I want to be able to react with my AI. And it's called a soul.md file. And in that file it says cut the bullshit. Like, be the devil's advocate as much as humanly possible. Because for me, I don't want people to be able to beat around the bush when I interact with them in person. And I definitely don't want my AI to be able to beat around the bush and basically tell me that I'm absolutely amazing and I should raise $20 million on Catter.com. It's complete BS. And I try to get that same thing into every single one of my team members' AI agents. Because an AI agent will agree with you the vast majority of the time. I was — one little personal story which has been wild, and it's connected to something else.

I have purchased a chalet up north in Canada. I'm Canadian, but it's way more up north, because honestly, I'm terrified of AI. It's got its own water, it's got its own electricity. I didn't think I was going to be a prepper six months ago. I remember when we were hanging out in San Francisco. My mindset has completely changed because I'm not scared of an AI doing something bad. I'm really scared of a bad actor using AI to do something really, really bad.

And so when we were purchasing this property, one of the sellers was very upset about a particular term and they wanted to talk to us about it, but they had said, well, we've engaged an attorney, and they gave me this big, long letter which was full of m-dashes and was very clearly written by ChatGPT. And I unfortunately got real attorneys, because if someone sends me a demand letter like that I have to respond with my real attorneys, which cost $2,000. And I remember my attorney said, like, Liam, why are you even bringing this to me? This is absolute bullshit. Just sign the deal, or don't sign the deal. And I said, well, you know, can you help me with this? And the response that he gave me was great. I responded back to this person saying, well listen, if you want to go the legal route, that's fine. This is my attorney, Patrick. I CC'd, he's currently going to be taking over, and your house is going to be held up in escrow for the next two years while we work this out. Oh no, no, our attorneys have told us that it's okay. Let's actually just sign the deal, right.

So a lot of people are having this mindset of, well, ChatGPT agrees with me and therefore I'm right. And the problem is all of these AIs are designed to make you happy. They're not designed to actually make you succeed. They're there to make you happy. I honestly think they're selling you candy instead of vegetables. And you gotta have more vegetables inside of your AI.

**JEFF FRICK (Host):**
Unfortunately, we live in the age of sycophants —

**LIAM MARTIN (Guest):**
I'm afraid so.

**JEFF FRICK (Host):**
we're not getting the greatest guidance from above. But let's finish on something a little more positive. And that's Running Remote and this community. It is a crazy passionate group of people. And like I said, it just so happens that all those same skills and behaviors that made them successful running distributed organizations are going to be the same things that help them get their organizations successful in grabbing onto this crazy new AI thing. So pretty serendipitous place to be right now.

**LIAM MARTIN (Guest):**
Yeah, I mean we've been doing it for — this is our 10th event that we've actually run, which is absolutely wild.

**JEFF FRICK (Host):**
Congratulations.

**LIAM MARTIN (Guest):**
We were running it for almost nine years, which is great. A couple of years during Covid we had more than one virtual event, so that's where we're counting that. And you're right. I mean, this community is so passionate about remote work and it's such a fantastic experience to be able to interact with them every year.

**JEFF FRICK (Host):**
Yeah. It'd be really interesting to do a compare of the top-level themes of those, how they have evolved over the last ten years, and especially the rate of acceleration and change over the last couple, versus, say, the first five or six or seven — which are probably all documentation, documentation, async work, documentation.

**LIAM MARTIN (Guest):**
Yeah, it was so 2018 and 2019. We thought we were this cool little group that had this neat secret called remote work. And then 2020 the entire world wanted to learn about remote work. So that community completely changed. And now we're kind of back inside. We're back to the middle, right, where it's like there are some people that we brought along for the ride from Covid that are really excited to be able to be here. There are a lot of people that went back to the office and that's fine too. They'll come back to remote eventually.

**JEFF FRICK (Host):**
Yeah. Well, Liam, thank you. First off, congratulations. Ten years, that's pretty — that's impressive. That's legit. And thank you for having us out. It's really a fun community to interact with, getting a lot of great insightful things. And like I said so many times, all this stuff is relative. Regardless of where people plug in their laptops, it really doesn't matter. And I guess nobody probably even plugs laptops in anymore, right, everybody's on their phone.

**LIAM MARTIN (Guest):**
Well, thanks for having me. And thanks for coming.

**JEFF FRICK (Host):**
Absolutely. Alright, he's Liam, I'm Jeff. Work 20XX is at Running Remote in Austin. We're happy to be here. Thank you for being here. Thanks for watching. Thanks for listening. Catch you next time. Take care. Bye-bye.

---

**COLD CLOSE**

*[Ooo woo — applause — cool.]*

---

Full transcript.