Why HR not CTOs Will Lead AI Augmentation - with Joshua Wöhle

I think the entire role of HR is about to become absolutely critical and going from a point where it used to be administrative. Actually it's not too dissimilar from what has happened to the cto. If you think about the CTO used to be the IT department, right before technology became critical to the company. You had an IT department and a head of it. And then companies were built where their entire value proposition was around the ability to leverage the technology. And so as CTO became a normal role and I think the same is going to happen from a people or needs to happen from a people perspective. Hi, I'm Joshua Wohler. I am the co founder and CEO at Mindstone. We are an AI adoption platform helping some of the biggest organizations in the world adopt AI. Specifically training their non technical people on how this technology can really, really make a difference in their day to day. And I am excited to talk a little bit about how I think HR is the function with the highest leverage in the organization to really get the benefits of all this technology today. And it's not the cto.

I've been so excited to have this conversation with you because I have come to know you over the last year and one of the things that's really cool about your perspective is you bridge these two worlds. On the one hand you work with some of probably the most advanced thinkers about AI that I'm aware of and on the other hand you run some of the most foundational, fundamental skills based training programs. And so I'm really excited to talk about some observations you had and things like that and that I want to start with kind of personal practices because I think it's a really fascinating approach to take your weekly habit of identifying opportunities to automate. I want to start there. Would you walk us through how and why do you do that?

Yeah, so I'm a. I'm a big believer in systems building systems to basically help myself improve and stuff. And so the. Well, one of the systems is exactly what you just said, which is that every Sunday I always plan my week ahead. So on Sunday afternoon I always have to make a decision on what is the next piece of my work that I think is highest leverage to either automate or enhance. And I have a slot reserved for every Friday morning. So Sunday I decide what I want to go and tackle that week. But Friday morning I have two hours dedicated every week where I then actually get to spend time to go and do that. Now I don't have to stick to only those two hours. And if I'm Honest, the last three weeks I probably spent two days a week on this, but it does mean that I have at least two hours a week to dedicate to that automation.

It's critical, I mean, blocking the time. I. I'm sure the stuff you're imagining now is kind of even beyond imagination. Well, can you tell us about a couple of the early opportunities that you identified, maybe early in your journey, and what the impact was of those early automation wins?

So probably the one that I come back to a lot and I've now been using for more than a year and a half is email feedback. It's been extremely high leverage. So a lot everyone talks about how you can use AI to write emails and stuff. And I'm actually not the biggest proponent of getting it to write emails, but I am a really big fan of getting it to critique emails. And so I have a. In this case, it's a cloud project. It used to be a GPT on ChatGPT, which has been trained with how I want to write emails, has been equipped with all of the knowledge of what Mindstone does and how we pitch and how we can add value. And then it's been set up to look at an entire email thread, not just an email, but an entire thread where the last or the top email in the thread is a draft reply. So it's got the full context of mindset and it's got the full context of the conversation. And then it starts critiquing the email, pointing out if I missed something, pointing out if I could reframe something or if I could add some of the data that it has in order to make it more impactful. And I can genuinely point to hundreds of thousands of dollars of business that we won based on that GPT alone. Like, and when I put that claim, it's literal business that I wouldn't even have addressed because I forgot it was further down the email chain. And I had totally forgotten that that was a piece of business we were talking about to begin with.

What are some of the things that you've learned, you know, based on this?

The biggest thing is that the power of iteration. And I think Jeremy talks about this quite a bit as well, which is the, the idea that with this technology, it's about identifying a use case that could be high leverage and then being okay with the fact that the first time around you use it, it might not be great, you make it a little bit better and a little bit better. And now I'm at the point where I've been using this One for a year and a half. It's pretty damn good.

And what's the form factor? It just sends an email to you after that. Or what's the. How does it. Can you talk us through the flow?

Yeah. So right now it's very simple. It's a copy paste flow. So I take the entire thread for in my Gmail. I just paste the entire thread into a Claude and then Claude gives me feedback on how the email could, could be better. Now it doesn't rewrite it and that is key. It doesn't rewrite this. Right. Because then you'd have to go around things of like using the exact wording and stuff like that. I prefer that it just tells me, hey, this section over here misses Personalization or this section over here. You're. You're missing out on a piece of business that you talked about three threads ago. I forgot one element, which is that I also have set it up that if I have had calls with the people in question, I just upload the transcript of those calls. And so that's additional context it has. Sometimes it'll actually suggest personalization elements which like, hey, you should mention this thing that was mentioned three calls ago, which is a good personalization hook for this email.

If I can just go back to your Sunday. I. I find it to be complicated to do that initial identification of things to kind of engage with. And so could you talk a little bit about what is, you know, what prompts you that Sunday? Do you. We've had other guests on the pod that kind of said, you know, sometimes they'll think of something throughout the week and they'll just make like a note in a trello board or, or to do or whatever they do and then back to it. I'm curious on the triggers or the methodology to identify high return kind of projects to, to throw AI at.

So I do have a similar thing where I just kind of. I have ideas that come at various moments and put them aside and then figure out where to go. So I leave. I always put them in my task list. So my flow is I have an inbox in my task list and once a week I then start triaging my task list to go to different days. And Sunday is that day for me. Now, as much as I'm a fan of that, I also really do believe in the, the habit of simply stopping for 15 minutes and thinking, wait, where did all my time go last month out of everything I did and where, where all that time went? Is there anything kind of big ticket items that I can really think about because sometimes like inspiration could, could happen on small issues, which is like really exciting. But then when I go down and say, okay, how often do I really do this thing? Twice a year. I'm not sure if that's the highest priority issue for me to kind of look at. So just forcing myself through trying to tackle the.

Can I just add one there, like little hack that I did. I was thinking the same way. So I now screenshot my calendar for last week and I just take the image and put it into ChatGPT and I have that kind of do that work for me.

That's actually not bad.

I wonder if one kind of hack perhaps that I've been thinking about is what's the thing that's keeping you from the work, so to speak. There's one way to think about it, which is what am I spending too much time on? Which kind of triggers a certain way of thinking. But I find especially in organizations, there's a lot of stuff that people feel like is imposed on them that's actually keeping them from doing their job. Maybe it's the bureaucracy, maybe it's the policies, whatever it might be, and telling folks to look for the things that you feel like are actually keeping you from being able to do your job. A lot of times they go, oh, well, it's this report or it's interfacing with this team, or it's always, you know, sending this information.

I think the interfacing is actually one super important component that people don't use that often. One organization prefer the input to be something different than another part of the organization prefers output. And so you can see like a sales team that wants to have designed a deck like the way that they would brief a designer is not necessarily how a designer like to get briefed. And so at Bark, we created like these like communicators that basically interview the salesperson and then, you know, ask them salesperson kind of questions and then it basically creates a design brief and spits it out in the right format. And so I think that's a super good point, Jeremy.

I think department. Department to department. Translation. Okay, so Josh, would I just. Just to recap what I just heard first, because the bigger question was what were some early wins in automating your workflows? And what you just said is, hey, I've built a cloud project to copy paste in an entire thread of a conversation with my draft reply at the top. And basically Claude has been instructed, knowing what you know about this conversation, knowing what you know about our other conversations that I've uploaded and knowing what you know about our business, Rate my reply and give me a critique. Right, and you're saying that workflow alone has led to hundreds of thousands of dollars of additional revenue that you wouldn't have.

Am I clear on that?

You are correct. Yeah.

Okay, what's another example of an early win that came out of this? Because I love your comment. Believing in the habit of stopping. What else came from the habit of stopping.

So from the habit of stopping. So on the early side there was the movie recommender, a movie and series recommender that came out, which was more on the fun side. So it makes a ton of sense retrospectively. But basically large language models are great with language. And so I think nobody is really happy with Netflix and Amazon and Apple recommendations. It's always, or at least I have yet to find someone who didn't find it very hit miss because they have a whole bunch of other agendas to try and push a particular series or show. But just using a custom GPT that had I think 20 or 30 of my favorite movies and TV shows to figure out what is the Netflix show that I should be watching has been like 100% hit rate. And I'm not even kidding, like literally insane.

By the way, I mean like, I think one of Netflix's call it Coca Cola formulas is their proprietary algorithm. Supposedly what you're saying is leveraging an off the shelf LLM and telling it 20 shows you like does a better job. That's, that's kind of astounding.

It is dramatically better. Like literally, you can't even compare the two.

Wow. Okay, so one thing I'd love, because you have a vision of what's possible that I think is kind of beyond most folks horizon, you run for folks who don't know, you run one of the largest AI communities in the world and you are seeing and hearing use cases on the edge and the fringe and the kind of bleeding edge of things. One thing I would love for you to do is talk about how you think about, call it levels of AI proficiency, so to speak. Meaning I think a lot of folks, if they regularly use ChatGPT, they go, oh, that's all there is. Right. How do you think about, you know, helping someone understand where they are in their journey and what the kind of levels beyond their awareness might be?

Yeah, it's a really, really good question and it's an interesting one because it's probably one of the first times that I feel that wherever you are on the ladder you think you're at the top of the ladder, which is just a weird situation to be in. So you've got the starting point. Those that still think that it's all hype and that the best choice is to not use it because it's only going to get in your way. You then have people that think that really mastered it, but because they're no longer using Google, they're using ChatGPT. And they're like, and they're using it a lot, they say, oh yeah, I really get it, I'm no longer using Google. And I was like, okay, that's another level. Then you have those that start to make the difference between, okay, well, Google gives you one thing, ChatGPT or an AI assistant gives you something entirely different. So you get to the point where you understand that almost by definition, if it was a good Google query, then you should not be using it in an AI assistant. And if you're using it in an AI assistant, it probably shouldn't be something you send to Google. Once you start making that difference, you start to look at a whole bunch of other use cases. This often comes with people looking at reasoning tasks. So that's when you're no longer looking at it as an answer engine, but you're looking at it as a kind of a starting point of reasoning through specific viewpoints. But you still are. You, you go into, I guess, conversation mode. That is probably the next step. And then the step after that is when you start using it as a proper thinking partner. And the biggest signal for that is when you start to use AI to ask you questions instead of asking questions to the AI and expecting answers. Right. And that's where it now starts to stimulate your own thinking. Step after that is when you start to go outside of AI assistants and you start to explore other apps like Perplexity, Notebook, gamma, recently PowerPoint, copilot. Starting to do some stuff that's interesting, but basically going a little bit outside of the, the general AI systems and starting to look at tasks that, that touch multiple apps. And then the step after that is when you start touching on building stuff, which is the what happens when you almost similar to someone who explores and for the first time understands the power of Excel. And for a very long time, like basically Excel runs entire companies in some cases. Well, that's where we're getting to now with app building. Like, you don't have to be too technical to be able to put together a general app that just manages all your employee holidays or something like that. Like that's. Very, very simple to go into. And then you've got another three levels behind that, I'd say that are more engineering led.

No, but when you get to building stuff, I mean, so there's a bunch of stuff here we could break down. And I know that, you know, building tools are not even the peak, but they're another kind of area. Talk for a second. I've heard you use this phrase before, Josh, which I really like, which is a utility threshold. Can you talk about how you define that and how a user can know whether they've crossed the utility threshold? I think that's one of the first kind of sound barriers a user has to break, so to speak.

Absolutely. And this is, this ties to some of my kind of pet themes where I hate people talking about how these models are not getting or how we're getting less out of their progress. But the, the, the idea you can think about the capability of AI and, and this AI threshold at the point at which an AI systems becomes useful, it is directly related to its level of intelligence. But it's basically the point where the time it wins you or the quality increase that you get is worth the time you put in. Now that means that if the model is 98% of the way there, which means you're putting 100% of effort, you get 98% back. It is useless. It can be extremely interesting and it can be mind blowing from a technology perspective, but it's useless because you've done the thing yourself. So if it creates a great poem, but it takes you. I don't know. Everyone has their own utility threshold. Here me, it would take me weeks to write a great pun. Somebody else would probably be able to do that instantly. The AI, if it takes them more time than it. Or if it takes me plus the AI more time than it would take me on my own.

Not you, but even like an email, right? I mean an email is a good example for most people. Most lay users aren't, you know, poets. But you go, how long would it have taken me to craft an email I'm proud of versus if what you're saying for the utility threshold is if collaborating with AI doesn't result in a call it meaningfully improved or meaningfully faster outcome than you on your own, meaning you get something from it, then you got to refine it and work it, whatever, it has yet to cross the utility threshold and it remains a toy.

Yeah. Now. And what is interesting is once you cross the utility threshold, a tiny increase in either your ability to wield the model, or the model itself yields dramatic increase in utility. Because now imagine that you're getting. Initially you get 102% out and it takes you a hundred, right? So you take 10 minutes and the AI does it in whatever, nine minutes, 58, whatever. So you get the two seconds. Now the model gets a little bit better and it does it in nine minutes, right? So actually what has happened is you just won that two seconds, but times. Times 50 in this case, right? So you're really, you're getting a dramatic uplift with a small increase in the model. And this is the bit that many people are currently, especially technologists, are just not getting in the conversation of how this technology relates to. To actual productivity, which is everyone talks about these benchmarks and how the AI is getting only slightly better on a particular benchmark. A 2% uplift on a particular benchmark might mean that there are a thousand different use cases that just went from being useless to useful. And so it's unlocking a whole bunch of other use cases, which then in turn gets those people to the other side of the utility threshold. Then they start to use it more. And when they start to use it more, their own proficiency becomes higher, which also has this effect on the utility threshold. Right, because the easier it is for me to use the model, the less effort it means I need to put, the more utility I get out.

You know, there's a phrase that someone used on the podcast we had, Bryce Shalomal, who's the head of AI at Moderna. And one thing he said which has stuck with me and I think resonated with a lot of our audiences, I can't imagine doing any part of my job without layers of AI baked into it. I think he said something like that. I can't imagine. And then he went on to say things like, it'd be so lazy and stupid and reckless, of which, yeah, I love the kind of impact of that. But I think speaking of the utility threshold, you have to get to the point. For a particular set aside all of your work, is there any part of your work for which you as a knowledge worker could say, I can't imagine doing it without AI. And I think for a lot of people, probably the truth is, oh, yeah, I can kind of go back to the normal way, but there are, I would say for myself, there are foundational work products that I go. It would set me back a long time if I no longer had the ability to collaborate with AI. And I think that to me is, I don't know if that's a. That's. It's evidence of utility threshold. But it may be like a mile marker, so to speak, on the journey. How many of those things can you point to that you can't imagine doing without AI?

Right. So honestly, at this point, from a, almost if not all of my work, no, I do live delivery, but my live delivery is showcasing AI. So I could, I could argue every single aspect of my job is knowledge work. In which case I don't think there is a single part of my job that doesn't get dramatically enhanced with AI. To the point now. So the last few years have been interesting where we went, I don't know when was it eight years ago or something like that. Through the whole wave of getting wi fi on planes and for a bit, extremely good. It really, it worked very well. Started to be able to do some emails and you were kind of in between because you were hoping, okay, you get some Internet. It's an additional productivity boost you go through. Now when I don't have wi fi on a flight, even the thinking work feels like it's wasted work because I know I'm going to have to go through the exact same steps again because I don't want to miss out of the input that the AI would have given in the thinking work.

The boost. Yeah, it's like, do I want a 10x think through this or do I want a 1x think through it?

Exactly.

Do you have anything? Because I think that's very true on a lot of different dimensions. The thing that I've been increasingly thoughtful on is what is it that. Where I have to force myself to kind of like lean into my humanity. Like, what would you say is the thing that, you know, even if you can create a bot that could do that, you would really want to do that yourself.

From a work perspective, I don't see it. It would be on the. At a personal level, it would be about the relationships I craft with the people I care about around me. I guess that would be the only. And even there, I would still think there was room for AI assistance in like making sure I don't forget a birthday, making sure I don't forget Mother's Day. But I would want for myself, for that relationship to build, for there to be a real connection. And that would. That can only come through me.

One might call it relationship capital. No kidding. Only because it's a core theme of a book that I have coming out August 5th. So, like, nice setup there. But it is fascinating, right? Like the people like you, who are such a super user, increasingly obviously would be completely deprived that it was gone.

Right.

So therefore, like, that's obviously, you know, interesting, but also increasingly have a little bit of a tough time seeing what is it that we as humans are completely unique to kind of do, except our ability to emotionally connect with other humans.

Yes.

And now I think we're seeing with the rise of mates and chatbots, you know, set aside, I mean, even romantic partners. But I think for a lot of people, there's a question of whether the, whether that's a uniquely human ability. That's not my personal feeling, but I would say as I observe the world around, there's an increasing number of people who go, you know, I saw in the paper the other day. Yes, I read the paper. It was like sitting on a hotel desk. It was like, somebody's marrying a chatbot. It's like. And so you start to see there's a lot of. I mean, think about someone who's always available. Right. Like your human partner has to sleep, for example, but chatbot could be available to have deep emotional conversation. Right, Right.

I think, I think you, you touch on something interesting here, because for me, there are two different components. So for me, the, the component that you feel your need. My need might be met in different ways. My need to feel heard in a way, or to feel like I understand myself, or to have an outlet where I can out my frustrations, whatever that is. I can imagine that need being met in some cases, maybe not fully, but in some cases to a degree by an AI. But that is not the same as the need that I have to express myself to somebody else. Hmm.

We talked a little bit about in an earlier part where, you know, you can get your GPT to count your calories, but for some reason it feels better to send it to your trainer, because there's a need for connectivity. If I can just, you know, speaking of need, like, you obviously are so good of kind of doing all these tools. And I think for a lot of us who kind of like are in this world, like, that's a little bit like catnip. First you talk. I saw on your LinkedIn about making all people, like 40% more productive. And I think a lot of people listen to the podcast. They sit with a role where it's their job, either because they're self selected or because their company have asked them to, to kind of like help their organization go through that. What is some of your current advice when somebody goes like, hey, I have an organization. People kind of Curious, but like it's being used, but only a little bit. What's kind of like the way to kind of kick it in the butt?

So. So there's two things. One is the simple realization that no technology has ever adopted itself. So it's very weird that we're expecting this one to somehow be that. And that is exactly why it's so treacherous. Because it appears we get back to these different levels of knowledge and you always think you're at the top of the ladder. And so it's so easy to get going that it's also very easy to fool yourself to think that you now know how to use it. And then the second bit is show, don't tell. I am genuinely very sick of the amount of people that talk about all of this stuff and that honestly, when you ask them right after, okay, so what, how did you use the. Use AI in the last, whatever, two, three days? And they can't come up with a single example, right.

Or they share their screen and you look at their history and it's like there's either there's three conversations in total there, or, you know, it says last 30 days. Like there's nothing today, there's nothing last seven days.

Right? Yeah, that's exactly it. But so. So to get back to that is when you get the chance. And it takes. Maybe we've gotten better at this over time. So it takes maybe 30 minutes, I'd say, where I'm. I can take any skeptic in the world, probably sit down and get them to the other side to realize, okay, shit, this is real and I really need to do something about it now. But it requires a live demo. I don't know of any other tool to be able to get someone past that gap.

I think that. And I think also, like there's increasingly like a gap, knowledge gap, right, between people who know a lot about it and the people who kind of like haven't really gone into it. I think what we found useful is to just kind of not personify it, but just think about it as agents. I think people are so used to thinking about colleagues that if you start to go, hey, there's like three types of organizational, basically agents that we can make. We can do one for the organization. So like your HR kind of bot, so that if people want to. You're talking about like travel days, you know, or it could be, you know, what's our rule for sick leave? There's like, I would say very department focused one. So you have a supply chain team, they Get a bunch of invoices in. They'd like to get it in like a structured format and so they don't. And then there's like the hyper individualized one where you have. For example, I'd like to read a lot of newsletters, but I often don't get to read them all. And then I feel frustrated. And so now I have like a script that kind of like summarize them for me. Is that a form that you use or do you have another way for people to kind of like get an easy mental model for how to think about it?

I think that is indeed it's that way of thinking about it is actually very traditional. The way that I hear that is every example you just mentioned is basically we could meet that need through building traditional technology. As a software engineer before generative AI, I would have been able to build a solution that, that automates that. What doesn't get talked enough about and I am also not doing good enough job at it, is the ability to dramatically expand our own thinking power using these tools. So the use cases that I think are much more powerful are helping me brainstorm solutions to a particular problem, helping me understand a particular problem from different facets, helping me understand how different stakeholders might respond to a particular situation ahead of time, helping me develop the strategy for Mindstone for 2025 and have AI as an actual thinking partner, critiquing what I'm doing, suggesting where I might want to go and improving various aspects here and there. And it also paints this picture of two different worlds. And I actually think this is probably one of the biggest risks we have, which is that AI has always been about automation. Up until this point. We very rarely. I know it sometimes gets off the. But, but we rarely talk about augmentation. And the problem is that often when people don't understand what is happening at the moment, the responsibility of AI often lands with the CTO or the CIO because it's technology. And then a CTO or a CIO is often very. They're great in operating in a deterministic world. That is literally why they're great at their job. Because that is what technology required you to be able to do, to extremely rationally be able to take a very complex system and derive a set of rules that will get you to an outcome that automates that process. That is the skill set you need to be a great technologist. But if you then take this technology whose power is specifically at non deterministic use cases, but you put that in the seat of the person who's great at all the deterministic scenarios, they end up using it for all the same scenarios, which is that they end up using it for automation. And so the result of this means that the organization ends up automating more rather than if. And this is one of the biggest battles that I'm trying to get to. This should be in the hands of a progressive HR team who are like, hell yeah, we are going to equip everyone in our company with the ability to do twice as much by the end of, of this year. And then you end up with an augmented workforce, which then means we are all reaping more of the benefits. And I really think that it might well be that difference. And if, if we don't get there, we just end up with a whole bunch of unemployed people. And if we do, we end up with a enriched society that actually is able to do that.

I think that's such an astute point. But what I find complicated is in large organization. And maybe we can have you talk a little bit about how you build big kind of teams before, and I assume now building them in a new way. So let's kind of like end there. But I think you're right. And obviously we see some of the new, some of the Ethan Malik kind of like research come out of basically how if you allow people to spill into other functions and their core kind of like function, then, you know, which I think is a little bit of what you're talking to. A lot of teams, of course, have a tough time doing that, because while AI will allow an engineer to suddenly think as a marketeer or a marketeer to kind of code some things that they can make it actionable, you have structures that makes that complicated. Do you have a view on how this HR manager, which I think is, you're totally right on how can they navigate the structures that they kind of work within. And then maybe you can talk about as how you've done it with your organization.

Yeah. So two very different things indeed. The first thing is just at organizational level, I think, one, the entire role of HR is about to become absolutely critical. And going from a point where it used to be administrative, actually, it's not too dissimilar from what has happened to the cto. If you think about the CTO used to be the IT department before technology became critical to the company. You had an IT department and a head of it. And then companies were built where their entire value proposition was around the ability to leverage the technology. And so a CTO became a normal role. And I think the same is going to happen from a people or needs to happen from, from a people perspective. They need to be elevated. But it, it doesn't, doesn't just work by creating the position you need the type of person that sits at the C suite claims their position and drives that function in the way that function can actually be driven and the value that that can add to the overall organization. You need both a slightly different type of person and you need the space for that type of person to be able to execute. Now actually, on the Mindson side, interesting enough, we are tiny. So luckily enough, we don't have that problem at the moment. Now we are tiny comes from two bits. One is that this is an early space and fairly newish company. Second is that we also walk the walk. I would say we probably do a dramatic amount of work for the number of people that are in the company because we live and breathe using AI every single day. It's not a single entity. In this case, it's me driving it.

Can we talk for a second perhaps about that, Josh? And this might be a fun opportunity for you to get into your backstory a little bit. Talk perhaps for a moment about your entrepreneurial journey, the company you built and then exited, and then how your approach to company building has changed now that you're AI augmented.

Yeah, so I built, so 10, 13 years ago now started a company called Super Awesome. I was a co founder. We were five co founders that became one of the biggest kids technology companies in the world. We were about 200 people when we were acquired by Epic Games, the creators of Fortnite. And with Mindstone, we haven't gone past 10 people. So I think gives you a, it gives you a good idea. But we are getting close to starting to look at real numbers now and the, the key difference. And there's. To be honest, I have hit a few walls myself as well because this wasn't like a straight path. I'm a software engineer by background. When ChatGPT happened, initially everyone thought I was going crazy. And when I say crazy, like I nearly lost the whole company over it because everyone thought why the hell's the CEO spending 50% of his time on this technology, which is clearly not having an effect on our business. And I did so for like four months until I hit a particular point where it started being useful for the business. And then GPT4 came out and it became clear that it was going to be useful. And then basically the year and a half after that were a Series of steps, realizing of how right I actually was and then trying to baby step at a time, think, okay, how sure am I now? Should we really go all in? And then over time we went all in and that, that we reconfigured the company as such. Now the other bit that was interesting is I, when I was building super awesome as a software engineer, I always was scared of losing touch or losing touch with a code, losing the ability to code. And when I would get frustrated by our, whatever we were not doing as a team, I would have these spurs every three years or so where I'm like, I'm going to get back to building stuff and then I would open my development environment, start to look at the code and about three or four days in realize what the hell are you doing? You can't get anything done. Things have moved on. This tech stack is too advanced for you. Like it's going to take you weeks before you get to productivity again. And by that time everything else you're supposed to be doing has moved on and it's more important. And so I think like literally three or four times during the nine years of building Super Awesome, I went through that cycle and I, I made it very clear myself and I thought I had learned the lesson, Josh, you no longer build. And then generative AI comes along and suddenly the possibility to build and the, the ease of building has just dropped through the floor. And so I, I basically got hooked again. And I got, I went through the same cycles. Okay, am I kidding myself again? Should I be doing this? And initially I thought I learned the same lesson again. I can build nice shiny prototypes, but they're not making it to production. Was that really useful? Yeah, not so sure now. I am entirely convinced of the opposite. I now have three full blown production apps that are live that I built over the last three months and they are handling significant loads in the company that otherwise we would not be as a team, we would not be able to do what we're doing without them. And I built them without writing a single line of code. It was all AI built. And so that is one of the lessons that I'm really thinking differently. This, and this is a whole other theme which I don't know if we have enough to dive into. But the capability for individuals inside a company to just build the software they need instead of buying it is going to fundamentally transform how all businesses will be built from here onwards.

There's kind of, there's like this age old question of build versus buy, right? Which the, the Fundamental algebra of that equation has totally changed.

Do you believe that basically SaaS, you know, classic SaaS is dead because of, of this? That.

Yeah, I, I really, I, I give it two years, three years maximum.

And so who do you think will be building this internally? Is that the HR team that will build their own stuff because they're now understanding how they can enhance themselves.

So you need some points throughout the company where someone has an engineering mindset, basically being able to take a set of requirements, understand the process that needs to be followed to the problem that needs to be solved, and then try and figure out what does a good solution look like. But they don't have to be technical. Right. They can be a product type person. They can also just be a very structured person in that particular team. Some companies might decide to outsource this to their IT team and then their IT team becomes the, the, the hub that ends up building this. Some companies might decide that this now becomes part of a cross functional team where you always have at least one person with that mindset.

Back to your point, like one of the issues sometimes with engineers is obviously their, their empathy UI is not always as evolved as other teams.

Right.

Their bedside manner is lacking. Is that what you're saying?

And, and so it is interesting, you know, like just, just look around. Like the stuff that I've seen is, you're right. Like the engineers obviously knows a lot about how this stuff works, but some of the best kind of like uses of AI is not necessarily the engineering team that comes up with it.

Yeah, just to be very clear, that's why I was specific in using the word engineering mindset. I'm not talking about engineers. I would suggest that, well, okay, I am an engineer by background, so for me it's easy to say that there are many more people that have an engineering mindset. Most founders have an engineering mindset because it's about taking a problem, deconstructing it into smaller bits, and then trying to execute on those smaller bits or putting them together into a solution to the problem. You don't have to be technical at all. And I do agree that that role, anyone that can sit at the intersection of understanding the real problem, understanding the people involved, and understanding how you might go around solving that is going to be in an extremely valuable position for the next few years.

I mean, just as one example, one very practical example we've got, I think all of us probably, you know, at least superficially know this person, but this person is in a large entertainment organization, has very kind of call it baseline technical ability, not a lot of technical ability, but identified a problem in the business around resales of their product that were basically cannibalizing their ability to be the primary seller. And through kind of a slick understanding of very simple AI building tools, built an application that is providing an offer to the marketplace to do something that you know, would have taken a team, you know, two years ago, six months to develop a product. And this person just did it in their free time because they, as Josh, as you said, they sit close enough to the problem and they have enough of an engineering mindset. They kind of said, well could I just kind of patch this together? And they're actually doing a beta launch with a bunch of customers and it's a fundamentally different way to approach, you know, vendors and, and, and product development and solutions then for sure, I guarantee this organization has ever done before.

Well, I can, I can point. So I had my own example two months ago where nearly signed a $15,000 SaaS contract until I realized I could build the whole thing myself in about three hours. And the next morning, rather than having signed the contract which I was initially going to sign the night before. But I stopped to try and figure out if I could do it. The next morning I'd built the entire software and I did it better than I thought the off the shelf state would have done and it saved me $15,000.

Wow.

Wow.

How do you think? Because I obviously have also built large companies where there's now hundreds and hundreds of people and then now I have a new organization that I'm most involved in where we can kind of build from scratch. I think it's obviously sometimes easier to build from scratch because you don't have like the legacy issues and so you can do all these different things. I, I'm still puzzled about like what would happen if at BARK when there's 700 people, somebody built something internally, you know, like how I bought for sure, because we're so all over AI they would be allowed to it. But I would imagine like if you sit in a large other organization and you suddenly were kind of like throwing data here and there and like if you would get stuck and so how do you kind of like convince people up the chain that this is something that is very useful.

We, we right now. If you find the answer to that, please let me know because the only thing that we found is like the executive team needs to. The rest is capitalism.

Yeah.

Literally it's just going to be that the team that leans in two months before the company Next to it is going to win. That's just what's going to happen.

I love it when the answer is capitalism. I want to zoom. Just kind of last question as we're wrapping here. One of the things I've heard you talk about, Josh, that but. But you haven't mentioned here is just your regular habit of just having a conversation with ChatGPT, just going on a walk and just having a conversation. Can you tell us a little bit about that practice, what some of the kind of conversational topics are and why you have chosen to do that?

So yeah, so this was actually an extension of existing habit that I had. Once a week I have, I literally have a calendar item which is weekly thinking. What? And it just forces me on a Friday afternoon to get out of the house, to go on a walk for 45 minutes to an hour and to think through whatever is on my mind at that point in time. Sometimes, actually most of the time it's a work related thing. Sometimes it's a personal issue that I want to think through and it's really just about giving myself the space to think it through. And then at some point I just started using ChatGPT voice to go and have that conversation and I just put AirPods in and I start talking about, I have this thing I want to think through. I want it to act as a thinking partner. I want it to ask me questions to help me think. And again, the entire thing, the purpose for me is to use the AI to help me think. Rarely do I ask it to actually give me real advice. I'm specifically asking it to ask me questions to help me explore aspects of a particular situation I might not have explored. Which indirectly still means it's stirring me in different ways. But the it's really about helping me think. And I do that for an hour at a time. I even started doing it now on my bike ride into London. So it's almost like an interactive podcast on whatever topic I want to think through. So on the bike in, I tend to have it more on exploratory topics, literally as if it's a podcast. So I'm like, today I want to learn more about how, how tokens work in large language model learning. And I literally ask it to just start with an explanation and then I just fine tune it. Oh, actually can you dive into this topic a little bit more and then like how does it interact with this other thing?

I've biked in London and I'm not sure that's safe, dude. I mean like I'm not sure.

Do you want. Don't forget I was born on the bike, right? So I'm Danish.

So I think I might be up there with you though.

But so you have the time. So what I'm hearing, there's actually two things. One is while commuting, but then there's another which is a dedicated time where you're. You basically have an augmented conversation partner to explore a topic of strategic interest or personal concern on your calendar. When did you discover AI could be a part of the conversation? And how has that changed your thinking patterns?

It's dramatically improved the quality and efficiency of those moments. Now, having said that efficiency wasn't the only thing I was looking for during those moments because part of it was about walking out and having some space for my thoughts to kind of calm down. One of the biggest unlocks for that, by the way, was when ChatGPT started managing pauses much better because before the early days it would always expect an answer and then it would switch off when like I didn't answer for a minute. Now I can just walk around for five minutes, basically let sit and think through the last interaction I had and then I'll just start with the conversation again. You mentioned. And sorry if I'm jumping a little.

Bit, but I just wanted to make sure. I think a lot of people ask me about slide decks and using AI for that. You were just. You had an off comment earlier about PowerPoint's copilot becoming better. Can you just put a few more words on that?

Yeah, so I mean, Jeremy will love this bit as well, because I came finally to the conclusion that I am happy for people to start using copilot. They release different things. I will show you what is possible now, Jeremy.

But yeah, all right, you have to convince me.

Finally, you can now properly draft and use Copilot to be an actual file partner using track changes in Word documents and everything. It can also finally produce and interact with Excel sheets, which up until now you had this nice interaction on the right hand side and it could tell you things, but it could not actually modify the Excel. Now it can go in and add columns, add calculations for you, and then at the end of that process do the visualization and stuff. So it's finally getting there. And the big kicker one, they brought back GPTs. Finally, agents are a thing that everyone can use. They call it agents even though they're not proper agents. They're not proper agents, but they call them agents and they are extremely easy to create. They are using the exact same interface as what used to be the GPT creator that they already had a year and a half ago, but they are now exposed again. You can use them in the enterprise version and it's one of the biggest productivity unlocks that you.

I mean like an. This might get too nerdy. I went whole down on the Power Automate for a while because, you know, wanted a lot of friends that were in a Microsoft kind of universe and. And it was just so difficult to do anything and so are you saying this is the time where you might want to try to dust off the good old Copilot and see what it does?

So yeah, Power Automate, that's not what I'm talking about. It's entirely different, honestly. I mean I literally just went through it this morning because I had the benefit of having to put together an entire a spin on our program that was specifically for co pilot. So I wanted to bathe back into it and make sure that I really understood what was happening there. And yeah, Copilot now I would say has actually gotten back to a point where it's actually useful and it's usable by non technical people again.

So it's moved out of this clippy kind of like moment.

It is.

I'm hesitant to declare victory for anybody too quickly. So let's not declare we're in a post clippy world. One, one last thing on my mind because you just used a phrase which I think is a key phrase. One of the biggest productivity gainers you mentioned is building GPTs, which it sounds like Copilot incorrectly calls agents. Can we go back? One of the very first comments you made was about the importance of iteration. And I think that would be a nice place to kind of land here because Henrik, I'm remembering our conversation with Russ Summers and one of the things he said, we asked him, how do you know if you're doing a job? And he said when you refuse to settle, when you keep tweaking the instruction set, you know you found something that's valuable to you. Josh, can you talk for a second about what you consider to be the biggest productivity gaining feature and how someone should approach creating whether Copilot calls an agent or GPT or whatever.

Yeah. So the starting point is to identify a task that you do often and that is of high value. The higher value the better because that means that there's enough space to explore and your utility threshold basically is fairly low. You can invest a lot of time and it would still be worth it if the AI ends up being regularly useful in a process that's high value to you. And that is key because it's rare. Even for me. It's rare that I get a GPT right on the first go. You get. You try it out once. The sales email assistant is one. Recently, I built one for creation of program content. So literally in the last few days, I had to build one that takes a particular topic, takes my train of thought, and then translates it into a series of specific pieces of content that we can use in the program as we teach people how to use these tools. It's extremely high value for me because this is what we do. But I think in the last two days alone, I went through 20 different iterations because every time it came through, it was not quite there. Now the iterations are extremely heavy at the front end, which means that I have to get it right or I have to tweak it maybe 10 times the first 10 times I use it. And then it starts to ebb out because I start to watch. That's good.

And I think just put a fine point, what you're saying, Josh, for the audience as well, one where you could think about interacting with the GPT is it gets called 80% of the way there. And then you handcraft an artisanal kind of last 20%. The important step that, Josh, you're getting at what you said you did 20 times, is instead of just handcrafting the perfection, you actually go back into the instruction set of the GPT or the agent or whatever and say, this part, this last 20% wasn't great. So I clearly need to change how I'm instructing it. So the next time it's 82% of the way there, 84% of the way there. But every time you get an output, you're actually changing the instruction set to improve the output. Is that correct?

That is exactly it, yeah.

I think it's a huge point. If you don't care that the model doesn't give you really great output, you probably aren't choosing a high enough value task.

Yeah.

I also find, and this is maybe lazy, that I then model jump. And so if I feel a little bit stuck, I'll just take the same thing or what the last one, just output it and I'll just kind of walk through the models like, you know, Cloud Gemini, stuff like that. And it's a wave often just for me to get unstuck because, like, I get, like something different back.

Yeah. So I would. That works. But that means that the next time around, you're going to have to Jump around again, if you incorporate that into the logic, you, you remove the need to jump next time.

Yeah. I do think though that there is some models that have different benefits. As you were saying earlier, you went from GPT to Claude. And I do sometimes find that by trying to almost like going to different kind of co workers and saying I know that you normally have a more empathic kind of like way or you're a little bit more, you know, better coding, you know, coding logic. So I agree with you, but I maybe just my, my wanting to talk to all the models.

Gentlemen, we've come to the end of another fun filled episode. Josh, thank you so much for joining us.

Jeremy, oddly, what your, what you're thinking.

I mean I've had a, I've had the privilege of getting to see Josh in action in a bunch of different contexts, on a bunch of different stages, so to speak. And he's someone who, I think he's on the leading edge in terms of his not only his ability to use AI but also his ability to help others. He's really, he's singular. He's a singular individual as far as I'm concerned. I loved hearing him talk about utility threshold. I loved hearing him talk about the necessity of iteration. I mean the fact that even if like take as a given just for a second, just imagine you're listening to a world expert on AI. Just take that as a given for a second and then you say this world expert needs to build a new GPT to accomplish a purpose. How many iterations do you think it's going to take? I think most people go, the more expertise you have, the less iteration is required. Right. And what you just heard is I bet the expert did 10x the iteration that you did as a less experienced person. And to me it speaks to some of the foundational behaviors we have to get right. One around, iteration around, not settling for mediocre output, recognizing that spectacular output is possible. And then of course the other really big piece I think is how he wields his calendar. I think it's non trivial that you know, on Sunday he has a weekly habit of looking for where are high leverage opportunities to automate his workflows. And then he has time blocked every Friday morning to actually build those automations. And then he also has time on Friday afternoons to do more of a strategic reflection conversation. The fact that those are actual blocks on his calendar that he could refer to as when he does that regularly I think speaks to how he wields his calendar as a weapon. Most People are the victim of their calendar. It's like, I can't do strategic thinking because of all these meetings. I can't automate anything because of all these meetings. And then you hear someone who's really got expertise say no, it's so important. I block the time. I don't take meetings.

Friday morning, I can't remember what book I, I read it in, but recently they talked about timeboxing and made the same point, not only with work stuff, but also with private stuff. Like, I mean like you have, you assume that you always have time to your partner or your kids, but the reality is that you're a slave to your calendar. And so I, for example, now like, you know, date night every Friday, you know, like my wife and I have these walks with the dogs, like, you know, scheduled. So I think, and obviously with learning, which is something that you don't necessarily feel that you can monetize immediately learning AI, I think same thing, you know, like time, box it in. So I, I, I noticed that too. And I do think it's one of those things that's good just to kind of pass on a little bit and, and kind of bring out because most of us don't do it and we probably could yield a lot of benefits for doing more of it.

The other comment when you spoke about augmentation versus automation, it reminded me a lot of Bryce. Do you remember in that conversation how he spoke about chief AI officer? You know, in five years is going to sound like chief mobile officer, right? There's, there's not going to be an AI officer, but I think there's probably a convergence point around, call it AI and hr. And Josh's comments around the most high value opportunities are the non deterministic functions where it's not that we're automating deterministic things, but we're augmenting non deterministic humans to me, reminded me a lot of how Bryce thinks about it. And it's how I think you and I both think about augmentation and thought partnership as the much higher value opportunity than simply automation.

But I think, you know, you know, as we heard this conversation, it's complicated because we say that and I, I do think that is a hundred percent correct. Meanwhile, when you come down to examples, it very quickly becomes examples where you took something that took an hour and now you can do it in 20 minutes because obviously the measurement of enhancement can be more complicated. Saying I, I wrote this email better. I guess he used the point of like having made hundreds of thousands of dollars more doing it this way. Or pounds yours.

Well, here's. So, Henrik, here's a cool example. I don't think I told you this. You'll love this. So, as you know, I've been working with the National Park Service here in the US Doing kind of capability building. And we have these monthly office hours where folks from the National Park Service will share examples. And then usually it's kind of half case studies from folks in the service and then half kind of Q and A time with me and with. And with the folks you presented. And I consider myself a front row student in the classroom this past week, we had a woman who said, yeah, I've created a tool that helps folks write their impact bullets. And I go, I'm sorry, what? What are impact bullets? And of course, I'm, I'm kind of out to lunch because anybody who's a federal employee knows that Elon and Doge have basically required every federal employee write bullets of what impact they're having, because their job depends on it. If federal employees can't demonstrate that they're having impact, they're being terminated. And this woman built a tool that a GPT basically, that basically scrapes all of Elon's public statements and Doge's public statements about what they consider impactful work and how they consider framing. She said, a lot of employees don't know how to describe what they're doing in a way they understand. So it's literally an English to English translation where a Park Service employee, she said, thousands have used it in the last couple of weeks.

That's awesome.

Park Service employee can say, here's what I did. How do I describe that in a way that sounds high impact? And then the GPT actually recasts their impact in a way that Doge can appreciate. That to me is like a quintessential example of it's not automation. It's actually augmenting those humans and helping them demonstrate the value that they're bringing to their work.

That's a great example. And I think, you know, like, maybe just to find on something super practical, because it is awesome when after these conversations, we have notes like, I think the movie recommendation is something I've never done, which just seems to be an obvious one that I just need to try. I think the holiday app, like, you know, manage all the holidays, scheduling, like, just seems to be an obvious thing that I haven't built yet. And then, you know, having had soured a little bit on Copilot, but obviously using Excel and PowerPoint I think it's, it's time to kind of like dust it off and then maybe in one of the future episodes we can talk a little bit about the results.

You hate to admit it, but it might be time to relaunch the Copilot app.

On that note, thank you so much for listening as always. Really, really appreciate it. And if you don't mind, we would be incredibly grateful if you could like, subscribe, share and all this good stuff to get our podcast out to a bigger audience. So thank you so much for this time and.

Oh, we didn't have it. We didn't have a code word.

Is it too late?

How about retry copilot hashtag Retry copilot.

Hashtag that seems like a promotion. Hashtag we just got a new sponsor.

Hashtag new sponsor. Either one's fine.

Awesome. Jeremy, thank you so much.

I'll talk to you soon.