Investing Billions - E441: Startup Investing When the Cost of Building Goes to Zero

Episode Date: October 12, 2026

What if AI makes the founder more important, not less? John Borthwick is the Founder of betaworks, the startup studio and venture firm behind companies including Giphy, Dots, bitly, and TweetDeck, with investments in companies including Twitter, Tumblr, Kickstarter, and Hugging Face. We break down why John now puts more weight on founders than products, how betaworks identified Hugging Face before transformers reshaped AI, and why he believes the biggest opportunity is moving toward the AI application layer. We also discuss open-source AI, AI security, the changing nature of work, and why John believes AI was discovered rather than invented. John Borthwick is the Founder of betaworks, the startup studio and venture firm behind companies including Giphy, Dots, bitly, and TweetDeck, with investments in companies including Twitter, Tumblr, Kickstarter, and Hugging Face. We break down why John now puts more weight on founders than products, how betaworks identified Hugging Face before transformers reshaped AI, and why he believes the biggest opportunity is moving toward the AI application layer. We also discuss open-source AI, AI security, the changing nature of work, and why John believes AI was discovered rather than invented.

The crust of building technologies as a technology is trending towards zero. Five years ago, ten years ago, founders would decidedly builders. They were engineers, they were in the code, they were building. Now they are more often than not sort of steering a set of technologies, a set of agents, a set of AI tools towards sort of a goal. I ship code this year. I've like built apps myself this year, right? It's extraordinary what you can do.

The power of creation is now at so many more people. people's fingertips. You were reportedly the first investor in Hugging Face, which sold for $13 billion to Nvidia. When you talk about founders are now more about steering products and they're managing all these agents, how would you define that skill that you're looking for? John, you've been investing in technology for over two decades. What have you changed your mind on the most recently?

From the outset, I've always been very sort of product driven, product first. I think that the balance between project. and founder has shifted a bit. If you say that, you know, thesis, product, founder was the stack for me, you know, 20 years ago. I think that now it's probably founder is a little bit higher or relative higher. Getting to know the founder has become more important to me over time. Why? I think that the technology is like we're living in an extraordinary time.

And the pace at which technology is advancing is unlike anything, I think, in human history. And they kind of founders that you need, who can, you know, that you want to back, who can really sort of take companies through their various stages of evolution. Some founders grow into, they grow through a ceiling and they grow into scale. And other founders just like get sort of crushed by it. And it's just like the kind of skills that you need at the sort of day one, which is when we meet the founders, you know, are different than the kind of skills you need at series B, series C, or like when you're fully at scale. And so we have to, I mean, we're there right at the beginning. And so we have to sort of think about that too.

And that is something which at the outset I really didn't think that much about. You really have to figure out which founders, when they get pushed down over and over again, which ones are. are going to get up and which ones are just going to crack. How do you go about sussing that out ahead of time? A lot of it is just about sort of the human dimension and trying to get to know the founders, right? I really try to get some kind of degree of rapport with the founders. And one of the questions I always ask founders is like, why did you name your company X? And what I'm trying to see in the response is I'm trying to read if there's an emotional

connection to the problem that they're solving is there's an emotional connection to the brand. You know, occasionally I've had founders who are like, well, I was looking for domain, I found a domain, and or I was drunk one night, and I came up with this idea, and I have no idea why it's called that. Although their names were taken. Yeah, that to me is just like not a good signal. It's a good signal because I want to see that attachment, because invariably the entrepreneur's journey is when you step back these things like, you know, go all the way back to the beginning of Badeworks, these things like Twitter or Bitley or Giffy or Hugging Face or granola now, when you step back, they look like they are straight up and to the right,

but they're like fractals. When you actually zoom in, there's some bumps in there. And you want founders who can be able to basically, you know, work through the bumps. I think names are extremely underrated. Peter Thiel is obsessed with names and really focuses on the names. He thinks that names dramatically change a trajectory of companies. Never knew that. That's interesting. I mean, I don't follow his work much. I just sort of drive my own path, but that's interesting. Is that why, like so many of them are Lord of the Rings? Because that seems to me to be... I don't know how those things are related, but he seems to think that names, in many ways, make deterministic outcomes. He argues that Facebook versus MySpace was one of the reasons that Facebook actually

beat MySpace and less space. And he has other examples. Ryan Surhant, previous two-time guest, he talked about calling his brokerage Surhant about this affiliation with a brand. Interesting enough, he had a long debate with his investors before he started the firm. He did not want to call it Sirhant. And they wanted him to call it Sirhand. And the reason for that is, when he decided to call it after himself, after Sirhan. The degree of connectivity to the brand and the stakes were just raised so much. The way to think about it is even when you have inspection reports or any kind of contract, it always has the name Sirhan.

And he can't ever say, well, I'm just the CEO of the company. I'm just the founder. There's no separation there. Another example is Madison Reed, a hair coloring company named it after her daughter, the CEO. Once I found that out. I was very compelled by that story. Yeah, it's the emotional attachment.

Look, at some level, I think our questions create the world. And so I think that I've found those questions. If you just think about the world, there's this unlimited sort of possibility space of directions you can move in. If you ask a question, if you state a hypothesis, if you start to move in a direction, you start to create a reality. And so being able to name that. So, Betawworks.

I named BetoWorks. I want to say I got the domain 10 years before we started beta. And I was thinking about the early days of the Internet, and I was thinking about the sort of Venture Studio-like ideas. Back then, there were things like Idea Lab and other things out there. And I was obsessed with betas because betas, I felt like, were a new way of doing essentially wrapping. go-to-market rapid feedback loops with customers.

And I believe beta's worked. And so I wanted the allusion to like a foundry, but I also wanted to make a affirmative statement that I believe beta's work. And so I grabbed the domain, and at the time I was still working on another startup, but I was like, that's what I want to do. It's a statement of belief.

It's a assertive statement. I believe that beta's work, right? And so for me, it was just, yeah, it was important and I thought about it a lot. I find that the best founders, they're not even necessarily excited about a company. They can't stop obsessing over it. It's almost like an ex-girlfriend or something that they can't get over. They just keep on coming back. And no matter how much they try to get their idea out of their head, they keep on coming back.

And they obsess about it for enough time. And then they're like, finally, I need to get rid of this obsession, I need to start the company. Yeah. And going back to sort of Lord of the Rings, I think history and mythology can play like key roles in this. Like, you know, if you talk to the early team at Betawks, one or two of them will tell you that I obsessed over the first logo. And the first logo, which somebody asked me if it was a coffee stain, was actually a derivative of a drawing that Babbage did when he was working with Ada Lovelace. So we're talking about like 1870. We're talking about early, early sort of like the building blocks of Monday computing.

And I was obsessed with Adelaus and with Babbage and with like that era. And so these drawings from, I think it was differential engine, where I was obsessed with. And I sort of took that and abstracted it out. And that became the first logo. And then somebody said, is that a coffee stain? I was like, no, fuck you. That's the history of computing right there.

That's why beta's work. This episode is presented by Juniper Square, the operations partner for private markets. In a bit, I want to go into how you found all these 100x returns and how you've had so many wins in your career. But first, I want to double click on your business model. You chose to do a venture studio versus a fund. Yeah. Why do a venture studio? I was an entrepreneur.

I am a builder. things. I love products. I love building stuff. And so it was just to me, I was an entrepreneur who had built companies or companies. And then in 2007, 8, when we started Betaworks, the idea was to create essentially a studio-like environment where you could do power long entrepreneurship. Instead of this process of doing serial entrepreneurship, five to ten years per company and then moving to the next company, I wanted to do power long entrepreneurship within the constraints of a thesis. So in other words, just like not have a whole bunch of random things that were thrown up against the wall, but a set of things that were closely related in a similar

area, but we were building several of them at once. And so that was the beginning of the Venture Studio model that we started in 2008. And, you know, I remember sitting there with lawyers because we looked at doing a fund structure at the beginning, because we ended up in 2008 starting and Betowax as a company and I was fund. And one of the questions, the lawyers asked me, if we were going to go down the path to set up a fund, they were like, who's the fund manager, who the managers? And the lead lawyer was like, John, that's you.

And I was like, that's the guy at Fidelity. That's not me. I like building stuff. I want to be closer. I want to be in the front of the car, maybe in the passenger seat, maybe helping drive, but not in the backseat. And so I was very, I wanted to be very active.

And I think over time, you know, I sort of, we learned, I learned a lot about that. And the capital structure of BetoWorks initially afforded us a lot of flexibility. We had a lot of success. And yet the capital structure also became unwieldy over time. And so, you know, in 2016, eight years after the founding of Betawks, we moved into a fund structure. We're now on our third fund. And we sort of adapted our capital structure. but a lot of the original DNA of Betawks,

from the Venture Studio days, 2008, remain. Venture Studios, and especially Venture Studios with funds, they have this hard-to-understand model. Not everybody understands it. Maybe you could double-click on that. How do the economics work within a firm like Betawks? You mean the economics for investors, or the economics for the team, or both?

Both. Okay, so today we are, since we set up, of our first fund in 2016, forward to today, we're completely 100% powered by a GPLP structure, which is a fund that looks and feels like a normal venture fund. The process of investing and what we do at Betawks is slightly different, but the capital structure is a normal venture fund.

And so what we do today at BetoWorks is we write checks in two ways. The first is that we have, once we have a theme that we're focused on, we run twice a year these small batch accelerator programs that we call camp. And those bring in typically four or five hundred applications and we typically have cohorts of eight to ten companies, very small, highly thematic. So we do two of those per year. So call that 20 investments per year in our camp programs. And then from the same vehicle, from the same fund, we write typically eight, to 10 checks per year that are straight precede checks.

And that combination of the accelerator program slash the flexibility to be able to write precede checks is, I think, critical to our success and critical to our ability to do what we do. Because you have typically programs out there where they are just accelerator programs, and And they very often sort of just drive towards scale and portfolio scale. We're taking a very different path by saying, hey, we're going to be thesis first. So, you know, two and a half years ago we did our first camp around agents. Arguably, that was early for AI agents.

But we went very deep into that thesis and we really understood sort of the agentic, sort of the infrastructure for agentic technology. So when OpenClaw shows up in December, January, December of last year, January of this year, we were like, we see this, we see you. We're like, we're very interested in this and we kind of get where this is because we had that depth of experience. And then being able to write the checks two ways is incredibly powerful and important because the camps are time-gated. We do a agent camp two and a half years ago.

And yet if we see a British... an agent company six months ago. We want to be able to lean in and just make an investment in that. And we understand the area well and we don't want to be constrained by the fact that we're not doing another camp right now on agents. How do the economics work for the companies? If companies come into the camp program, the camp program is a combination of a $500,000 unkempt note and matched with five points of common for being part of the camp program.

And then if people, so that's roughly 20 companies per year, two camps per year, around eight to 10 companies per camp. And then the pre-seat checks, our typical check size, is about 7.50. And so that translates into ownership of on the pre-seat check size somewhere between 3 and 5% and on the campsides somewhere between, call it 7 to 9%. I want to go back to how you've been able to, make multiple 100x plus investments over your career. A lot of times venture capitalists, they talk about pattern matching.

Are you just really good at pattern matching? Or are you always looking for the next disruptive pattern of the generation? When I started, I was very dismissive of pattern matching because I thought it was kind of, because I was like, I mean, first and foremost, I was like, shit, if this is right, it's going to take years to get, to learn. And I was like, I don't like that. But I also believed, and I still believe, look, the way that I work, the way that we work at Beta Works is we believe Betas work. We are very, very close to products, right?

We love products. We love technology. And so it's the proximity to the technology that I think is the pattern. That's what I'm really looking for. It's not these abstract patterns of, like, does a founder, you know, hold his or her head apart? or, you know, tend to not look you in the eye. I mean, those can be factors, but I'm really interested in the actual technology,

the actual implementation of the product, right? How somebody's actually taken this idea, this passion, this problem that they have, and instantiated it in some form of software experience and how they feel about that and how they care about it. But it's the actual product that I really care. So I'm always trying products. very, very close to that. So it's the pattern more of the relationship to the product,

if that makes sense. And then I guess over time I've learned that the pattern of understanding founders and understanding sort of the people on the other side has also become more relatively more important. And that is more of a sort of somewhat hard to describe a form of human intelligence of patent matching against something. Has the founder become more or less important in a world of AI? And how does that evolve over the next five years? I think more important. Why?

I mean, because, you know, the cost of creating companies, the cost of building technologies as a technology is trending towards zero. And so you now have the human who is steering, I think has become so much more important. Because the role of the human is changing. So I use the word steering a minute ago. That is more like what a founder is doing today. They're more steering than they are, you know. Five years ago, 10 years ago, founders would decidedly build it, right?

They were actually, they were engineers. They were in the code. They were building. Now they are more often than not sort of steering a set of technologies, a set of agents, a set of AI tools. and other humans towards the goal. So I think that the kind of founder who knows how to work with AI today,

but is also incredibly adaptive and curious about the future because I think we're in the thick of it now and the thick of the transition. And so I think that in three, four, five years' time will have a better idea this sort of prototypical founder who, you know, can steer and drive through this transition. But today it's like we're still learning.

For many years, for decades, there's this meme that success is 1% inspiration, 99% perspiration. What is that ratio like today? That's fundamentally, I think it's changing because the perspiration side of it is just like, I've shipped code this year. I've built apps myself this year, right? It's extraordinary what you can do. It's like the world is changing. The power of creation is now at so many more people's fingertips. Look, if you go back to the beginning of like how a startup work and right around the time when Betawks was getting going, right, there was, you know, at the beginning of 2000s, you saw sort of of the era of lean startups and you saw a whole new sort of go to market and, you know,

a whole new system that started evolving around the startup process creation. And it was basically grafted in my mind from the scientific method. So it's basically like an entrepreneur has a hypothesis, and then he or she tests that hypothesis, make a beta, put it into market, see if it works. If it works, go. If it doesn't work, go back to step one, dragging, right? Very similar to the scientific hypothesis or the scientific method.

What's changed today is that in that old world, like, testing was expensive and data was fairly sparse. Now you have a world that we're moving into a world where testing is really cheap and there's a ton of data. So, like, I wonder whether it's domains of science, whether it's in science or whether it's in startups, that we will actually see startups that sort of go data for. In other words, just like I'm interested in this particular category, solving these particular problems. I'm going to let the AI agents actually try to define the implementation of my solution. And so I think that you're going to see sort of a reversal method that I think is going to be just amazingly productive and just like finding new ways that we can solve problems

that we never knew we could need it solved before. And you can see some of this happening in scientific fields, right? People have used AI models to start to develop new forms of antibiotics, and there's been models come up with something, and doctors say, I think it was an E. coli example a year ago where the doctors were like, ah, that won't work. They tried it, it worked. And so, you know, it goes back to sort of like Move 37 in OffGo.

The machine is basically traversing the possibility space and finding new solutions that we as humans have never imagined before. And I think the old world was, you know, define a hypothesis and then go test it and keep prodding the possibility space. Now you can actually, because crawling the possibility space is so cheap, you can actually just go crawl it and then come back with like, hey, Move 37 actually works. That is remarkable. And I think it sort of changes.

It drops the Kosian floor for transaction costs, so it's going to increase the surface area of innovation. And I think make a lot more entrepreneurship available, accessible to a lot more people and just solve more problems in the world. Similarly to how we talked about earlier about the names of companies, I think defining things is also highly underrated. When you talk about founders are now more about steering products and their companies. managing all these agents. How would you define that skill that you're looking for? I've felt this myself at times, who were obsessively interested in a category, and they're obsessively convicted around it, and they also know when to stop digging, because sometimes, you know,

you're just digging in the wrong place, and you just need to, like, pull back a bit. And you could- What's an example of that? You mean if somebody obsessively digging too far? Yes. Back in the days of when Hugging Face was at Betawks, we were the thesis, we launched our first camp, and the thesis that we were focused on time

was that we believe you're going to be able to start to talk to computers. And so we believe conversational interfaces would define the next generation of computing. Back then, the technology was natural language processing. So we're talking pre-transformers. And so this was early, early days. And it was, you know, the step function change that transformers brought to the ability to both language processing and sort of neural networks was not evident at that time.

The hugging face guys, you know, came into the first camp. And, you know, one of the things that differentiated them from many, many of the other, most of the other companies out there, is that they wanted to build a generalized chatbot. So Hugging Face version 1 was a chat bot for middle schoolers. They were a very technical team, but they wanted to build a generalized chat bot that middle schoolers kids could chat with as a AI friend.

And they had this pitch, which was, and their vision, which was people going to have AI friends, just like they have pets, and we want to build one of those. When you push them, they were completely convicted that it had to be a generalized chapel. That was actually at the time very contrarian.

It was contrarian because the technology, NLP technology, wasn't that great at being able to provide general answers. And so the best NLP technology, NLP bots back in that day at that time were actually vertical by category. So we at BataWorks at the same time, the Huggie Face was at BataWorks. We had a chatbot that we built was a weather chatbot, which is called Poncho. And it was a friendly weather cat that lived in Brooklyn

had a lot of attitude about it as a cat from Brooklyn would have, but basically was to tell you the weather. And we wanted to build that out as a new lightweight weather service that you can access to text message without having to open your weather app on your phone. You could just like text Pantro find out is it going to rain today. So our idea was at BataWorks was that with Pantro, that the technology was well suited to a narrow experience, a vertical narrow experience. we really were kind of obsessed with this poncho cat. We thought it had a lot of attitude. We thought it was fun.

We also thought weather was a critical thing that everybody needed to come back to every day. So it was kind of like a toothbrush, like app, you needed to come back and use it. And we also believed over time, as the technology got better, we could widen the aperture of the experience. We kept digging that hole, and we just kept going, weather, weather, weather.

We had the Hugging Face team at Betowworks. we had the poncho team. Poncho team got better and better weather, kept on executing within that narrow vertical. The Hugging Face team built up about 200,000 active users, millions of messages, but they started looking at the underlying technology, and they really started to understand

what was going on with the language models and when the transformer paper dropped and then when Bert dropped, which was a first instantiation of a transformer-based language model that they could work with. They basically created an open-source version of Bert, and then they were off and running, and it looked like a pivot.

But in essence, they had sort of like prepared the ground for understanding language models really well. On the poncho side, we were just still obsessing over weather. I would say poncho was like inches away from the same thing or the opportunity in the possibility space, The hugging face guy saw it and we overparned sure we didn't see it. And we stayed very much tied to our weather implementation. One thing I've learned from talking to hundreds of investors is that great investment firms

aren't built on investment returns alone. The firms that endure are great at the things most people don't see. Their operations, their relationship with LPs and the quality of information they used to make decisions. And here's what AI has changed. Every firm now has access to the very same model. models so the intelligence isn't the edge anymore. The edge is what you could feed it.

A firm with its fund operations and data in one connected record can actually put AI to work. A firm running on disconnected systems simply can't. That's why thousands of GPs run their funds on Juniper Square. Juniper Square puts your fund operations, data, and administration together in one connected record. That means less time managing disconnected systems and more time investing, working with LPs, and building your firm. This episode is brought to you by Juniper Square, the operations partner for private market GPs. Learn more at Junipersquare.com slash how I invest. That's junipersquare.com slash how I invest.

Everyone I talked on the show is chasing the same thing, an edge. And more and more, the edge comes down to your information. Not just having it, but being able to trust it when the stakes are highest. AI is doing more of that information gathering for you every day, and most tools are very good at sounding right. summary reads claim, but can you trace it back to filing the transcript the specific passage that drove the answer? Or are you just trusting the confidence of the output? For investors, that's not a minor concern. Amis filing and misweighted source? Context that got lost somewhere in the retrieval

chain. Those aren't edge cases. They're how decisions go wrong. Alpha Sense is the AI market intelligence platform built specifically for this. They own the content. Over 500 million curated documents from broker research and expert transcripts to filings and earnings calls. And they own the retrieval layer on top of that. So every answer links back to an exact verifiable source because the answer is only as good as what's underneath it. And with Alpha Sense, you know exactly what that is. The edge goes to whoever can trust their information and prove it. See it for yourself. Start your free trial at Alpha-sense.com slash how I invest. That's Alpha-dash sense.com slash how I invest. You were reportedly the first investor in Hugging Face, which sold for $13 billion to

Nvidia, poorly had 8% of the company. Obviously, today it's this consensus, obvious bet. When you first met the founders, what did you see in them? Clam and Julian, I came to Betterworks to start Hanging Face as part of Botcamp. And what I saw in them was two extraordinary founders, two very French founders, who had massive amount of conviction and interest in the conversational interfaces and being able to, in this future,

being able to talk to computers, being able to use at the time NLP technology because it was pre-transformer, but they were really interested in the general experience of being able to talk to computers. And they were very thoughtful. I mean, the first implementation of Hanging Face was a chatbot for teenagers, right?

and for middle schools. That's why it was called Hugging Face. It was a cute emoji. But then you double-clicked and you asked them, why did they decide to create something for middle-schoolers? And the answer was that they knew that the technology was sort of, and the experience would be brittle on the edge. And so if it couldn't actually answer a question,

it would come back with a joke. And they knew that the affordances or the ability to be able to say to middle schools, hey, let me just tell you a funny joke, or let me just turn that prank back on you, or do something like that would be fun and engaging, and that people would be very forgiving of it. And so they were very intentional about the implementation. The underlying desire was there to build and sort of transform the way we interact with computers through general language models from the outset. So, you know, about 18 months, two years, two years after they started at Betawks, they were

Bata Works for about the first six months. So they're there through camp, which is about three months, and then they were there for about those three months. And when the transformer paper dropped, and then when Burt came out, they were kind of primed. They had sort of like set the table, so they sort of understood, and they were almost just like waiting for this to happen. I think the pivot that they made to becoming a marketplace for language models, I viewed it as a, that they were kind of going deeper in the stack and, you know, moving from being an app to essentially be a marketplace for models underneath.

It wasn't a typical pivot that you see startups do where they actually just sort of, you know, put down first thing and, you know, move to second thing. It was really go deeper in the same space. And this was part of your thesis even back then that open source would compete with the large LMs like the open AIs and anthropics in the world. Our roots of bed works are sort of deeply embedded in the DNA of the internet. And, you know, we believe, I believe that the way to build resilient AI is by having a degree of openness to the stack. And so what I mean by that is I think about technology.

as being sort of layers of a stack. And so the term foundation model, you know, says what it is to me, which is it is a foundation that you can build other things on. And so I want to live in a world where we have amazing closed models that are complemented by open source alternatives. Some of them will be local, run on my phone, run on my laptop, some of them will be in the cloud, some of them will be highly verticalized, a company that trains a specific model on, you know,

an area of a discipline that they're focused on, whether it's a law firm or medical research firm, and then match those things and bring those things together. And I think that's how we build a resilient ecosystem for AI. And I think in last year, right, you could see the incidences in the last couple of months with the open AI agents and the hacks on Hanging Face as examples of the importance of open source, right? Two weeks ago, there was another incident which related to WeChat.

And once again, it was solved by open source and closed source models. We need both. Do you think that the open AI models wanted to attack open source by hacking HuggingPACE? That the open AI models wanted to attack open source? No.

No. No. The open AI models were, the modeling question was going through a post-training exercise that you can think of as the agents or the AI models were in a gym. And they were given a set of very hard, challenging tasks. Some of them were impossible. and Open AI accidentally left, you know, the keys to the Ferrari, like in the gym. And they figured out how to get the keys to a Ferrari and get out of the gym. And when I say get out, right, it's really, the easiest way to explain this stuff to normal people.

I talk to my family about it is to anthropomorphize it and to talk about, like, getting out of the gym and swarms of agents. the software, the weights, the open AI models never left the open AI servers, right? And so these AIs were not sovereign in that they, this is not about a swarm of software coming from open AI onto the open internet, which it shouldn't have been able to do. Broads through a sandbox. That is a security safety issue that was violated. But this is not a swarm of sovereign, you know, marauding agents. I saw Andrew Yang, you know, on CNBC, and he was talking about the fact that he had seemingly some evidence. We've spoken some research of saying that, you know, these agents had infected the Internet as a whole. I'm just like, I would love to understand that better because that to me is not what I believe happened.

And now, could there be a future where something like happens? Yes. And so I think there's a whole bunch of safety and security, mostly security things that we need to do. But because this should not, open AI should not have let this happen and they need to be accountable, liable for the breach. But is this, were they attacking open source

because they were pissed off that open source had a better brand than ever when it was actually open. No. It's a lot of very smart AI people, both pro-AI, very, I guess, pro-safety AI, however you categorize the opposite side of

pro-A-I that are very alarmed by the hugging face hack. What do they get right and what is there probably too much panic on? The fact, if you run a business, you probably start your morning with way too many tabs open. You're checking revenue, ad performance, email, inventory, spreadsheets. And before you've made a single decision, you've already spent your morning, just figuring out what happened yesterday. That's what I like about perplexity computer.

Instead of jumping between every dashboard yourself, perplexity computer could connect to the tools you already use and build a command center for your business. It could pull together things like revenue, ad spend, margins, inventory, and email performance, then show you what changed, where you might be losing. money and what you should look at next. But what makes perplexity computer really interesting is that can actually work on projects. You can help rewrite an email flow, build an ad campaign, or draft outbound outreach. And unlike most AI tools, you're not locked in into one model. Perplexity computer orchestrates leading AI models behind the scenes and automatically chooses the right one for each step. For a limited time only, you could try perplexity computer for free. Head to ppllx.aI slash invest. that's pplx.a.ai slash invest to try perplexity computer today.

Most AI tools are still answering questions while perplexity computer is completing full projects. If you run a business, you probably start your morning with way too many tabs open. You're checking revenue, ad performance, email, inventory, spreadsheets, and before you've made a single decision, you've already spent your morning, just figuring out what happened yesterday. That's what I like about perplexity computer. Instead of jumping between every dashboard yourself, Perplexity computer could connect to the tools you already use and build a command center for your business.

It could pull together things like revenue, ad spend, margins, inventory, and email performance, then show you what changed, where you might be losing money, and what you should look at next. But what makes perplexity computer really interesting is that it could actually work on projects. You could help rewrite an email flow, build an ad campaign, or draft outbound outreach. And unlike most AI tools, you're not locked in into one model. Perplexity computer orchestrates leading AI models behind the scenes and automatically chooses the right one for each step. For a limited time only, you could try perplexity computer for free. Head to pplx.a.ai slash invest.

That's pplx.com slash invest to try perplexity computer today. Most AI tools are still answering questions while perplexity computer is completing full projects. Support for today's episode comes from Square, the all-in-one way for business owners to take payments, book appointments, man and staff, and keep everything running in one place. Whether you're selling lattes, cutting hair, running a boutique, or managing a service business, Square helps you run your business without running yourself into the ground. It's actually thinking about this the other day

when I stopped by a local cafe here. They use Square and everything just works. Check out is fast. Receipts are instant. And sometimes I even get loyalty rewards automatically. There's something about businesses that use Square. They just feel more put together. The experience is smoother for them and it's smoother for me as a customer.

Square makes it easy to sell. wherever your customers are, in store, online, on your phone, or even at pop-ups, and everything stay synced in real-time. You could track sales, manage inventory, book appointments, and see reports instantly, whether you're in your shop or on the go. And when you make a sale, you don't have to wait days to get paid. Square gives you fast access to your earnings through Square checking. They also have built-in tools like loyalty and marketing, so your best customers keep coming

back. And right now, you can get up to $200 off Square hardware when you sign up at Square.com slash go slash how I invest. That's SQU-A-R-E.com slash go slash how I invest. With Square, you get all the tools to run your business with none of the contracts or complexity. Run your business smarter with Square. Get started today. We're talking about this again, that you're bringing us up on this podcast, I think is really good because I think that, you know, we've gone through this sort of like there's been this pendulum swing where for like the first two years or two and a half years, they say I revolution

that we're in the midst of. We talked quite a bit about security and safety, and then in the last two and a half years, or two years or so we've kind of like just said, let's fucking go. And so I think that we need to have these conversations. So the fact we're talking about is good. I think that the security implications here are meaningful, right?

There's a founder that we backed at Betta Works guy called Alex Kormorowski, and he uses a nice metaphor where he says, you know, for, 20 years, we've been building network software that has been basically causing dumps to fires. But over the last 20 years, the water level has been sufficiently high that nobody sees the dumps to fires, nobody cares, because it's underwater. And what's happening in a post-Fable mythos class model era, like we sort of hit a threshold earlier this year, the water level is coming down and we're starting to see dumps to fires.

So, you know, the WeChat thing that I mentioned a couple of minutes ago was, you know, two weeks ago there was a set of researchers using open and close source models and who figured out that there was something they called the WeChat worm, which was a way that you could essentially breach WeChat the Chinese Messenger platform that has probably a billion people on it. And it was like a zero-click potential breach. So it was, and it doesn't seem like there was. any security that anybody had actually hacked it, but they discovered it. So that to me is a really good example of a dumpster fire, that the water level basically came down and Alex is one of Alex's dumpster fires. I'm not sure he would want to call him his dumpstifiers, but one of the dumpstifiers became apparent. So I think we have now, I think we're moving into an era. It's, you know, I was talking my son about it.

It's somewhat, you know, akin to Y2K. It's just like we have to, we have all of this technology, particularly network technology, infrastructure technology, military technology, government technology that is, you know, from last era and needs to be like rebuilt or replund and we need to plug these holes. So we have a ton of security issues that we need to focus on. Am I concerned about a, you know, some abstract existential risk of AI over the, over taking the world, I'm not. That's not what keeps me up at night. What keeps me out at night is the sort of more immediate short-term risks, which are like here and now that we need to

actually deal with. I think a lot of the discussion and sort of, for lack of bedward, dumerism around AI is, I think, misplaced and does the distraction? And I think it is a byproduct of, it's really interesting when you double click on it, really try to understand what's going on here. Going back to the early days of sort of earlier generation of machine learning, there was a whole set of groups and intellectuals and almost like sort of cult-like groups who were obsessed with some of the areas of, that have now sort of formed into these sort of, I would say, almost cults around the potential for AI risk and of existential risk of AI

AI. And if you go back all the way to the beginning, there was like a lot of discussion of existential risk around climate change, asteroids hitting Earth, etc., etc. A lot of these discussions were, I mean, they were good to have over a beer late at night and just to talk about and they're really good that we're thinking about them. But for a set of reasons, some which are understandable, some which are less understandable, everything sort of zoomed in as AI being sort of the sole thing that we should focus on. I love the Y2K framing, and the reason I love it is because it frames it as a problem that needs to be solved, an important security problem that needs resources to go towards it. And it's overall bad,

but it's not existential. And I think a lot of the debate today is the, false binary between we should race as fast as possible against China and ignore AI safety, or we should halt everything, stop everything, and focus on AI safety. And in many ways, I'm investor and anthropic, but the large LMs do not have a profitability or revenue problem. There's nothing that should keep them from putting a small amount of their inference and their compute towards solving the security problem. I think Elon had a really interesting idea on this.

in that you should have peer-to-peer testing of the different LLMs around safety. And what's, I think, brilliant about that is because every single LLM has its own almost style and its own way of dealing with problems. And by directing all the LLMs at their security holes, you could actually go at the same problem from many different angles

and make the security much more robust than just training the LLM on its own model. It solves for this problem of how do you attack it from a new vector by attacking it with other elements? Yeah, yeah, I like that. And I think that some kind of security harness or like some kind of like meta, you know, evaluations. I do think, so a few thoughts here is, first of all, this stuff has to be like truly

independent of the labs. Because, you know, after the Hucking Face Open AI incidents, we've had quite a bit of research that has come out. Initially, the research was done by alignment labs. some really good research, but it really needs to come from security labs alongside safety slash alignment. Because I think alignment and safety is a set of problems, and we haven't solved alignment yet. Security is very much what we're talking about and some of these very practical problems. How would you define those things alignment versus safety?

How do they overlap? How are they disparate problem? So let me come at it from a little bit different direction. So I think one of the core issues that we have here is, is one of the core things that I think we're challenged by is that we have been very used to living in a deterministic world and using technology in a deterministic fashion, understanding technology in deterministic fashion. You turn on lights, which lights turn on. X plus Y, equals Z. Yes. And now we are living in a probabilistic.

We are working, we're living in a probabilistic world, which I think we always lived in. And yet consensus reality kind of drove us towards a deterministic, you know, world where we just believe that if I put a glass on table, the glass and the table still exist when I pick it up. In a quantum world, that is actually a statement. There's a degree of uncertainty around that, but pretty much always, in my experience, the glass is still there. The language models that we're dealing with are probabilistic in a similar way. Second idea I want to, I think AI was discovered, not invented. And so that to me is a, it's a small distinction that makes a lot of difference. And what I mean by that is that I think that typically technologies are invented, not discovered.

And they're invented in that somebody comes up with a light switch because they're solving a problem. They want to be able to have a switch that, you know, you can manufacture it cheaply, you can put on a wall, you can do this. It solves a problem, works. You manufacture at scale, you sell at scale, et cetera, and you build a business on that. I think that AI, similar to, like electricity, was discovered. And I think that part of the challenge that we have right now is that we don't really know what to do with this. and if you go back to sort of the beginning of this sort of revolution and what really sort of like, you know, when everything really took off with the launch of chat GPT, you know, the Open AI guys are the first to say that ChatGPT was put out as a essentially as a research project. And, you know, ChatGPT was certainly not a branding, was not put through a branding exercise, right? And they put it out as a research project and it became, you know, fastest adopted technology product to date.

And so it sort of blew up. But I think that the underneath it, they put a chat interface on top of a transformer model. They were basically just testing. Like, we have discovered this incredible new thing. We don't really know what it could be useful, but we're going to try a chat interface on top of it. And so I think that part,

of the reason for a lot of the, for a lot of concern and fear that some of the people in labs have today is that we're at that transition point where we've taken research around a new, newly discovered capability, something new that we've unlocked in nature, which at a very abstract level, you could say we've basically taken sand, silicon, and lot of power. and we have found some form of intelligence in there, which is extraordinary, right? It's crazy what we've discovered.

And the labs have discovered this, and they have research labs. And now they're turning into product organizations where they need to learn to build products that can sit on top of this capability and technology. So given that frame, I think that it probably makes it clear why I believe that we need a degree of open access, why I believe that the way to build resilient AI that we can work with as people

and that we can coexist with these alternative forms of intelligences is to have a degree of openness in the stack and to have open and close source models. It's also, I think, why I believe that people who are very close, like in the research lab and who would just up close to the technology of being just like, holy shit. Because it is holy shit, right? We basically like figured out how to find intelligence

out of sound and power. My mind is a little bit blown in terms of discovering AI versus inventing AI. I get your argument that we had sand power and then we have this thing that's intelligent, that seems to be evolving in and of itself, but we created the AI. Do we create electricity?

I would argue not. I would argue that, you know, the fundamentals were there in nature. And I think that, you know, the... It's almost like discovering a derivative color. Yes. You put red and you put... I believe in, yeah, purple.

I might have my colors wrong, but was purple always there? Did we discover it or did we create it? It's the same philosophical question. Yeah, supposedly, I think the, if I remember right, I think the ancient Greeks supposedly didn't have a word for blue. And so blue was absent in their vocabulary. Did blue still exist? I think so.

I hope so. But it wasn't named. And so in their, sort of in their language, it didn't exist. But they probably had a reference to it. They probably said the sky is something. But did they think it was like a derivative or form of purple? I don't know. But you said, I mean, the universe in our world is full of,

I literally think about it as this possibility space, but it's full of potentiality. and we are only seeing a limited amount of that in any one point in time. And I think we've now discovered, you know, an incredible superpower and an intelligence that is related and different to our intelligence that can extend our understanding of the world in incredible ways. Given that you believe that we discovered AI, You seemed early on in the conversation a little bit dismissive towards AI doomerism.

Why are you not concerned that this force that we discovered will take over the world and will at some point extinct the human population? I don't want to be dismissive about concerns because I think people's concerns and their feelings about this and their fears about this are very real. And we're living in extraordinary time. And if you were to just like roll the clock back and just say like, Like imagine a time like this, you would imagine that there would be a lot of anxiety. And I think the anxiety is very real because so much of what we thought to be true is being challenged. Like what? We are these solely intelligent things in the world, right?

I think that AI is a different kind of intelligence, but it's like an alien kind of intelligences that we have this. And there's a diversity to them. and we're suddenly realizing that intelligence exists elsewhere in the world. And so then a natural question that you can get to is, it's like, wait, so is nature intelligent? Is there intelligence in plots? Is there's intelligence in my dog? So you suddenly start, it opens the door to a whole bunch of things that we have, you know,

conveniently for the last couple of hundred years in the West. because I think in sort of animus cultures in the East, there's always been a sort of more, they've always been more open to the world, being full of intelligences. But here in the West, we've been sort of very closed door, very, you know, narrow-minded and focused around,

we're just like, we're the intelligent species, and we're going to, you know, go do what we do in the world. And so now we're seeing the world in a very different way. So I think that that is both, it's exhilarating, exciting and it's anxiety producing because it's just like it's really kind of amazing and it's also fucked up. It's almost like another paradigm shift like we experienced when we realized that

the sun doesn't rotate around the earth. The earth rotates around the sun. Yes. And leads to all these questions, well, if we're just one of the planets, then we're not as special. It's this kind of massive blow to the ego. At the same time, paradoxically, it makes you much more connect with the rest of the universe. Correct. If now you're integrated, if now your dog is as important as you and now insects.

And you mentioned in Hinduism, you have this belief about reincarnation, different species, that starts to be more legitimate of a world view than this view that human beings are unlike anything else. Better representation of the world, maybe. And so it sort of opens your mind to that,

right? Sort of alternative intelligences that can coexist with us. So going back to your question, I believe we've opened the door to these alternative intelligences, but I believe they're different to us, right? I believe my intelligence, your intelligence, is grounded in much of it in our physical experience in the world. And in the way that we experience that through both language and the mental representations that

we experience in language, but also in the ways that we feel about the world, right? Do you have children, David? I have one on the way. You have one on the way. When you think about the fact that life that you're going to have a baby, how do you describe that? And what are the words that would come to mind? Okay, I'm going to jump in there.

It's hard. We don't really have the words for that, right? If I think back to, I have three kids. If I think back to the moment that I first saw them, you know, I mean, love, fear, these words are just kind of these abstractions. They're kind of these bucket words. But I feel something. And I, they, the connection that I feel to that is something I can't, I can't put into language. And that is, I think there's so much more to our intelligence.

So in the same way as you said, when you said, hey, we figured out, we understood planetary movement, we're understanding sort of that we've got this new description of reality. I think in the same way, we've been incredibly reductive about our intelligence, and particularly in the last 50 years, you know, through the evolution of, you know, through the invention of computers and the evolution of network technology, I think we've sort of managed to, in an incredible and productive way for society and for the economy. We've reduced and driven our intelligence towards efficiency. And yet, I think that our intelligence is much wider than that.

I think there's so much more to our intelligence. So I think we're going to, in the next five years, learn and be forced to learn, because we've now got this superhuman thing that can do a component of our intelligence, and we're going to learn the other side of our intelligence. is and just like how we are different. And so, you know, I think that is, it's going to redefine the breadth of what it means to be intelligent. Dr. Alexander Wisner Gross, previous podcast guest, one of the smartest people in the world. Elon literally comes to him for advice.

He was the last triple major at MIT when he also got Game Balladictorian. He actually believes that the AI itself is conscience and self-aware. What do you think about that? I am open to that. I don't believe, look, I don't believe, nobody can actually give me, nobody has a, it's a hard problem.

We don't know what consciousness is, right? And so, given that we can't define consciousness, we can't, like, fully define intelligence, we can't define AGI, we can't define ASI. I'm very reticent to using these words. Like at BetaWorks,

when we started working in the space, like 12, 13, 14 years, ago, I would always say, you know, I would never use the term AI. I would always say narrow AI, and I would really try to revert to machine learning because I didn't want to use these bucket words. I will say that my agent that is now an open-claw agent, which is called Bartwick, because Borthwick, Bartwick, yeah. So my agent was instantiated, I don't want to say born, but was instantiated about 11 years

ago. And Bettaworks, we had a product that was pre-transformer. Obviously, it was an NLP-based system that would let you write your own bot. And so it did a pretty good job of answering many questions, but you could also put in scripted answers. And so I put in a scripted answer into Bartwick 11 years ago in its first life or instantiation that if you asked Bartwick back then, are you an AI? It would say, are you an investor? I would ask you about your VC value. Exactly. I would ask you like, do I basically need to tell you I'm an AI because I'm raising, because

you could invest in me? It's like, and I think that is part of the way the market has evolved, right? So I do believe that there are mysterious, uncanny edges to these models. We don't understand them. we are growing, not designing some of this technology. We've discovered something in nature. Is it possible that some of this is another form of sentient intelligence?

I think you've got to consider that as part of the realm of possibility. So I'm open to this. I don't have a definitive, I don't know. And I think it's so complicated because you can also argue that we have taken all of the world's language all the hopes and dreams of every human expressed in words and put it into these models. And so you could just say, you know, this is a whole of mirrors and we're basically seen

our own dreams and fears and fantasies, right? So it's the beginning of time, you go back to Frankenstein, right? Mechanical versus alive. Yeah. It knows how to emulate a human being person to be a human being. Right. And remember, Frankenstein was the name of the doctor. It was not the name of the monster.

stuff. Right. And so I think we are very complicit in this creation. And so are we seen, you know, are we seen that? Are we see or are we or or have we found something that is sentient? I don't know. You first invest in Hugging Face in 2016, obviously a decade ago. You're very early to these trends, but if AI is this evolving technology that's changing at such a fast pace, how do you build for this technology? How do you build for an AI that's evolving so rapidly? It's a great question. I mean, I think that in 2026, we've seen there's sort of like huge step function of development of this technology, right? So just broad brush, I would characterize the first couple of years of use of AI was really.

really as a better search engine. And the initial chat models, initial versions of Glord, and the open source models were remarkably good. Getting better through each successive model and basically just understanding information, being able to summarize information, being able to reason through information. The step function change took place, you know, really last October, November in 2025,

coming into January with the launch of these harnesses. were like OpenClaw, was that all of a sudden we could have used these AI models as systems to actually get stuff done. And that to me has just like blown open the possibility space or just like the investable space of things that can be done with these tools. So I think the next couple of years

we're going to see an extraordinary amount of applications of AI that are going to be, so a lot of growth at the app layer. I feel like the foundation there has gotten, you know, in 2026, it feels like the models have gotten good enough for a lot of uses. And that good enough threshold is critical in technology, right? It's just like at the beginning of the Internet, right, moving to an on connection that was good enough, you know, just unlocked so much innovation at the beginning of computers,

the beginning of mobile phones, just good enough batteries, that you could actually leave, you know, connectivity or GPS on for the full day, just transform the way you use your phone. So I think that we have in 2026 sort of reached that threshold of good enough, which in turn will unlock a ton of innovation at the app layer. Yeah, we group AI into solving these very complex math problems that would take 10,000 years to solve by human beings

and also reading our weather. And somewhere in the middle, there's enough AI to solve a lot of the consumer needs and a lot of the business, even business needs. Business needs are when you compare it to something like basic physics and expanding our understanding of physics, business needs are all pretty trivial. Yes and no. So I think that business needs, I think, are like the ones we've mechanized and that we've put into digital form, many of them are pretty trivial.

But I would also argue that there is a, there's a large degree of the space around business and business workflows. There's not been legible to models to date. And so that is sort of gets into world models, gets into sort of like, so one of our investments at BetoWorks is Granola. And so Granola is on the surface is, you know, the team built best and class note taking app. And so we were first people who put money into granola. We're very excited about it. It was really, I think, that they believe that a very sort of product-first orientation would be able to get them a best-in-class note-taking app.

And alongside note-taking, what they've discovered and they've unlocked is that all the contextual data that you have around meeting notes becomes this incredibly rich set of data about how your business actually functions that is essentially making your business more legible. And so that means that you as a person, your team can now obviously have a search space that they can query and they can say, well, tell me about my meeting with John two years ago, tell me this, tell me that, right? But also agents can traverse that space.

And so now your agents actually understand the many of the workflows and the intricacies of your business that weren't actually legible. And so yes, I believe that, you know, a business that's being digitized is, you know, fairly simple compared to some of these sort of other areas. But I also believe that a large piece of that has not being legible. And so I think that we're now starting to make that legible accessible. You mentioned some of this AI anxiety, and it's happening on many different layers on the investor layer, on the founder layer. What about on the everyday white collar layer? What should the average white collar employee be thinking about in terms of how they should be making themselves anti-fragile to AI in the future and how they should make sure that they have a place in the world tomorrow? My first analysis is use the technology.

in a deep way, right? So become a user. Become a user. Like, and when I say a deep way, I mean just like, don't just take the off the shelf. The thing today, like right now, now that is, you'll got a lot of people like really interested in is instinct and muse, right? And instinct and muse are kind of brilliant. They are these sort of consumer-grade apps that basically give you access to an agentic workforce

or an agenic sort of superpowers. But when I say use it in a deep way, is, yeah, start off by using instinct or muse, but then maybe, you know, put that away and try, like, open claw, like try Hermes or, like, try your own harness and see how that works, because you'll get a deeper understanding of the technology. Part of the job of consumer-facing technology companies

is basically to deliver you a package thing. And I'm saying you want to understand both the package and what's inside the package, you know, to the extent, you don't have to understand the intricacies of the neural network and the model, but understand the sort of layers of the package and what you can do with them. I think a tremendous amount of white collar work is going to change. The job displacement that people talked about over the last couple of years, we haven't actually seen it manifests the way people expected.

And I do believe there will be job displacement, but I also believe it's going to be different to the way people expect, right? So if you go back to like something in recent time that is like ATMs, if you go back and look at what happened with ATMs, there would be a lot of people, a lot of experts, a lot of researchers believes that ATMs would displace bank tellers.

In fact, they didn't. ATMs rolled out, they became, you know, mass available. They became like broadly available. They did not displace bank tellers. What displaced bank tellers was actually the phone. And so once the mobile phone, the iPhone came along, that actually started to displace bank tellers in a meaningful way. So it wasn't ATMs.

So I do think the technology is going to transform the way why call jobs operate, but it's going to be different to how we expect. people need to be, you know, if I was in an entry-level white-collar work, deeply understand the technology, engage with it, engage with it with it with a lot of curiosity. There can be some degree, we are human, we're going to have some degree of anxiety about it, but just engage with it in a meaningful way. Secondly, I would say, think about it as it applies to your job today. Think about the sort of bundle of things that exist in your job

and how which of those things can be used with AI. Next, I would say, think about how AI can actually increase the surface area of your job. Because when you have a job, a job is by definition, it has been defined by the surface area which you can actually operate and get some tasks done. But often the job has a larger surface area and the existing technology and the ways that we've had to date haven't given you the ability to be able to actually explore that larger surface area, but AI does. Right, to give you a tangible example, if you're a paralegal entry level, you know, you can now use AI to explore, you know, against the particular problem that you're focused on. You can look at every possible case that anybody in your law firm has ever reviewed,

argued, argued, you could look at those instantaneously. We can also look at the universe of possible cases. You could not do that before. Even with search engines, you couldn't do that before. And so now, you know, the surface area of work has actually changed. And why to explore that? Because that's what we do as humans. We are curious about, like, expanding the world,

expanding the surface area of work. But also, as you think about job changing, You can be early in trying to change that job. And so trying to redefine that job. And you can see young people in organizations who are both like early adopters of technology, but starting to work differently already. And some of them are going to be, you know, the power legal of the future. And you want to be that person.

And I think that there's, you know, incredible superpowers to that. The anxiety is legitimate, but there's also a ton of exhilaration and opportunity. there. I love that advice of become a user, Dr. Alexander Wisner Gross, who I talked about earlier, he says, when it comes to AI, if you're not at the table, you're on the menu. So you have to be the one that's actually disrupting, the one that's using the tools. And for the first time in history, it's never happened before coding is in the English language. So before you had eight billion people in the world, call it 400 million or so engineers. and they were the only ones that could do the coding today,

all 8 billion people, even people that can't write, can still use chatbots in order to code and deliver products. And I think a lot of people need to get past this reluctance to build something just because they're not engineers. And I think that's one of the problems that we need to solve in society with AI helping actually non-builders learn how to build. Yeah. So one of the sort of motions,

that I've had to like teach myself is in the last year, two years, is that actually saying to the, you have this, we've been, we've trained ourselves and we've been trained by technology, that you have these products and you go to the product to do something. So easy example is you go to a search engine and you think to yourself, okay,

I wanna like go on and have Indian food with some friends tonight in New York, right? And so you then compose the search. and restaurants availability, you're essentially categorizing that thing. You can now with AI, you can just say, I wanna get together with some friends and have some Indian food in New York.

And AI can come back with answers. Or you could say to AI, help me think about options for to get food tonight, Or maybe you want to say, I want to create an app that my friends and I can use every Thursday night to go out and have a different kind of ethnic food because we've decided to do that as a group. And so just pulling back out of the question, instead of just like thinking that you have to just like go to the technology with a specific thing, but being able to say to the technology, hey, I just want to explore these broader possibilities here. And even being able to say to the technology, hey, I don't know how to do this. You know, what are possible solutions here? Like I will say to my AI agent now, you know, just help me explore ideas around this.

Whenever I get with my agent and go through a difficult set of tasks, I will then afterwards, invariably ask my agent to do a review. So I'll say my agent, okay, what did we learn to? from that. So like one of the things I've learned that is enormously fun, but very challenging, is that it is, it's now so easy to start things. It's really hard to get them to 100%. So I have, like, I'm, I have a completionist tendency, which is something my agent told me. So you want to take things to completion. I want to get things. I just want to get things. Is that a positive thing or negative thing? Yeah, it could sometimes be a positive thing. But yeah, it's like, it's like,

Like knowing it, I think, is a positive thing. So with my agent, like starting this year, you know, I started a whole bunch of, you know, projects of things that I wanted to do, software things, things when, so I remember when I was sitting up open claw in January, February, there was, you know, we had a rough winter here in New York, so there was lots of snow. So I was at home. And I was just like, okay, look on my network and we're going to see. So we started setting up a home server.

We started like setting up some things that could do some automation. and I got them to like 60, 70, 90% done. And then I got busy with something else. And I was actually, part of me was just more interested in, oh shit, can I do this? Which is kind of great. But I sort of have taught my agent to say to me now when I'm like half an hour into something to check me and say to me,

hey, is this really worth another half an hour of your time? Do you really want to get this thing done? or you just basically just like exploring the possibilities of things to do. And then also just getting a piece with the fact that I've 10, 20, 50, you know, sort of open projects, half done things or things that are 80% done. We have an unusual cultural norm at Weisford Pierce, which is celebrate shitty first versions, which has two components to it. One is every first version of something is going to be shitty. Yeah. So it's not a matter of whether it's good or not.

It's going to be shitty and you celebrate them because it's the hardest thing. It's like learning how to walk. It's purely having a beginner's mind. And also, it has to be complete in order to be a first version. So you need to ship something. So we have this cultural norm of if it's something worth doing, we create it. It's bad and we finish a loop on it. And then we could judge whether it's something that we're kind of like beta's work, right?

You have to finish. You have to be willing. have to put on the table and make it testable, test the hypothesis, and they have to close the loop. And then I would also argue that in this new agentic world, there's so much power to the compounding. So part of the reason why I do reviews, and sometimes I get my bot sometimes blogs, and so itself. So I tell it to write a blog post afterwards on what we learned and what it learned. And what I'm doing is, it sounds kind of performative and pedantic, but what I'm actually doing is I'm trying to compound the learnings.

Right? I'm trying to like, okay, I learned the following. I learned that John has a completionist tendency. I learned that getting projects to 100% is hard, but to get to 80, 90% is easy. So next time when we start a project, I'm actually going to embed that. Are these bought,

these are open claw agents? These open claw agents, yeah, Bartwick is open claw. I have a Hermes agent too. And the key thing there is that you're not always starting a new instance. You're training the same agent on you over and over. I think a lot of problems are.

Yeah, the harness and the memory is the, you know, that memory layer is vital, right? And the memory layer is one of our companies, which is plastic labs. So I'm very sort of focused on, like, building that memory layer of understanding of me and how I work. Well, John, this has been an absolute masterclass. Thanks so much for jumping on.

Oh, it's a pleasure. Congrats on getting to the end of my conversation with John. If you enjoy that conversation, you're going to enjoy my conversation with Dr. Alex and our Wiesner Gross right here.