We have the best entrepreneur I know, Dharmesh Shah, co-founder of HubSpot. He's got a brand new project. He's going to take us through a tour of what the heck is happening in AI right now, why everybody needs a solo CRM and Second Brain to work on a day-to-day basis, and more importantly, where is all this going? And stick around because he's giving you a super secret discount code for his new app, Uspot, that is exclusive to the show. This is going to be a banger. So, let's get right into today's show.
Okay, Dharmesh, welcome back to the show. Always exciting to have you on. Where we wanted to start, Dharmesh, is that you've been deeply in the weeds of AI for some amount of time. You've built a bunch of different things and you're seeing what's not working and basically what do users love? What's working in AI now? So, can you maybe just give us the landscape? You know, how is AI working today and what are you excited by?
Well, thanks for having me again on the show. If we think back to prehistoric times, which is November 2022, we had the ChatGPT moment, right? Everyone sort of woke up to ChatGPT. It's like, oh, this is amazing. It can do amazing things. And it could. And the interaction model at that time was you sort of typed in a prompt and something often useful and sometimes amazing came back. And although no one really named the moment, there was a similar moment that happened November of last year. And what happened was multiple models around the same time frame got much, much better from a reasoning perspective. So back then you issued a prompt, it would do its best. It would do like a first pass, like here's your thing, and you could obviously interact with the output. But in November of last year, what happened is the model sort of got better at reasoning and self-reflection. So you could give it a higher order task that would take longer. So it would kind of take its time with it. And sometimes you could even instruct it, like, hey, you know, think deeply about this one. You take your time, it's important, and it would go do that. The primary use case that sort of really blew up as a result of that capability is agentic coding or vibe coding, which is now you can give it a coding task and it can go reflect and can go do things and produce something, not just a line of code, not just one module, but build entire products or apps or features and things like that. So, that's the sort of curve we've been on. And I've been kind of all over the AI thing, but now it's like I was up till 4:00 a.m. last night, agentic coding on, you know, the new project, which we'll talk about. And the reason is it feels like I have this, you know, superpower. And it would be like a waste of the universe's time for me to go sleep. It's like I've got access to this thing. How can I sleep? Like, that's not, you know, it's like that's, you know, five hours I'm going to lose, you know, building this next thing that I can build. And obviously that's not sustainable. I'm not suggesting, you know, everyone, you know, work 24/7 or something like that, but it's just so much fun because now you can sort of will into existence a feature or an idea or a project or an article or a book or anything because the models have gotten, you know, just so good now.
Yeah, you know, you co-founded HubSpot, built HubSpot, and then you have built a bunch of things since then. One of them was agent.ai, and it was a marketplace for agents. And then, as you said, like, things have evolved even more since you had started that project. And so November recently got a lot better. What are things that you can do now that you had dreamed about that you couldn't have done pre those models? Like, what has really changed for the things you can bring into the world?
Yeah, so agent.ai is now, you know, part of HubSpot in a new product we just launched called Agent Builder. So we lost sort of the marketplace. That's the part that didn't really work. I don't think the world was ready for an app store for agents just quite yet. I think it's going to happen someday.
But Agent Builder is now part of HubSpot. And what I've been working on is this thing I've had in my head for a while now called Uspot. And Uspot is a CRM for one-person companies.
But at the heart of Uspot, the part that's sort of really fun and interesting about it is this thing called Second Brain.
Yes. And, you know, there's lots of different definitions of it, but I'll give you mine since I own the domain secondbrain.com. I get to...
Oh, you own that officially. Oh, I'm going to be so jealous of you now. Of course you do.
So, and I've had it for years, so it's not, you know, something that just, you know, crossed my mind, but...
How many people have tried to give you money for that domain?
A fair number, actually. It's on my list of domains that outranks, like, I own vibecoding.com also, which is, um, and I own prompt.com. Anyway, bunch of AI-related domains.
If you think of any AI word, Dharmesh owns it.com. Yeah. Well, he either owns it or has tried to own it, and he'll always succeed.
Hey, if you're enjoying today's show, we have put together some of the best AI resources on the planet to help you get better at marketing with AI. You can scan that QR code above or you can click the link in the description below and get those right now. But the idea of a second brain is sort of almost self-descriptive, right? The idea is this kind of digital auxiliary to your own personal brain, right? So we have awesome brains, very energy-efficient, and they're great, and they build memories over time. And, you know, when we sleep, they sort of kind of refactor the stuff that you've remembered and things like that. There's all this kind of magic biology that happens, but there are also limitations to our human brains, which is, okay, well, it can only hold so many things in current memory at a single time. Sometimes it sort of loses the plot because there's only so much, you know, capacity it has. So there's limitations to it. It's also limited in the sense that we only have one of them. There's only so much it can do. And so the idea of a second brain is a power boost for your brain. It's like, now that we have AI, can we say, oh, well, AI has been progressing and that's been awesome. The models have gotten smarter. So that's like the IQ getting higher and higher. But the reality is the model itself is sort of frozen in time, right? Whenever the training happened, it's like, okay, the weights got established and now it's this black box that you can kind of feed things into and get things out of, but it has no concept of memory on its own, right? And what the second brain does is it kind of supplements an AI model with an actual memory and things that it should know about you, your business, the things you're trying to do, and what's happening in your world, and can sort of bring that in. And then it can be layered on top of a frontier model like a GPT-6, Astra, or, you know, Fable, or anything like that, and you can sort of do like much more amazing things than you were able to do even a year ago. That's the part that's exciting. So in terms of the second brain, this was like a thing that caught on after Andrej Karpathy, ex-OpenAI, I think co-founder, talked about creating an LLM wiki for their agents. So the agent can kind of have its own little wiki, which means it has memory and can get to know things over time. And I think the concept that maybe you're excited by is that you as a person collect information. You collect information in the emails you receive. You collect information in the LinkedIn contacts that you get. You collect information from the meetings that you have. Like, meetings have become much more valuable, but now because instead of trying to jot down a couple of notes, you actually have the full transcript and AI can parse that and get smarter over time. But is the way you think about the value of the second brain as pre-AI, you had all of this information that you were collecting that you probably were either unaware of how valuable it was or you just could not keep on top of, and now you have a real agentic system that has persistent memory and can actually make sense of that and value of that for you as the user?
Yeah, that's exactly it. So, you know, Karpathy's kind of LLM wiki idea, which, you know, really blew up, especially within the kind of geeky, techy circles. The idea was to be able to take things that you might have in Obsidian or on your file system on your local Mac or the Mac Mini you bought just to do this kind of stuff, and you can run an LLM against it and it sort of builds this wiki with, you know, pointers between documents and things like that. So it sort of creates this network-style or graph-style structure using the power of an LLM, and then you can sort of chat against that kind of personal wiki. So the underlying concept is kind of very second-brain-ish. It had a couple of challenges to it. One is you sort of had to run it on your local machine, and there are benefits to that for sure, but it also sort of kind of presupposes that you're going to be able to reduce the things that you want in the wiki into some sort of file that's resident on your file system, because that's the thing that was used to kind of slurp in data via the LLM. And my thinking on this, and I'm going to geek out for just a little bit. In that concept with the LLM wiki, you know, the file was effectively the primitive, and then you can use that to kind of build out the wiki and do inference around it. The way I'm thinking about it in the second brain that's inside Uspot is there's a set of primitives, but the primitive that I think about is it's an object, and there are different kinds of objects, of which file is one. You could have an email, you could have a LinkedIn post, you could have a LinkedIn comment, and all those objects, if they're related to each other, have connections between them, forming this very, very rich graph of nodes and edges, connections between these objects. And so what's cool is that then you can say, okay, I've got this way to hold things in this kind of object-with-links-between-objects format. Now I can build connectors to say, okay, I'm going to bring in information from this thing, from my email, from my LinkedIn, from my Twitter, from my whatever. And the brain sort of gets bigger and bigger over time. And then as you interact with it, even the prompts that you put in, even the questions that you ask, even the response emails that you send are all objects. So the kind of phrase I have in my head is like everything in Uspot is an object. Everything, right? It's like your history, everything, all the memories that it gleans, everything is sitting in this object. And now it's like you can expose that second brain across multiple interfaces. The most natural is Uspot has a chat interface, as you would expect, very much like a ChatGPT or a Claude, and you can interact with your second brain. You can put things in, you can take things out, you can have long-running tasks. You can run cloud agents against it. This is a very kind of OpenClaw or Hermes-inspired cloud agents that say, oh, I've got your second brain, I'm watching for this kind of stuff. And you can kind of program your own agents to run on top of it. And then you can also just provide that second brain to any AI app that you're using over MCP. So the idea is build this second brain. Its value compounds over time. It learns from you and the things you plug into it. Then you make that second brain available to yourself directly and then yourself indirectly, possibly through a ChatGPT or a Claude cowork. So it's sort of portable. So you build it, you own it, it's yours, and you take it with you wherever you want. And you can build agents on top of it and things like that.
I brought it up because I want to make sure people get that concept because it is something you have to wrap your head around. If you think of this as a personal CRM and you're thinking like contacts and companies, because when you go into your second brain on Uspot, you see like tweets that I've sent out and then briefings, and you can see our prompts. You can see all of the interactions happening. So Kip and I did a really popular YouTube episode and had a bunch of LinkedIn reactions, and you can actually go into like who's liked it. It stores everything as an object. So it's like this giant knowledge graph of everything happening across the data sources that you want to populate.
Well, there's a couple magic things happening here. One, Dharmesh, the whole second brain concept, I think, was kind of relegated to dorks for a while. Like, oh, I've got my Obsidian database and I know how to set it up, and it's like very hard.
Should I be offended, or I'm not sure that...
You should definitely be offended. No, no, you shouldn't be offended. You're a dork and you like being a dork. It's fine. Can I give a shout-out to someone here while you say, because there's this guy called Eric Morrison, Dharmesh, you should check out. I don't know if you've seen this guy on YouTube yet, Eric Morrison. He's, uh...
He basically does like these amazing like 10, 11-minute videos making fun of AI influencers, and he did one on like second brains, and it's one of the best and funniest videos you'll watch. And it's all of the people who have created second brains on Instagram and like just dorking out on them. So...
I agree, Kip, I do agree about it. As like, I would count myself as like a second brain dork. The reason I'm bringing this up is, Dharmesh, I've had the good pleasure of knowing you for, gosh, 15-plus years now. And you're really good at taking this thing that was for these really technical people and making it accessible for everybody. You know, the every-person version. And I look at what you've done with the second brain, and people be like, oh, well, he actually put the second brain thing, made it easy, put it in the cloud. I'd actually argue you're reimagining what people should think a CRM is. It used to be the CRM is like, oh, I have this person's exact name, address, email, phone number, and that's how I think about my relationship with them. And it's like, oh, they liked my LinkedIn post or maybe watched my YouTube video. That's like, it's a completely different way to think about the relationships we have with each other, especially if you're like a solopreneur and you would have never been able to like fathom tracking this stuff before.
Yeah, that's exactly it. I'm, you know, thinking through the dork reference, of which I'm definitely...
Sort of the, you know, course of my life over the last now 30-plus years is I take things that...
Only dorks could use or would use, and democratize it for others, effectively.
That's true.
You know, 20 years ago when HubSpot started, there was this thing called the internet. And, you know, you can now build applications. It's not just static web pages or something. You can actually, you know, build websites that serve customers and build applications or whatever, and there should be millions of businesses benefiting from this thing called the internet. Let's go build something and democratize kind of access to what was then called inbound marketing. And now it's the same sort of thing. It's like, okay, there's this new thing called AI, and we're just on the, you know, very, very precipice, and we sort of get to, you know, geek out on these kinds of things. But, you know, there are tens of millions, hundreds of millions of people that are sort of using AI to some minimal degree, but there's so much more they could get out of it. And like even from a business context, like CRM, I think the evolution of CRM will be at the heart of it, this kind of second brain-style context. And HubSpot's building that for go-to-market teams, and I'm sort of experimenting on the fringes. And Uspot is specifically focused for one-person companies. It's like, if it's just you, you're a solo operator, yeah, this is the CRM for you, right? It's like a very lightweight, second brain-oriented CRM. And then, yeah...
One of the very first emails we exchanged post the ChatGPT moment is, like we said, there will be more smaller companies. And it turns out that that started to happen to the extreme, where like there's a lot more solopreneurs and very, very small companies because you can now do a lot more with AI. And I think the needs for those people are very different. You're kind of showing like even the context in Second Brain is very interesting because it's like those are all people who have interacted with you in Uspot in some way across your different social accounts or email or what have you. But the other thing, Dharmesh, is that it's like very proactive. Like my Uspot daily email has like become one of my most-read emails just because it has all this knowledge about me and it's like kind of managing me and it's telling me where the opportunity is. You know, we're launching this Loop book in a couple weeks, and it's like a lot of that email is about that, and it's like fascinating how good that stuff now is.
Yeah. Two things. One, on the kind of small business front, some stats I think are fascinating. So the world has seen an increase in the number of new company formations that's been happening now, and the US Census Bureau puts data out. Stripe puts data out. In 2025, we had the record highest number of new company formations in the United States than we've ever had in history. Roughly just a little under 5 million new businesses were formed. Through July of this year, we've already exceeded 4 million. So 2026 will be another record year. So that's actually interesting. What's even more interesting is the US Census Bureau breaks this out. They bucket all the new company formations into two types of companies. They're the companies that are one-person companies and will remain that way. They have no plans or are unlikely statistically to hire an employee in the next two years. That's how they define these kind of one-person, you know, solo companies. And then they have the other bucket, which is people that are likely, yep, they're one person right now, but they're likely to grow into companies that will have employees and will grow beyond the one person. The percentage of new company formations right now that are not expected to hire any employees by the US Census Bureau's definition is now 86%. Are not expected to hire. 86% of all new companies started are not expected to hire.
Are not expected to hire.
And so, and they may be right, they may be wrong.
