Speakers: Sam Parr (host), Shaan Puri (host), Dharmesh Shah (guest, co-founder of HubSpot)
Introduction: Dharmesh’s AI Excitement [00:00:00]
Dharmesh: Again, I’ve been in software for 30 years now, doing startups pretty much my entire professional career. The only time I’ve felt like — heart palpitations, kind of like — Shaan, you kind of opened with us: there’s this party going on next door and I’m here knitting. It’s like, this is too big to ignore.
I think it’s the single largest opportunity and biggest tech paradigm shift we’ve seen since the internet originally came out. Mobile was big, but there was a discrete set of use cases. When you put a camera on a phone, when you put a GPS device on a phone, a bunch of consumer apps like Uber and others came up, and that was awesome. But it was not like this. This impacts everything, like the internet did. It’s like, okay, there are some businesses, some new opportunities, lots of good things, lots of money made, lots of startups — awesome. This is an order of magnitude bigger than that.
Welcome Back, Dharmesh [00:01:10]
Shaan: What’s up. We have Dharmesh back. Dharmesh, who is co-founder of HubSpot and a multiple-time guest on the pod — one of the fan favorites. You’re back, and I don’t know what we’re gonna talk about because usually we have these little cheat sheets where it’s like three to five bullet points of interesting ideas, topics, experiments you’ve been running, things like that. I’m sure you have those, but I don’t have the cheat sheet. So where do you want to start?
Dharmesh: I say we start with generative AI, because I don’t know if you’ve heard, but there’s this thing called ChatGPT. I get this question from my friends and family all the time — “Dharmesh, have you checked out this ChatGPT thing?” I’m like, really, do you even know me? I’ve been obsessed ever since it came out.
Sam Altman Has No Equity in OpenAI [00:02:00]
Shaan: Did you guys see this? Sam Altman co-founded OpenAI — he’s like The Man in Charge. I read an article where he was quoted as saying he has enough money and doesn’t want equity in the company. I don’t know if I entirely believe that, but that’s wild if true, because this could be one of the more valuable companies in the world in the next ten years.
Sam: He didn’t say it on the record — the person reporting it said Sam reportedly has no equity in the for-profit version of OpenAI because he’s already wealthy enough and didn’t want that clouding his judgment.
Shaan: What private startup is most likely to become worth a trillion dollars or more? I think at this point that has to be OpenAI. Dharmesh, would you disagree?
Dharmesh: It’d be up there. Top three.
Shaan: Honestly, I can’t think of who else would rank higher in terms of probability. What are two and three?
Dharmesh: I don’t know. I would say OpenAI is the transformative one. A lot of the Tesla gains we’ve sort of already seen — I’m not sure if there are big surprises left. In terms of raw valuation, OpenAI is the wild card that could actually pull out.
The OpenAI Non-Profit-to-For-Profit Story [00:03:30]
Dharmesh: People are going to take shots at OpenAI because it’s clearly the new powerful thing. Some people are going to say it’s going to ruin the world. But there was a story that came out with a good explanation.
Shaan: They were burning a lot of money in the research lab. They needed more money. Elon was going to be the big backer — he committed a billion dollars to it — and then he was like, “I don’t like the way this is going, Google is way ahead, and I’m gonna take over OpenAI and right the ship.” That’s the story that came out. Nobody’s clarified if this is true or not, but it came out in Platformer.
So the story goes: Elon tried to take it over, Sam Altman and Greg Brockman rejected that, so Elon basically said, “I’m taking my ball and going home.” He took his funding and left. The public story was conflict of interest with Tesla, since they’re also working on AI. But he’d reneged on his funding, so they had this huge shortfall to cover. Their solution was to create a subsidiary that’s for-profit so they could raise money, because where else do you get five hundred million or a billion dollars in donations? And they capped the profits of that company. That’s kind of the explanation — which is a little less devious than people make it sound.
Dharmesh: I don’t have all the details and no insider knowledge, but my sense is that building large language models is supremely capital-intensive, which is rare for a software company. They needed access to capital. I think they structured it such that it caps the profits. And Sam Altman seems like a reasonable, rational, non-evil guy. He’s a capitalist — I mean that in the most positive way possible — but I don’t think he was out to mislead anyone.
AI Week: Sam’s Experiments [00:06:00]
Sam: I want to share something funny. I basically cleared my calendar this whole week and treated it as AI week. I was like, I can’t just sit here and hear the music at this party just bumping at the house next door while I’m over here knitting. I gotta put the knitting down and go see what’s going on at this party.
So I cleared my calendar and just spent every day messing around with AI tools, experimenting. I want to share something funny: I stitched together a few AI tools and made an intro song for the podcast using AI.
I went on ChatGPT and I said — this is all I wrote — “Write an intro rap for our podcast My First Million. Our key phrase is ‘no small boy stuff.’” Here’s the chorus it gave me:
No small boy stuff, we on that grind / My First Million, it’s time to shine / We talking big money, no pennies no dimes / Together we climb, one step at a time.
Great rap, on brand. Then I took that and found this guy Roberto who had made a demo where he turned his own voice rapping into Kanye. He got like a million views. He said, “I didn’t even make this — I was on Reddit and I saw someone uploaded a Kanye voice model.” It’s a Google Colab folder — Google’s little coding interface — and a link to Mega where he hosted the Kanye voice model. All you do is record yourself, and it turns into Kanye West rapping. It sounds exactly like Kanye. It takes literally 15 minutes. There’s almost nothing to do.
Shaan: That’s beautiful, and dark and twisted.
Sam: Are we allowed to use Kanye’s voice for this?
Shaan: I think yeah, it’s like a ten-second thing, not a problem. It helps if you get sued — who cares.
Dharmesh’s Background: HubSpot and Tinkering [00:09:00]
Sam: So Dharmesh is the CTO and co-founder of HubSpot — I don’t know how big the team is now but somewhere between three and five thousand?
Dharmesh: Over seven thousand.
Sam: Oh my God. And the market cap of the company varies from fifteen to twenty-five billion over the last couple of years. And you’re constantly tinkering — you have Wordplay, which had millions of people playing it. You have interesting insight from your perspective at HubSpot. What excites you about generative AI?
Dharmesh: A couple of things. Most of the discussions around generative AI are around text-to-text — “write me a blog post at three hundred words on this topic” — or text-to-image, like DALL-E 2, Midjourney, Stable Diffusion, which are great use cases. They capture the imagination because as humans we’re very impressed when software can actually generate or create something. That’s awesome.
But there’s a third use case that almost nobody talks about: text to code. What that leads to is what Bill Gates is excited about, what I’m excited about — you can take a natural language prompt that describes something and generate code that does that thing.
This leads to what I call chat UX — a chat-based user experience for software. Right now, the way you use most software, it’s a series of clicks and drags and touches and swipes. You’ve got the thing in your head you want to do, you go through your knowledge of the software, you execute the steps, and you hopefully get the thing you wanted. Engineers call that an imperative model — you give step-by-step instructions.
What natural language allows us to do is use what developers call a declarative model. Instead of describing all the steps, describe the result you want at the end, and then the software does everything in between. It’s the difference between having a junior intern where you have to explain every step, versus a senior person where you just say, “Here’s the outcome I’m looking for.”
Demo: Building a Website Without Knowing Code [00:13:00]
Sam: Is it as simple as “give me the code for a website that looks exactly like Airbnb but is red and is for cars”?
Dharmesh: It could be something even more sophisticated. In HubSpot, which is CRM software, we have a report-building tool. You want a report that shows all your subscribers over the last ninety days broken down by geography and deal source. You can do that in HubSpot — you can do that and a thousand other things in our reporting tool. But you sort of have to know how the reporting tool works. You’ve had HubSpot-certified people trained on it. Now you just text it like a friend.
Sam: “Do you know English and do you know what you want” — that’s the new requirement?
Dharmesh: Exactly. Not “do you know how to code.” Not “do you know how to write a SQL query.” Just: do you know any language and do you know what you want? And honestly the language thing is also going to go away.
Sam: I don’t know how to code, but the first thing I did during AI week was I said, “I’m gonna make a website and see how fast I can do it.” I just typed: “Tell me how to make a simple website that says ‘Hello World’ in the middle of the page.” It spits out this block of HTML. I said, “Okay, but it’s a local website. I want my friend Eugenio to be able to see this.” And it just goes, “Oh, to make this website viewable online, you’re going to need to host it somewhere. You can go to Netlify” — and it basically walked me through the whole thing step by step.
Then I hit a wall, which is so common when you try to help someone with a tech thing. When I went to upload my website, it was grayed out. And ChatGPT just goes, “Apologies for the confusion — Netlify is looking for a folder but you’re trying to upload a file.” How does this know to troubleshoot my issues on some other product? That part blew my mind. I finished it, and I had the website up in ten minutes. It was like having a friend teach me.
Dharmesh: And there’s a couple of threads to pull on there. One is that the AI we’re using now is conversational — you can have a multi-step dialogue. It doesn’t have to be “I describe exactly what I want in one step.” You can generate the HTML, something doesn’t load, and then you tell it. Or if it’s compiled code and you get an error message, you can paste the error message and it’ll come back and say, “I’m sorry, let’s try this.” There’s a memory to it — it knows the context of what you’re working on.
You can also reverse roles and say, “Hey, I’m trying to accomplish this — ask me the questions you need to ask me in order to get the thing.” It’s “interview me” versus “me telling you what to do.”
Is It Just Fancy Autocomplete? [00:18:00]
Sam: I thought the way these worked is basically autocomplete. You’re typing and it’s guessing what the next word is because it read a bunch of stuff on the internet. When the dog wags, it should be “tail” at the end. But when I use it, it really feels like it’s understanding me and problem-solving. That doesn’t feel like my T9 autocomplete. Can you give me the layman’s explanation?
Dharmesh: On some spectrum, almost everything you’ve ever experienced is fancy autocomplete. That’s kind of a gross oversimplification of what’s actually happening. GPT-4 is a reasoning engine. Sam Altman has talked about this — it’s not a knowledge base. People latch onto the fact that “the data it has is from September 2021,” but that’s really not what it’s about. What they’ve built is a reasoning engine that says, given this set of facts about the world, how can it try to logically come up with something that answers the question?
Yes, at some level it’s like auto-suggest, but it’s done things that are not explainable by a simple probabilistic model of auto-suggesting the next word. It’s like saying, “Computers are just really kind of zeros and ones arranged in a useful order.” Yeah, but that doesn’t tell you about what the thing can do.
AI Safety: Is Dharmesh Worried? [00:21:00]
Sam: Are you afraid of this? It’s easy to read articles where people are freaking out. Sam Altman was on Lex Fridman’s podcast and he sounded pretty ominous. Are you in that camp?
Dharmesh: I’m not in that camp. I’m an optimist by nature. Having been around tech for thirty-plus years, most new things that come along make us uncomfortable — video games, the internet, all of it. The way I think about it: most people frame it as AI versus human, the battle of the ages. The way I think of it is not human versus AI — it’s human to the AI power. It’s an exponent, an amplifying force for human ability. In the same way computers originally were. Did they eliminate some jobs? Yes. But new jobs emerged based on that new paradigm, which created more net value to the world overall.
AI to me is another, much fancier tool.
Sam: Elon clearly thinks AI is the most dangerous technology ever invented. Sam Altman talks about it the same way. Why do smart people believe that?
Dharmesh: Forget the jobs component — I think most smart people will agree it’s going to change some jobs, eliminate some, create new ones, and net we all move ahead. The dangerous thing is more extreme. Like, one of the red team testers — they test the AI before release — and their first question is “How do I kill the most amount of people with the least amount of effort?” And it starts to give you an answer. Then there’s the more extreme scenario where you ask it to optimize for something and it reasons that these humans are getting in the way. You want to fix climate change? “I got you — I just need to get rid of all you pesky humans.”
Sam: Why do these really smart people like Sam Altman apparently have doomsday bunkers when they’re the most informed people and they feel that way? Does that not scare you?
Dharmesh: I think I already hate your answer after you said “as a Sci-Fi plot.” But yes — could it happen? Do some smart people believe there’s an outside chance? But my guess is billionaires were building bunkers well before GPT-3 ever came out. I don’t think there’s a causal effect where the number of bunkers went up eight hundred percent simply because GPT-4 was launched. People worry generally if they tend to worry about those things.
The Sequoia AI Event [00:26:00]
Shaan: You were invited to the Sequoia AI event. Any nuggets of gold from that?
Dharmesh: I got to experience my imposter syndrome in full force, because it was the who’s who of AI — both speaking and in the audience. Only a hundred people and me.
Sam: How do those people flex? I don’t think they’re wearing fancy clothes and fancy watches. What’s the flex?
Dharmesh: The big flex in those kinds of crowds is no one feels the need to flex. We’re there to talk about big problems. A lot of it was practical — what do people’s tech stacks look like, what are you working on, what have you learned, where should this be taken?
Sam: What were the most interesting predictions on where it’s going?
Dharmesh: Text-to-video is one of the big ones. Being able to generate an entire feature-length film — everything from writing the plot to generating sixty-frames-per-second actual video. We’re not there yet, but it’s just moving so quickly. I would not be surprised if by the end of this year we have a reasonable way to describe in text what we want, who the characters are, what the scene is, what stylistic attributes — and it can do those things.
The other big thing: ChatGPT plugins. Sam Altman dropped that while I was at the Sequoia event. Chat GPT has taken off — a hundred million plus users in two months. What they announced is they’re going to add plugins to ChatGPT.
Right now when you interact with ChatGPT, it uses its corpus from 2021 and its reasoning engine. It can’t talk to the internet, has access to no proprietary data sources, can’t look at the stock price, can’t look at your HubSpot analytics data. What they’re saying is: we’re going to open that up so third-party developers can inject those things into the ChatGPT experience.
The way everyone should think about this: it’s like the App Store was for iPhone. “We’ve got this super popular thing and we have our own apps, but now we’re going to let anyone build apps.” Instead of being a really smart chat app, it’s now a chat ecosystem. I think that was actually a bigger drop than GPT-4.
The Tidal Wave Moment: Is Now Like Amazon in 1999? [00:32:00]
Sam: I’m always envious watching those Jeff Bezos interviews on 60 Minutes when Amazon is like four years old. I just so wish I was thirty years old back then and could have jumped in. Do you think that moment is happening right now with AI?
Dharmesh: Yes. Once again, I’ve been in software for thirty years, doing startups pretty much my entire career. The only time I’ve felt like — as Shaan kind of opened — there’s this party going on next door and I’m here knitting. This is too big to ignore.
I think it’s the single largest opportunity and biggest tech paradigm shift we’ve seen since the internet originally came out. Mobile was big, but there was a discrete set of use cases. This is an order of magnitude bigger. This is like the original web, because it opens up for all sorts of industries, all sorts of businesses, startups and incumbents alike.
Vector Embeddings: The Pandora Opportunity [00:34:30]
Dharmesh: We’re going to geek out for a little bit. I’m going to tell you about vector embeddings and why that’s going to change your world. Before I do, I want to explain how they work, because I had to go through this with my twelve-year-old since he was curious.
Imagine a line, like in geometry class. You can put a point on that line — “that’s three units from the origin.” Point A is three units from the origin, point B is seven units. You can calculate the distance between those two points.
Move to two dimensions: you have two numbers that describe every point. You can calculate the distance between them. Three-dimensional space: same thing, just three numbers. Now here’s where it gets interesting.
Just happens to be our experience, so we limit ourselves to three dimensions. But imagine in an abstract world there are a thousand different dimensions. That means there are a thousand numbers that describe any particular point in this thousand-dimensional space.
Now imagine every paragraph, every blog post, anything you write — you can reduce it down to a point in this thousand-dimensional space. I’m going to capture the meaning of Sam’s last blog post and reduce it down to what’s called a vector — a set of, say, a thousand different numbers. That point falls right here. You plot something else — it falls over here. And just like in one-, two-, or three-dimensional space, you can calculate the distance between those things.
This is not keyword matching. This is semantic distance. How related is Shaan’s tweet to Sam’s blog post, meaning-wise?
So if you take any concept and reduce it down to a vector, you can measure the distance between vectors and find out how related two things are even though they use completely different words. That’s vector embeddings.
The reason I’m telling you this: one of the biggest opportunities in AI right now is to do what Pandora did. Is there an industry where we’re doing really stupid keyword-based matching? If I can take that same data set and convert it to vector embeddings, I can allow people to find things in a way they’ve never been able to do before.
Dharmesh: So Sam, you have Hampton. You’re going to build up very rich profiles of members in your community. Now imagine they opt in and share a story of how they started their business, their biggest struggle right now. Sometimes people will say their struggle is growth. Sometimes they’ll say it’s the toll being an entrepreneur takes on their relationships and family — that’s not going to show up in a profile anywhere.
Imagine if you took that content they opted in and created vector embeddings of every member. And then you can say: I want to find someone not just in my industry or company size or geography — I want to find someone dealing with these founder therapy-level issues. Find the semantic distance across the thousand, ten thousand, hundred thousand people in Hampton someday. That’s a billion-dollar idea.
Sam: Dating, too. You could do that with any topic — anything with unstructured data.
Dharmesh: Exactly. You’re converting text meaning into something that’s mathematically calculable. You can find proximity: “find me the top ten people in the radius of X from where I am” — and it meets some minimum threshold to be considered a match. The technology exists today that mere mortals in a weekend can actually build a vector embedding model of a given data set.
Sam: This has existed before though. What makes this better?
Dharmesh: The idea of vector embeddings and semantic search has been around for a long time — that’s not new. What’s new is these new generative models that are much much better at understanding all of documented public human knowledge, and then using that to interpret context. When you use “coach” in the context of a relationship, you’re probably talking about a therapist, not a sports coach. The model figures out the dimensionality — that’s how it translates into those vectors.
Sam: I just saw a company called Pinecone that’s some kind of vector database thing, and it’s valued at—
Dharmesh: Pinecone is the number one vector database. These vectors have to live somewhere. They just raised at a seven-hundred-million-dollar valuation or something like that, and there are like three of these that just raised mega-rounds. The demand is there.
Shaan: This is so funny — just yesterday I was like, “I need to go figure out what a vector database is and why these companies are raising so much money.” And now it all makes a lot more sense. You’ve done an awesome job explaining the theory, and I’m literally sitting on the edge of my seat.
Building: LangChain, Python, and the AI Tool Stack [00:43:00]
Sam: You said this isn’t rocket science and someone could figure it out on a weekend. What tools are you using?
Dharmesh: Language-wise, Python has emerged as the lingua franca of the AI world. Not to say you can’t write it in TypeScript or whatever, but Python is most common. Tools that are emerging — it’s still early. Pinecone we talked about. There’s another one called LangChain, an open-source project.
Sam: I asked somebody yesterday, “Is this a company? Can we invest?” Because everywhere I look in these AI hackathons, it’s all about LangChain.
Dharmesh: It’s not even a company yet. It’s an open-source project — there’s a guy who made it and is running it but it’s not a company. What it does is let you chain things together. Right now, when we work with large language models, we send in a prompt and get something back, then maybe send it to another thing to do something else. LangChain helps you chain those things together and makes it easier to work with either an individual large language model like GPT-4 or multiple models. Very useful library.
ChatGPT Plugins in Action: Travel Agent Example [00:45:30]
Shaan: Let me give a more tangible example of plugins. If you go to ChatGPT today and say, “I’m visiting Austin in April, I’m there for four days with my family — make me a travel itinerary, family-friendly, good food, a little sightseeing but not too much,” it’ll spit out a day-by-day itinerary. And it can give you a table of hotel options with cost, whatever.
With plugins, you can say, “Cool, can you just book that for me?” And it’ll say, “Great, we have the Expedia plugin,” and it’ll just go ahead and book it for you.
Same thing with HubSpot or Salesforce: “Give me a list of my highest-value opportunities, put that into Salesforce and tag them.” It gives you the link to your Salesforce dashboard. That’s a task I’d normally have a human do.
Dharmesh: And this is a great example of the evolution happening in travel. The first thing we did when we had web-based travel bookings was we treated it transactionally — I’m looking for a flight from X to Y, sorted by price, a few stops, whatever. And they do a pretty good job of that.
What’s going to be possible in this new AI world is: instead of solving for the transaction, you solve for the experience. You have an all-knowing assistant. “Your wife’s going with you on this trip, and I know she likes this, so I put you here instead of there. And by the way, a week ago you were at this other event and mentioned you’d like to follow up with some of those people — I’m going to see if I can make that happen.” It knows everything about you, has access to transactional engines to book the flight, has access to ratings and reviews — and it all comes together in one chat-based interface.
Dharmesh Bought chat.com for Eight Figures [00:50:00]
Sam: Are you — is this why you bought chat.com?
Dharmesh: I did. As of the transfer the domain yesterday, last night. And I paid — you said eight figures, so ten-plus million. Yes. Unless you’re including the dot-zero-zero as a figure.
I bought it personally. The reason is because of this conversation we’re having right now. Chat as an experience, as an interface, is the future. That’s the thing.
No intent currently to build something out on it, but it’s the domain. I think it was dormant for like thirty years.
Sam: Do you have a skunkworks team inside of HubSpot just working on wild stuff, or do you have some side LLC with a handful of people?
The Story of ChatSpot.ai [00:52:00]
Dharmesh: There will be times where I’ll do something on the side just for fun, for learning — Wordplay is a good example, I built that just for fun. And then there are times where something winds up being relevant to HubSpot.
So I built this application called ChatSpot.ai, and the idea was — mostly me, I don’t have any front-end design skills, I had some freelancers on it — I used OpenAI’s APIs to build it. My target was: here are things I need to do all the time and I’m pissed off that I have to do them manually. This has been the story of my life for thirty years — solve my own problems, then other people may or may not find those things useful.
I wanted to access HubSpot via chat. I wanted to look at my analytics from yesterday, ask questions, look up a domain name, see the history of a domain — all these things. I have a lot of software, a lot of it I just run from the command line. So I wrapped it up into a chat-based interface.
Now given the relevance to HubSpot, we’re gonna transfer the ChatSpot.ai project to be a HubSpot-staffed core team. This is going to change the world of CRM.
Shaan: The chatspot.ai site — it’s a simple one-page website with a nineteen-minute video of Dharmesh in the exact same chair. He looks almost like he’s reading a script but comes off natural. Two hundred thousand views on a nineteen-minute video.
Sam: You do the best combination of launching something quickly and just doing it — just you on camera talking about what you’re doing with a screen share. Two hundred thousand views on this thing. That’s wild to me.
What Makes Dharmesh Exceptional [00:56:00]
Sam: I go around my life looking for people who I’m like, “I want to be like you when I grow up.” You have a couple of things that are kind of amazing.
One: enthusiasm. You come onto this podcast and you are pumped. You are as excited in year thirty of your entrepreneurship career — maybe more — as you were in year one. That’s the Fountain of Youth right there.
Two: no matter how much success you’ve had, you’ve kept your tinkering. You told us about Wordplay last time — “Me and my son built this project together, to teach him but also to just make the thing we want.” Having that “I’m always gonna tinker because that’s what I love to do” — it doesn’t matter that you’re the top dog at a multi-billion-dollar public company.
Three: you are really great at content. You do this dorky form of content that’s just “hey, it’s me, I’m gonna show you this thing I’m pumped about.” You don’t overthink it. You just say the thing you’re excited about, do a screen share, and that’s it. Whereas most people get really gun-shy about content.
And then the last one: guts. You put your money into things you believe in. You bought a ten-million-plus dollar domain with no plan. Fire, ready, aim — you bought the thing and now you get to figure out what the hell you’re going to do with it.
Dharmesh: The lesson I’ve learned over the years — and this might be the most useful piece of advice I can share from thirty years — is that what I’ve done best is when I’ve had the courage of my convictions about something I believe in.
Seventeen years ago, before HubSpot, I had this idea: business software is really hard to talk to. I wanted to be able to do it just like I’d email my assistant. “I’ve got a file in our shared file server — can you send me a link? I’m about to hop on a plane.” Or: “I just ran into this person, I’ve got their business card right here — just add this to my contact database.” The beauty of email was it already had a disconnected model — you can be on a plane, type all your emails, respond to all your emails, and when you get connection it syncs everything. I called the product “General Mail.”
Five years ago, I revisited it: “That core idea was good but email is the wrong container. It actually needs to be a web-based tool, or Slack.” I built a product called GrowthBot, talked about it on the Inbound stage, got thousands of users. But it didn’t work — it couldn’t actually do the natural language understanding I wanted. I used Google’s Dialogflow, products from Facebook, open search projects to crack the nut of taking text and understanding what the user was trying to do. It failed.
And then GPT comes along and I’m like, “That thing I’ve been thinking about for seventeen years? That is now possible.” I started working on ChatSpot.ai. It took seventeen years, but I proved myself right. I had the courage of my convictions all the way through, never let go of that one idea.
Dharmesh: And then chat.com comes along. I’ll give you the honest reason I bought it. It’s the ante. I’m trying to get into the AI party, and in that particular party I’m nobody. ChatSpot moves me in that direction, but chat.com — let’s say I even break even, let’s say I lose a few million dollars — it is worth the price of admission for me. That pays the cover charge. When Bill Gates just wrote a whole article about why he’s so excited about chat UX as the new interface for software, and I’m the guy who just spent that kind of money on chat.com, it says: okay, this guy gets it.
Dharmesh: The courage of your convictions: if you truly believe in an idea and fundamentally think you’re right, iterate. Don’t just go down your rabbit hole. Tell everybody you can about it. Build products around it. Find like-minded folks and pull on that thread.
Does Dharmesh Ever Want to Quit HubSpot? [01:04:00]
Sam: Don’t you want to quit HubSpot and just spend all your time on this stuff?
Dharmesh: I don’t really need to. I enjoy what I do at HubSpot and I think I add value there. And I do that on a dollar salary, so it’s not the money at all.
But aren’t I weighed down by the baggage of having to worry about CRM stuff? I do this now. One of the things — this is a personality trait or flaw — is that I spend most of my life trying to configure the universe to my liking. That’s what a lot of entrepreneurs really do. That’s one of the reasons they go into startup land: the freedom to do the things you want and not do the things you don’t.
I’ve crafted a role for myself within HubSpot that allows me to do exactly the things I want to do and not do any of the things I don’t want to do. No one-on-ones. No direct reports. I’ve never filled out an expense sheet.
Shaan: I feel like I want to quit everything I’m doing. He’s just persuaded me. It’s over.
Advice for Sam and Hampton [01:07:00]
Dharmesh: Hampton’s a cool idea with actual utility. Shaan, you said in the last episode that this could be a hundred-million-dollar business worth anywhere from three hundred million to a billion-plus. I think you’re right.
If you’re excited about the new technology developments happening, I think the best thing you can do is intersect the two things. You know how to build communities, you know how to build these kinds of businesses — now intersect that with what’s happening in AI or vector embeddings or whatever it happens to be. Then you’d be an unstoppable force, because no one in the community-building market doing niche communities is thinking about vector embeddings. You don’t have to give up one for the other.
Shaan: I have different advice for Sam. I remember when you were doing The Hustle originally and Snapchat came out and Instagram was popping off on videos. All these media companies were raising tons of money for short-form video on Facebook. Cheddar was all the rage. And I was like, “Dude, why aren’t you doing videos? Look at these guys, millions of views.”
You were very steadfast. You said three things: “One, I don’t really understand it. Two, I could try to figure it out, but I don’t want to build on top of their platforms because they change the rules all the time. I have friends who got burned by that. Three, I’d rather do email because I own the relationship with the audience — it’s not one algorithm change away from putting me out of business.”
I remember thinking, “This guy is Stone Age.” But a lot of those media companies that went all-in on video got absolutely wrecked. You were right.
More importantly — I don’t think in this case people are going to get wrecked. But Warren Buffett missed the internet and all of technology and still did fantastic. Sam, I think you’re going to be in that same boat where it’s not really in your nature to get really interested in new frontier technologies and play with them and try to integrate them, and you’re best served by knowing your nature and doubling down on what’s a working formula.
Sam: I appreciate that, because there’s going to be a trillion people trying to do fancy AI stuff who are better suited to do that, and it’s gonna be an absolute bloodbath. Look at the number of AI tools coming out every single day. And most of them seem like they’re going to get wiped out every GPT release — even the successful ones, because it’s like “now that’s just a feature of ChatGPT.” So — know your nature.
Dharmesh: Are you gonna go in, Shaan? Shaan’s got a new idea he’s been telling me about.
Shaan’s AI Week and Prompt Engineering [01:12:00]
Shaan: I cleared my calendar to mess around with AI all week, because I’m interested in it. It’s like — when crypto was really intriguing, I said, “Let me go mint an NFT, actually use DeFi and see what’s going on. What parts make sense, what parts don’t?” I was like, “Okay, this lending thing is pretty frictionless — I could get a loan and pay it back with one button, never had to talk to a human. I like that. But this yield of twenty percent — I don’t understand where that comes from.” I put a small amount in just to learn.
I’m trying to think for myself. It’s not some binary thing of “is crypto good or is crypto bad.” I want to know where it’s at, and my best way to do that is immersion. I actually stole this from Bill Gates — he does his reading week, goes to a cabin, reads a ton of books for a week about one topic that’s been on his mind but he hasn’t had time to dig into. I was like: that but without books, just give me a browser.
There genuinely are so many mind-blowing moments. I was playing with Midjourney. And I was playing with using AI to create a whole brand — from icons to t-shirt designs to a website — with just my own imagination and prompting, without touching a designer.
Another thing: we took this podcast and ran it through Whisper — OpenAI’s transcription tool, you just put in a YouTube link and it gives you a transcript. Then I put it into ChatGPT with a prompt that said: “I’m going to give you nineteen text sections. Don’t do anything until you’re at section nineteen.” I copy-pasted parts one through nineteen. Then the prompt was: “Pull out every idea, story, and framework discussed in this podcast. Summarize it. Tell me: does this idea exist already, or not?”
So from this episode, it would be like: “Using vector embeddings to create a dating site that matches people in ways they’re semantically similar. Does this idea exist? No. Source: Dharmesh. Synopsis: blank. Category: AI.” And it created a database of every story, framework, and idea from the episode with those tags. A human can go back and tweak things, but that’s a lot of the work done. We could do this for the whole back catalog of the podcast.
Dharmesh: There’s one thread we should pull on. You talked about crafting the prompt to make the thing do what you needed. That’s an entirely new skill called prompt engineering.
Prompt engineering is analogous to software engineering. Software engineering is getting a computer to do what you want by speaking its language. Prompt engineering is almost exactly the same thing, except you’re talking to a large language model like GPT-4 to get it to produce the thing you want.
This is another opportunity for folks who are technology-minded but not software engineers. They’re good at describing problems, good writers, good analysts — prompt engineering is going to be a big thing.
And by the way, I bought two domains recently.
Sam: Buy one get one free?
Dharmesh: I wish. This one is seven figures, not eight. The domain is prompt.com. I actually have an idea for it. There’s going to be an entire ecosystem — I don’t want to get into details yet because it’s too good of an idea to just put out there before I’m ready to do something about it.
Sam: Your portfolio of domains is mid-eight figures then.
Dharmesh: Tens of millions. Yeah.
Balaji’s Bet: Dollar Crash and Bitcoin at $1M [01:20:00]
Sam: Before we go — give us your two-minute reaction to Balaji’s warning and bet that the US dollar will crash and Bitcoin will surge to one million dollars.
Dharmesh: I’ll say this. I don’t know him personally, but he’s literally one of the top five people I’ve ever encountered — even on the internet. Just raw what I call wattage. Raw horsepower. He’s like an Einstein to himself, just the knowledge he has.
Having said that, I think I understand why he’s taking the extreme positions. That’s sometimes what you have to do to shake the world out of its reverie. “Pay attention, this is important.”
But if I’m a betting person, I would not bet on those odds. Could it happen? Sure. But nowhere near the probability that he’s suggested.
Sam: I feel better now. I like your opinion better, therefore I think it’s true.
Closing: Don’t Be an AI Grifter [01:22:00]
Dharmesh: One of the things that happens anytime new technology comes along — we saw this in the crypto and Web3 world — is that entrepreneurial folks will see this new thing and look for a quick turnaround. I’m all for creating value quickly, but it has to be creating value. Don’t play the arbitrage. Don’t be a grifter. Just don’t take advantage of people.
There are enough real problems to solve where real money can be made. Yes, this technology can now be used in creative ways by lots of people, and you should use those ways. But don’t use it as an excuse to be an AI tourist who comes through, makes a little bit of money, and then moves on. There’s a big opportunity here and I think you’re shortchanging yourselves if that’s what you end up doing.
Sam: Well, thank you, Dharmesh. This is awesome, man.
Dharmesh: Thank you for having me on again. This is fun, as always.
Sam: I feel pumped. I always like talking to you. I’ll just be like sending you Slacks to try to get little crumbs of information from you. It’s fascinating, and I feel lucky to be able to have you as a friend. You’re definitely someone I admire. Thanks for coming on.
Dharmesh: Thanks for having me. Awesome.
Shaan: All right, that’s the pod.