WEBVTT
00:00:00.000 -->
00:00:00.540Brief history.
00:00:00.540 -->
00:00:09.869When HubSpot started, the fundamental thesis was, Hey, there's this thing called the internet that millions of small businesses should be benefiting from, but they're not, and the reason they're not is because it's scary.
00:00:09.900 -->
00:00:10.800It was very daunting.
00:00:10.800 -->
00:00:13.230So our thing was, how do we make it approachable?
00:00:13.230 -->
00:00:15.599How do we democratize this thing called the internet?
00:00:15.630 -->
00:00:21.179We're doing the same thing with AI that says, okay, there's this new thing called AI that millions of small businesses should be benefiting from.
00:00:21.210 -->
00:00:23.309What do we need to do in order to kind of get them into the fold?
00:00:23.339 -->
00:00:25.350Before the internet, you didn't have the internet, but then you did.
00:00:25.350 -->
00:00:26.940So you'd be an idiot not to use the internet.
00:00:27.054 -->
00:00:27.984AI is the same thing.
00:00:27.984 -->
00:00:36.113If you build a reputation online in the algorithm's mind that you are crappy content, not trustworthy, it's hard to dig yourself back out of that hole.
00:00:36.113 -->
00:00:37.945And the same thing applies to ai.
00:00:38.003 -->
00:00:39.354You still code yourself.
00:00:39.414 -->
00:00:42.174It almost seems to me like it's gotta be a waste of your time.
00:00:42.174 -->
00:00:44.875Is there a reason why you're like deep in the trenches?
00:00:44.905 -->
00:00:45.265Yeah.
00:00:45.265 -->
00:00:46.945And I'm going a mini rant here.
00:00:46.973 -->
00:00:51.295So Dharmesh, thanks so much for doing this.
00:00:51.534 -->
00:00:52.314Thanks for having me.
00:00:52.795 -->
00:00:54.744Uh, so you do a magnitude.
00:00:54.759 -->
00:01:03.039Of things, uh, but you're best known as the co-founder and CTO of HubSpot, and you guys were really early to inbound marketing and blogging.
00:01:03.820 -->
00:01:10.959If you were to go back in time and start HubSpot today in the AI search era, what do you think inbound would look like?
00:01:12.069 -->
00:01:13.420So let's take a step back.
00:01:13.420 -->
00:01:13.840So the.
00:01:15.284 -->
00:01:21.584Early days of HubSpot, SEO search engine optimization, as we all know it now, um, was a relatively new thing.
00:01:21.584 -->
00:01:24.405And that was, uh, a key part of HubSpot's product.
00:01:24.974 -->
00:01:36.825And, you know, I've been thinking a lot now about, uh, what many call a EO uh, answer engine optimization that says, okay, well this is like SEO, but now we're solving for these large language models in the, in the AI apps.
00:01:37.305 -->
00:01:38.685Uh, so a couple couple things.
00:01:38.685 -->
00:01:41.025One is, um, you know, when we started.
00:01:42.329 -->
00:01:47.700With, with the, uh, SEO product that we had, SEO was a bit of a dark place.
00:01:47.730 -->
00:01:59.219Um, you know, we had just gotten over a lot of black hat tactics and keyword stuffing and all these things that, uh, you know, some companies and, and, uh, people were playing effectively trying to trick the algorithm.
00:01:59.670 -->
00:02:06.840Um, and our approach was the opposite of that, which is that we don't think that that, uh, really sustains, uh, long term.
00:02:06.840 -->
00:02:09.060And what you should really be doing is helping the algorithm.
00:02:09.465 -->
00:02:18.254Deliver great results, um, that their consumers will be happy with and, and trying to fight Google, uh, which was a pre predo, you know, obviously, and still is a predominant search algorithm.
00:02:18.974 -->
00:02:26.129Uh, I think all of those things, uh, still apply in an AI world that says you wanna create value for your users, you want to actually.
00:02:26.860 -->
00:02:35.050Help the algorithm, uh, and, and the way you did it back then, uh, you said, oh, well, I'm gonna have content that's actually useful, that answers a user's question.
00:02:35.409 -->
00:02:40.419And there was a phrase I used back then, uh, called, the best way to rank is to be rank worthy.
00:02:41.469 -->
00:02:51.129And it's as simple as that, that says, okay, if I were out there like searching for X, Y, Z, whatever the keyword or a search phrase is, would I want to see this webpage as the very first result?
00:02:51.819 -->
00:02:55.210Honestly, if you ask yourself, you know that question, and if the answer is no.
00:02:55.919 -->
00:02:58.829Job one is to make that the true answer.
00:02:58.829 -->
00:03:03.209Like, this is the best thing that deserves to rank above the other things that are being ranked right now.
00:03:03.569 -->
00:03:16.224Because once you do that, once you, um, create content that deserves to rank higher than the things that are currently being ranked, you're effectively kind of swimming with the tide over the fullness of time you will rake because Google's got.
00:03:16.889 -->
00:03:23.219Hundreds of engineers that, you know, that were working on the problem of getting the best content to actually be, uh, you know, highest in the rankings.
00:03:23.490 -->
00:03:37.830So I think all that kind of philosophical approach still holds true now, but I think we need to sort of keep in mind is that, you know, back then the atomic unit, uh, for, uh, for the internet, uh, at large was web pages.
00:03:37.830 -->
00:03:39.569So that's what Google was indexing.
00:03:39.629 -->
00:03:40.620That's what was ranking.
00:03:40.620 -->
00:03:43.680It was a link to a webpage and there was the link graph and all those things.
00:03:44.219 -->
00:03:45.479Now, the, the unit.
00:03:45.868 -->
00:04:01.588Is actually smaller and we have a, another degree of indirection, uh, because it's the LLM looking at content that's out there using search like capabilities, passing it through an LLM, synthesizing the answers, quote unquote got, and then giving you the actual answer, right?
00:04:01.588 -->
00:04:06.838That's what makes perplexity and, uh, the search in ChatGPT, they're, they're so amazing.
00:04:07.349 -->
00:04:13.319And so the way businesses should think about this is that how do I take the content that I have or produce new content?
00:04:13.979 -->
00:04:16.829Uh, that helps the LLM arrive at an answer.
00:04:17.670 -->
00:04:19.500And there are a couple of tactical things you can do.
00:04:19.500 -->
00:04:26.129One is, so back in the day, you would, you might write an 800,000 word blog post on whatever topic of interest.
00:04:26.129 -->
00:04:28.860And, and Google was really great at saying, oh, that's high quality value.
00:04:29.069 -->
00:04:30.060Click throughs are great.
00:04:30.089 -->
00:04:31.290Uh, the bounce rate is low.
00:04:31.500 -->
00:04:37.560All the things that, you know, Google is looking at in order to kind of, uh, determine it, determine it to be high, uh, high quality content.
00:04:38.069 -->
00:04:40.589Now what's considered high quality is.
00:04:41.084 -->
00:04:54.735Things that are easily quotable, like snackable content by the lm, uh, that are pithy and kind of precise and accurate, that fits well inside the answer that, that the LM is going to provide.
00:04:54.735 -->
00:04:54.884Right?
00:04:54.884 -->
00:04:56.894So if you, if you have a 400 word answer.
00:04:57.194 -->
00:05:09.375That's sort of meandering that sort of gets at the, the question that's much less likely to be cited than something that provides exactly the answer in a kind of nice, uh, soundbitey sort of forum in a, in a way.
00:05:09.435 -->
00:05:21.134Um, and so I think the way we need to kind of think about it is like, you know, all, all of us, uh, have content assets, uh, that were kind of effectively written for humans to consume, which is great.
00:05:21.163 -->
00:05:22.125Uh, I love humans.
00:05:22.663 -->
00:05:25.004Uh, but now it's like, okay, we're gonna still solve for that.
00:05:25.468 -->
00:05:33.449But, um, in the same way that we would take a webpage and optimize it for SEOs, like, I want fast loading, I want semantic tags, I wanna use my H ones properly.
00:05:33.449 -->
00:05:34.139All those things.
00:05:34.468 -->
00:05:41.788Now they're a new set of tactics to say, okay, how do I take that content or those ideas and make them, uh, so possibly put them in question, answer form.
00:05:41.788 -->
00:05:46.228Let's make it easy for the LM to know exactly which questions this article answers.
00:05:46.620 -->
00:05:48.269By just, just tell them.
00:05:48.300 -->
00:05:49.769It's like, yo, here are the 10 questions.
00:05:49.860 -->
00:05:50.730Here are the answers.
00:05:50.819 -->
00:05:52.110Like, nice, snackable form.
00:05:52.110 -->
00:05:58.800That's, uh, you know, that's citable I guess would be my, uh, so thank you Kko for coming to my, uh, Ted talk on, uh, answer optimization.
00:05:58.800 -->
00:06:00.389But anyway, no, that was incredible.
00:06:00.449 -->
00:06:02.189Uh, there's a lot I wanna unpack there.
00:06:02.218 -->
00:06:11.189So, a EO, it's this brand new distribution channel, and, um, it seems like HubSpot's pushing hard into this.
00:06:11.279 -->
00:06:13.259A lot of companies are trying to push hard into this.
00:06:13.995 -->
00:06:23.625If someone were to start today and they, they watched this video and this kind of like was the light bulb moment for them, what are the first two to three things they should do?
00:06:23.625 -->
00:06:25.935Should they be rewording their blog articles?
00:06:25.935 -->
00:06:27.584Should they be creating new content?
00:06:27.959 -->
00:06:28.800What's the first step?
00:06:29.100 -->
00:06:29.730Uh, both.
00:06:29.730 -->
00:06:33.240So find out, um, which you can do with existing tools.
00:06:33.269 -->
00:06:37.800Which of your content is already kind of considered high quality by the search engines themselves?
00:06:37.800 -->
00:06:37.980Right.
00:06:37.980 -->
00:06:46.769That's a, um, there are lots of free tools and paid tools that, that will do that, that content that I think should be, um, I'll say reformulate.
00:06:47.625 -->
00:06:57.283Um, so that it has sort of an LLM mindset in the same way that we said, oh, we're gonna, we're not gonna write content for the search engines, we're actually writing it for the humans.
00:06:57.913 -->
00:06:59.473We're still gonna be writing it for the humans.
00:06:59.473 -->
00:07:04.423But now we have this abstraction layer, which is the LLM that's going to consume the content on the human's behalf.
00:07:04.423 -->
00:07:07.245It's still a human that's gonna read it and they're gonna get that little snippet.
00:07:07.245 -->
00:07:12.973But how do you, um, you take the content of the ideas that you have that you know are high quality, and how do you.
00:07:13.470 -->
00:07:15.959Uh, I'll say, almost translate them.
00:07:15.959 -->
00:07:18.629This is actually, um, a decent metaphor.
00:07:18.629 -->
00:07:22.019I'm making it up as I go, but, so let, let's imagine you had, uh.
00:07:23.564 -->
00:07:27.675A hundred blog posts in, uh, in written in English, uh, for, for human readers.
00:07:27.735 -->
00:07:28.425Uh, great.
00:07:28.634 -->
00:07:37.694And then you discover all of a sudden that you have a bunch of people in Japan that are reading your content, that are potential customers, potential subscribers, uh, based on the business you're in.
00:07:37.964 -->
00:07:46.845What you would do is you say, I'm gonna take my English content, same ideas, and I'm gonna translate it to this other language, which is Japanese, in order to better serve that particular audience.
00:07:47.764 -->
00:08:10.235So what we're talking about now, moving from an SEO mindset, uh, solving for Google, the search engines to an LLM mindset solving for um, AI is to say, oh, well, how can I take the ideas and the content that I have and translate it, quote unquote to make it more easily consumable, uh, by large language models versus consumable bubble by the search engines and, and humans directly.
00:08:10.654 -->
00:08:11.524That makes total sense.
00:08:11.584 -->
00:08:11.884Is there.
00:08:12.478 -->
00:08:18.899Maybe this is completely inaccurate based on what you first said, which was essentially just create good content.
00:08:19.019 -->
00:08:19.168Yeah.
00:08:19.168 -->
00:08:20.819And the LLMs will, will rank you.
00:08:21.089 -->
00:08:28.199But are there any like AI workflows you can use to translate, I know it's not perfect, but it can get you like 80% there.
00:08:28.560 -->
00:08:35.490Is there anything HubSpot or any of your startups or are are doing to basically rank an LLM search using.
00:08:36.134 -->
00:08:38.083AI to get them there.
00:08:38.323 -->
00:08:46.605Yeah, so we have not, uh, uh, it's not meant to be a HubSpot sales pitch, but the, the, the approach here is, uh, you do want some amount of automation.
00:08:46.605 -->
00:08:47.174The nice thing is.
00:08:47.909 -->
00:08:55.948Uh, yes, AI makes things harder because they're different now than they used to be at SEO, but AI also brings a lot of kind of technology to kind of help with that process.
00:08:56.339 -->
00:08:57.179So, a few things.
00:08:57.179 -->
00:09:06.119Uh, if you're a like existing startup or an existing business, uh, one of the first things you need to do if you're not already doing it is, uh, tracking your traffic.
00:09:06.479 -->
00:09:14.309Just like back in the day, we would track and say, how much of my traffic is coming from organic versus paid, you know, for PPC versus social versus whatever.
00:09:14.609 -->
00:09:17.039Now you need to have a bucket around ai.
00:09:17.404 -->
00:09:21.695It's like how much of my traffic is coming from the AI engines ChatGPT, perplexity of the top ones.
00:09:21.725 -->
00:09:24.485Um, and there are tools out there that do that, but you need that.
00:09:24.485 -->
00:09:25.115So you have a sense for.
00:09:26.009 -->
00:09:39.120And the good news, bad news is I will bet you money that, uh, your organic traffic is down because, um, uh, you know, human attention is moving from the search engines, um, to these, uh, answer engines to things like chat, GPT and perplexity.
00:09:39.659 -->
00:09:43.679Uh, so that's kind of thing number one is to get a sense for where you stand, um, and how it's trending.
00:09:43.679 -->
00:09:47.549So as you do things, uh, you can kinda get a sense for what's working and what's not.
00:09:47.969 -->
00:09:50.399Uh, number two is to have, uh.
00:09:51.720 -->
00:09:58.980And you can use, um, AI tools for this existing ones or, uh, you know, HubSpot's building an ai a EO tool is to help with that workflows.
00:09:58.980 -->
00:10:05.730Like, how do I take my existing content and rethink it, reimagine it for, for a EO, um, and then you can kind of track it.
00:10:05.789 -->
00:10:14.610Um, so that would be the step two, but it's like, my, my advice would be this stuff is just like, you know, uh, search engine optimization was, you know, early 25 years ago.
00:10:15.000 -->
00:10:17.700Uh, this stuff is early, uh, changing fast.
00:10:18.075 -->
00:10:20.745Uh, but I think the fundamentals actually don't change.
00:10:20.745 -->
00:10:33.434So even if you think about, uh, where LLMs will source the information that are kind of going to be considered for citation, considered for being included in the answer they provide, uh, provide the person, um.
00:10:34.139 -->
00:10:35.610All those fundamentals are still gonna apply.
00:10:35.639 -->
00:10:40.110In some cases, they're actually using, uh, classic search engines or search engine technology.
00:10:40.110 -->
00:10:41.549So those things are still going to matter.
00:10:41.610 -->
00:10:49.409Uh, the, you know, the fact that you have lots of inbound links, the fact that you're, uh, referred to by other sources, um, is still gonna apply.
00:10:49.409 -->
00:10:50.700So that's not wasted effort.
00:10:50.970 -->
00:10:53.490But now we just kind of need to, uh, take up a level.
00:10:53.820 -->
00:10:56.100The other thing to kind of be prepared for, um.
00:10:56.759 -->
00:11:07.740And this is one of the, uh, advice we give, uh, give HubSpot customers is think about, uh, taking your content and reducing the primitives down, uh, that make it easy for you to kind reformulate.
00:11:07.828 -->
00:11:13.318Uh, so what I mean by that is like, think in chunks instead of webpages to say, okay, here are the, here's the atomic unit.
00:11:13.350 -->
00:11:21.029The atomic unit might be like a question, answer or pair that says, I have, here's the, and there might be 5,000 questions you answer across 200 blog posts, whatever it is.
00:11:21.389 -->
00:11:24.360But if you can get it down to those chunks, that AI can definitely help with this.
00:11:24.855 -->
00:11:40.664Then what happens is as we learn more about, uh, what LMS are favoring or what they're looking for, what actually works, it's easier to kind of retake those chunks and, and, uh, reassemble them in ways that make sense versus taking a, a long form narrative and trying to kind of translate that on the fly.
00:11:40.664 -->
00:11:46.184So that would be another kind of piece of tactical advice that should be part of your flow is to take, um.
00:11:47.039 -->
00:11:53.460Kind of reduce down to kind of snackable bits, uh, the key insights or ideas that are, uh, embedded within your content.
00:11:53.460 -->
00:11:56.340Uh, either you're producing or, uh, going to produce.
00:11:56.639 -->
00:11:59.789I look forward to reading a guide on how to do this.
00:11:59.789 -->
00:12:00.840It seems like it's a lot of.
00:12:01.259 -->
00:12:05.519You know, take, take workflows from here, take workflows from there, and kind of mash 'em together.
00:12:05.578 -->
00:12:05.789Yeah.
00:12:05.879 -->
00:12:11.220And um, there's like this weird, uh, kind of term that I like to call it, which is just AI intuition.
00:12:11.370 -->
00:12:11.549Yeah.
00:12:11.578 -->
00:12:15.120Just this kind of skill of knowing when and when not to use the ai.
00:12:15.149 -->
00:12:18.659'cause we're kind of at that weird point where it's not end to end yet.
00:12:18.809 -->
00:12:18.870Yeah.
00:12:18.870 -->
00:12:23.309You gotta have a human in the loop in some places and some places you can hand it over to ai.
00:12:23.490 -->
00:12:23.698Yep.
00:12:23.789 -->
00:12:28.620Um, like, have you, what do you think about AI intuition or just like that broad term of like.
00:12:29.565 -->
00:12:32.235Knowing when to have a human loop and not, and Yep.
00:12:32.264 -->
00:12:37.934Are you like training new people at HubSpot to, to have that kind of sense?
00:12:37.934 -->
00:12:43.424Or is there any kind of like courses or training, training programs you put new employees through?
00:12:43.965 -->
00:12:50.504Yeah, I mean, what is, um, and this was kind of the genesis of the inbound marketing thing that HubSpot started back in the day.
00:12:51.419 -->
00:12:54.719And this is gonna sound very meta and very philosophical, but it's still true.
00:12:54.750 -->
00:13:05.339Um, so one thing we sort of instill, not sort of, we do instill in, uh, in the team is that at the end of the day, after all is said and done, uh, our job is to create value.
00:13:05.789 -->
00:13:10.500Uh, and the reason we want to be kind of cited is because once again, just like you need to be rank worthy.
00:13:10.500 -->
00:13:11.458You need to be site worthy.
00:13:11.458 -->
00:13:16.169It's like, okay, you're not trying to trick, uh, Chad GPT into like providing your answer or whatever.
00:13:17.220 -->
00:13:23.250Do the work to come up with, do the research, whatever industry you happen to be in, to come up with, uh, really good answers.
00:13:23.250 -->
00:13:27.720And then the i and then to kind of, um, you know, automate the pieces of it.
00:13:27.720 -->
00:13:32.340That, for instance, um, I think it's well accepted for good reason.
00:13:32.549 -->
00:13:35.669You know, I use the, the LAN language, uh, translation metaphor.
00:13:35.669 -->
00:13:38.250It's like, oh, I wanna go from English to, to Japanese or whatever.
00:13:38.669 -->
00:13:42.870Um, in a large body of use cases, it's completely okay to use AI for that.
00:13:42.870 -->
00:13:45.330You should be using AI for that because it allows you to do that at scale.
00:13:45.330 -->
00:13:46.059Now is it going to be.
00:13:46.784 -->
00:13:50.294Exactly as good as a, the top-notch, uh, translator.
00:13:50.985 -->
00:13:53.384No, but you don't necessarily need that in all cases.
00:13:53.384 -->
00:13:53.565Right?
00:13:53.565 -->
00:14:02.475So this is where you're, uh, and I like that phrase a lot, that kind of building that ai intuition, but always have in the back of your mind, like, what actually does the human value?
00:14:02.774 -->
00:14:05.475Uh, this is kind of the, uh, like a golden rule thing.
00:14:05.475 -->
00:14:09.345Like how would I want technology to work if I were the human on the other side?
00:14:09.345 -->
00:14:09.524Right?
00:14:09.524 -->
00:14:15.195And, and a lot of the times right now, it's like, and you know, and the AI swap is an increasing thing that we can talk about that as well.
00:14:15.600 -->
00:14:18.690Uh, but that's, that never really works, right?
00:14:18.690 -->
00:14:24.929It's like it, uh, so back in the, in the SEO days, people would do keyword stuffing, uh, duplicate webpages.
00:14:24.929 -->
00:14:34.740They would do link circles and say, I'm gonna have 50,000, uh, you know, webpages that I spit up on these kind of crappy domains or crappy website all pointing to, because I learned that inbound links are good.
00:14:34.740 -->
00:14:37.169Well, no, inbound leagues are a proxy.
00:14:37.815 -->
00:14:39.465Uh, for a value determination.
00:14:39.465 -->
00:14:43.365And if the value is not there, then you're effectively swimming against the tide.
00:14:43.365 -->
00:14:48.195It's just a matter of time before you, uh, before you figure things, uh, before things come out.
00:14:48.195 -->
00:15:00.705So, uh, and while we're on the topic, just, uh, because I think it's a very, this is like a first principles primitive, which is, uh, back in the day, the SEO days, uh, when folks were doing, uh.
00:15:01.500 -->
00:15:10.409You know, either black hat or gray hat, SEO things, one of the arguments that the agencies or companies providing those services would say is like, oh, well we, we should do this.
00:15:10.409 -->
00:15:14.639Because the worst that's gonna happen is that it won't work and you'll get like zero return.
00:15:15.360 -->
00:15:18.870And I don't think that's a true statement because even isn't then and isn't now.
00:15:18.870 -->
00:15:21.480Because the worst that could happen is actually not.
00:15:22.230 -->
00:15:22.980Zero return.
00:15:22.980 -->
00:15:25.470The worst that can happen is negative return.
00:15:25.769 -->
00:15:37.409That says, if you build a reputation online in the algorithm's mind that you are crappy content, not trustworthy, uh, black hat, it's hard to dig yourself back out of that hole.
00:15:37.409 -->
00:15:38.879So you actually reduce your value.
00:15:38.879 -->
00:15:45.389So now you have to dig yourself out of the hole that you dug yourself into, and then add value in order to kind of get to where you, you know, want it to be.
00:15:45.389 -->
00:15:48.330So this is not, oh, we're gonna throw a bunch of things out there or whatever.
00:15:48.330 -->
00:15:50.250It doesn't really matter where that's gonna happen.
00:15:50.250 -->
00:15:50.759Is that like.
00:15:51.299 -->
00:15:52.230You know, it is zero.
00:15:52.230 -->
00:15:54.299And the same thing applies to ai.
00:15:54.299 -->
00:16:01.649So the reason we need humans in the loop is that we still, um, have a better sense for what humans want.
00:16:01.649 -->
00:16:04.769And some people call it taste, uh, instinct, whatever you call it.
00:16:05.309 -->
00:16:06.600Uh, we need to kind of tap into that.
00:16:06.600 -->
00:16:14.639That's the thing we need to be developing, is figuring out we are the things that are basically a grind that AI is as good or good enough at doing.
00:16:14.909 -->
00:16:15.299Fine.
00:16:15.480 -->
00:16:25.830And the reason to do that, um, partly is, you know, drive efficiency, but is to be able to elevate our creativity so the, the humans are spending more time on the things that humans are uniquely good at, right?
00:16:25.830 -->
00:16:31.710So let's take our time away from the grind towards the thing that, you know, creates legitimate value.
00:16:31.710 -->
00:16:32.519That's differentiated.
00:16:32.519 -->
00:16:35.250That's the, yeah, I really like that example.
00:16:35.279 -->
00:16:37.679Um, it reminded me of a tweet I saw.
00:16:37.865 -->
00:16:39.330I, I think it was, had to be.
00:16:40.289 -->
00:16:41.309At least a year ago.
00:16:41.339 -->
00:16:48.448But, um, it was that you have to be careful with your tweeting now because the training data will kind of live forever.
00:16:48.480 -->
00:16:48.659Yes.
00:16:48.719 -->
00:16:53.100Um, so if you're kind of just like, shit posting a bunch of random stuff, the LMS will remember this.
00:16:53.100 -->
00:16:53.159Yeah.
00:16:53.188 -->
00:16:58.289Uh, so that, that point that you made, that like the worst you could do is like negative.
00:16:58.289 -->
00:16:59.578It's, it's totally true.
00:16:59.850 -->
00:17:04.920Because whatever you're putting out there in the world now, it, it's getting almost remembered.
00:17:04.920 -->
00:17:06.358And the elms are training on it, right?
00:17:06.598 -->
00:17:06.659Yeah.
00:17:06.720 -->
00:17:06.993So like.
00:17:07.875 -->
00:17:11.384You do gotta be careful with what you put out, but at the same time, you gotta put out content.
00:17:11.714 -->
00:17:13.125Uh, it has to be really high quality.
00:17:13.335 -->
00:17:13.605Yes.
00:17:13.664 -->
00:17:17.384Um, so like, I think the quality bar is just going up like crazy.
00:17:17.384 -->
00:17:19.785So I I really appreciate that example.
00:17:19.875 -->
00:17:20.234Yeah.
00:17:21.315 -->
00:17:25.934A related thing to this is, um, is so the, the, the reverse is true.
00:17:25.934 -->
00:17:27.944I tend to be a, a positive leading.
00:17:27.944 -->
00:17:30.434I know you are too, uh, you know, optimistic type, but.
00:17:31.140 -->
00:17:40.349Uh, the, the good news there on that particular thing is that that, um, value and reputation, authority brand, whatever label you wanna associate to it, uh, compounds.
00:17:40.680 -->
00:17:54.630So what happened is, like, let's say you were a business and you, uh, create great content pre AI and, and, uh, and your WebP pages were out there and you were ranking because, and, and you had, uh, a, a great reputation online as a result of which you were ranking in the, the search algorithms.
00:17:54.990 -->
00:17:57.119Then perplexity comes along, and this is an actual.
00:17:57.464 -->
00:17:58.095Tangible things.
00:17:58.095 -->
00:18:08.325So, uh, perplexity, perplexity early implementation was, it was using the Bing search engine to come up with list of links of, uh, here are the possible things that will answer the question that the user is asking.
00:18:08.924 -->
00:18:12.164Pass it to an LLM and basically it did the work for you.
00:18:12.164 -->
00:18:12.345Right?
00:18:12.345 -->
00:18:22.904So it was effectively a, that was the kind of the pipeline this says, okay, user typed in this particular question I'm going to find from being the related webpages that are out there that are high quality based on the kind of old.
00:18:23.279 -->
00:18:26.009Uh, classic, uh, link structure and other things that search engines used.
00:18:26.309 -->
00:18:32.160Then I'm gonna take those web pages, I'm gonna pass them to an LLM with a, a nice prompt that then kind of comes up with the answer.
00:18:32.164 -->
00:18:33.900And, and Perplexity was brilliant at this, right?
00:18:33.960 -->
00:18:36.630Um, by the way, full disclosure, I'm, uh, friends with the founder.
00:18:36.630 -->
00:18:39.269I'm an investor in the company, but, uh, I would say these things regardless.
00:18:39.269 -->
00:18:47.730Um, and so the, the kind of point here is that whatever you did pre ai, uh, to create that value, that built that reputation.
00:18:48.134 -->
00:18:52.815Serve you well because you started showing up in perplexity as because perplexity was using Bing.
00:18:53.085 -->
00:18:55.275And so all of this is like turtles all the way up and down, right?
00:18:55.275 -->
00:18:58.664It's like, okay, well value accrues upwards and across.
00:18:58.994 -->
00:19:05.744Uh, you may reformulate, you may translate, but that brand sort of carries through and it's gonna carry through, uh, with the thing you do right now.
00:19:05.744 -->
00:19:06.704So it's, yeah.
00:19:06.734 -->
00:19:15.105Um, so the investment is worth it, I guess, in, in, in developing that human, uh, intuition, uh, around when and when not to use ai.
00:19:15.105 -->
00:19:15.555And it's like.
00:19:16.365 -->
00:19:25.365And sometime I think we have a natural sense, I think for like whether something is kind of shady or actually high value or something.
00:19:25.365 -->
00:19:27.615It's like, you know, it's, we get it right, we get it wrong.
00:19:27.615 -->
00:19:35.115But some deep down inside that I think we all have consciousness that sort of sort of know when we're sort of dancing on the edges or, you know, gone on the dark side.
00:19:35.474 -->
00:19:43.934Just avoid those temptations because in the long run it compounds that kind of goodness and, and value creation, focusing on your audience, your customers, uh, I think accrues value.
00:19:43.980 -->
00:19:45.150Yeah, totally.
00:19:45.420 -->
00:19:47.519Um, I wanna touch on Aiop a little bit.
00:19:47.519 -->
00:19:50.549I know we mentioned it a little bit before, but this is kind of related.
00:19:51.210 -->
00:19:55.440Uh, HubSpot is investing really heavily into like the podcast network.
00:19:55.440 -->
00:20:00.150This has been going on for a few years now, but you guys just have been wildly successful at it lot.
00:20:00.150 -->
00:20:04.650A lot of other companies are now investing into their own podcast networks, kind of copying your model.
00:20:04.799 -->
00:20:04.980Yep.
00:20:05.160 -->
00:20:13.170How do you think about the role of personal brands, company brands, or just content in general in, in the next.
00:20:13.513 -->
00:20:16.544You know, decade with a bunch of AI content flooding the market.
00:20:17.744 -->
00:20:18.105Yeah.
00:20:18.134 -->
00:20:21.763Uh, I think to me it's, it's, we've had variations of this.
00:20:21.763 -->
00:20:29.355It's, um, you know, in terms of, uh, inferior content, I'll call it, for lack of a better term, uh, from since the pre AI days.
00:20:29.565 -->
00:20:38.804And what we found, uh, we collectively, the industry have found is that, uh, the internet has a nose for what is actually good content over the fullness of time.
00:20:39.224 -->
00:20:41.950Um, and so the way we think about it now is that, um.
00:20:42.855 -->
00:20:48.884And we're big believers at HubSpot in like the personality behind content and the human behind the content.
00:20:48.884 -->
00:20:56.234And we think that's going to continue to be valued, hence our, uh, you know, continued investment both with our internal resources and with our external resources.
00:20:56.595 -->
00:21:09.849And so I think, um, like what I value, like my hope would be that this conversation that we're having is going to create value for, you know, uh, for many people out there and the reason they would gravitate towards this versus, uh.
00:21:10.815 -->
00:21:20.234Something that maybe had even exactly the same words or some, or maybe even the same ideas, uh, maybe they even crawled this video and created, uh, you know, a derivative film based on it.
00:21:20.234 -->
00:21:23.505It's like, it's just not the same.
00:21:23.505 -->
00:21:30.674I think we, uh, humans appreciate the craft and there are things that we enjoy, um, and, and we think that's going to continue.
00:21:30.674 -->
00:21:39.704So I think what's gonna happen and, and maybe we get, you know, um, power to steering and power tools in order to, uh, produce more content more easily.
00:21:40.138 -->
00:21:41.309Uh, which is fine.
00:21:41.309 -->
00:21:51.660It is like in more content needs to exist, but, um, but I think the, the, you know, top 1%, top 5%, uh, of the content is still gonna get rewarded for that high quality.
00:21:51.690 -->
00:21:53.549That's, that's my, my sense, uh, yeah.
00:21:54.148 -->
00:21:54.509Yeah.
00:21:54.898 -->
00:21:56.700I'm, I totally agree with that.
00:21:56.878 -->
00:22:04.019Another thing, so on top of running HubSpot, you actually coded and launched a no code agent builder called Agent ai.
00:22:04.769 -->
00:22:09.539Uh, what's the, what's the long-term vision for the platform and why did you personally go all in on it?
00:22:10.034 -->
00:22:16.454Yeah, so it started as a personal kind of nights and weekends project, um, and the idea behind it.
00:22:16.454 -->
00:22:18.464And I, I do these things sometimes.
00:22:18.464 -->
00:22:22.154I have this kind of theory in my head of where the world is going to go.
00:22:22.544 -->
00:22:24.194Uh, and I'm super impatient.
00:22:24.253 -->
00:22:28.273Um, and I have the, uh, good fortune of being able to write code.
00:22:28.335 -->
00:22:31.575I'm classically trained, uh, software developer, so I'll just go do things to try.
00:22:31.575 -->
00:22:34.154It's like, okay, well let's just see where this road leads.
00:22:34.214 -->
00:22:35.119Uh, and part of it is just.
00:22:35.849 -->
00:22:40.650And I recommend this for everyone is to, uh, to a degree that you can like, follow your curiosity.
00:22:41.009 -->
00:22:42.000Um, it's like, okay.
00:22:42.000 -->
00:22:50.759It's like, and some of the best things that have happened to me in my, in my professional career have been as a result of dots that kind of later connected that I wasn't quite sure what I was gonna do with that.
00:22:50.759 -->
00:22:51.869But then over time.
00:22:52.170 -->
00:22:53.940It's like, oh, well that connects with this, or whatever.
00:22:53.940 -->
00:22:56.339And, and uh, you know, something valuable emerges.
00:22:56.730 -->
00:23:10.109So on the agent AI front, my, my theory at the time, this was, you know, years ago now, um, is that agents were going to be the next, um, kind of evolution in, uh, the development of AI that we were used to these kind of conversational interfaces, uh.
00:23:10.470 -->
00:23:13.200But they were going to, uh, become multi-step.
00:23:13.200 -->
00:23:19.500We were going to have, um, you know, much better, uh, models, uh, which you all of that's kind of borne out to be true.
00:23:19.500 -->
00:23:21.660So, uh, so agents were gonna be a thing.
00:23:21.690 -->
00:23:22.019Okay?
00:23:22.529 -->
00:23:32.700And, and so then the, the weird thought that leapt into my head was, okay, agents are gonna be a thing and they're going to be, uh, able to take on increasingly higher order tasks so you can give them a higher order goal.
00:23:32.700 -->
00:23:33.900So right now we can say.
00:23:34.214 -->
00:23:39.914In a single prompt, we could say, write me a blog post about the rise of the Roman Roman empire and why it fell, or something like that.
00:23:39.914 -->
00:23:40.575And you can do that.
00:23:41.023 -->
00:23:52.784Uh, but what you want though is you want it to be able to say, oh, well actually I'm launching a new product, or, uh, a new event, uh, four months from now, I want you to do that for me.
00:23:53.204 -->
00:23:54.134Like, okay, well what does that mean?
00:23:54.134 -->
00:23:55.214It's like, oh, we're gonna need a website.
00:23:55.214 -->
00:23:56.174We need a social media account.
00:23:56.174 -->
00:23:57.585We need to have a social media calendar.
00:23:57.585 -->
00:23:59.595Like these are all the pieces that go into.
00:23:59.924 -->
00:24:01.335Serving that higher order goal.
00:24:01.365 -->
00:24:01.634Right?
00:24:01.694 -->
00:24:10.125Uh, and we're very, very far from being able to do that, but that's in my mind what the, uh, agents, uh, will be able to do someday.
00:24:10.154 -->
00:24:10.394Right?
00:24:10.394 -->
00:24:17.625And so the way my mental model was and is that agents, um, will effectively act like teammates.
00:24:17.684 -->
00:24:20.954Um, and they may be junior teammates, they still will need to be trained.
00:24:20.954 -->
00:24:22.095We can talk about all those things.
00:24:22.634 -->
00:24:33.855But then the first thing I kind of jumped to was like, okay, well in this kind of what I think of as a hybrid world of both, uh, carbon based life forms and, and, uh, AI agents, what does that world look like?
00:24:33.855 -->
00:24:36.434It's like, well, how are you gonna find these agents?
00:24:36.434 -->
00:24:42.704Well, right now you go to, uh, if you were gonna hire a human, you go to a place like LinkedIn, uh, it's like, okay, what's their history?
00:24:42.704 -->
00:24:43.484What's their experience?
00:24:43.484 -->
00:24:44.234What's their reputation?
00:24:44.234 -->
00:24:45.644You can see things they've posted online.
00:24:46.184 -->
00:24:49.095Uh, there needs to be an equivalent of that, uh, for agents.
00:24:49.305 -->
00:24:51.494So the very first pass@agent.ai was.
00:24:52.410 -->
00:24:59.549It is gonna be a professional network for agents, because agents are gonna need a professional network too, to be found, to demonstrate their experience, have a profile, all those things.
00:25:00.390 -->
00:25:03.119Uh, and the thing I learned subsequent to that, so that's how it started.
00:25:03.750 -->
00:25:06.599I'm like, well, there's just not enough agents out there.
00:25:06.660 -->
00:25:07.680Like, it's like, fine.
00:25:08.009 -->
00:25:12.420Like I have this site that has, uh, the ability for agents, uh, you know, kind of to be listed.
00:25:12.750 -->
00:25:14.460And so that's when I'm like, okay, well.
00:25:15.434 -->
00:25:16.785More agents should exist.
00:25:16.845 -->
00:25:20.714Um, and the reason they don't exist right now is it's just too hard, uh, to build an agent.
00:25:20.714 -->
00:25:21.674No one knows what to do.
00:25:21.674 -->
00:25:31.934So there's lots of domain experts that know about X, Y, or Z, but there's no real way to kind of encapsulate that knowledge into something that's a, uh, productizable thing that you can make available to others.
00:25:32.295 -->
00:25:38.744And so that's where the, the low code, no, no-code, uh, agent builder, uh, evolution happened, uh, for agent ai.
00:25:39.164 -->
00:25:41.805And that's going like really well, right?
00:25:41.805 -->
00:25:43.035So then, uh, and there's.
00:25:43.589 -->
00:25:45.720So you have the ability to build an agent for yourself.
00:25:45.720 -->
00:25:48.720You have the ability to share your agent that you built with the community.
00:25:48.720 -->
00:25:53.970In 2000 agents, of the 50,000 agents that have been created have been shared with the community.
00:25:54.509 -->
00:25:58.710Uh, but you know, with, uh, you know, all, with all candor.
00:25:58.710 -->
00:26:00.329I will say this, these are the early days, right?
00:26:00.329 -->
00:26:02.160So it's like the early iPhone app store days where.
00:26:02.460 -->
00:26:08.099You're gonna have things like the flashlight app all the way through the, you know, sophistication of something like, you know, Photoshop on, on the iPhone.
00:26:08.160 -->
00:26:09.690Um, and so we have something similar.
00:26:09.690 -->
00:26:13.470We have sophisticated agents and we have, you know, very simple kind of hello world style agents.
00:26:13.470 -->
00:26:15.240People are still still learning.
00:26:15.269 -->
00:26:16.980Um, so that's where we're headed.
00:26:16.980 -->
00:26:25.109The, the idea now, the vision for agent.ai is to be a platform for agents to exist because we think there are lots of these discrete tasks.
00:26:25.694 -->
00:26:27.644Uh, that we will want to give to agents.
00:26:27.674 -->
00:26:31.694Uh, someone's gonna need to build those agents and doesn't have to be built with our agent builder tool.
00:26:31.694 -->
00:26:34.394It can be we have an SDK and APIs and things like that.
00:26:34.394 -->
00:26:39.734But, um, then now what I'm spending sleepless nights thinking about, um, so that's moving.
00:26:40.214 -->
00:26:47.265Well, we sort know how to solve these kind of concrete individual tasks, but the next step is around agent to agent, uh, collaboration.
00:26:47.325 -->
00:26:48.825To say, okay, well I talked about that.
00:26:49.065 -->
00:26:50.025Oh, I'm launching an event.
00:26:50.025 -->
00:26:51.255I need to do these seven things.
00:26:51.555 -->
00:26:52.214Now, there may be.
00:26:52.710 -->
00:26:54.119Agents for those seven things.
00:26:54.690 -->
00:27:10.019Well, whoever that high order, uh, we'll call it the event manager person in your organization, if there's a agent equivalent to that, will sort of need to know how to find the agents and delegate the work and then orchestrate the work and pay the, those sub-agents to do the whatever things they are.
00:27:10.019 -->
00:27:13.740So that's going to be the, the ultimate, uh, vision.
00:27:13.740 -->
00:27:15.029This is why I'm excited about it.
00:27:15.029 -->
00:27:16.259I'm very excited about networks.
00:27:16.785 -->
00:27:22.430But imagine a, and this is what agent AI in its fullest form will be, is like imagine hundreds of thousands of agents.
00:27:23.234 -->
00:27:30.044Uh, with that are all doing specific goals, but they themselves might be using other agents on the network in order to accomplish the goal.
00:27:30.044 -->
00:27:32.054This is an idea as old as software, right?
00:27:32.054 -->
00:27:40.574That says, oh, I have these libraries or APIs or functions or things like that that I can reuse and build on the top of the, you know, it's like building on the standing of the shoulders of giants model.
00:27:40.994 -->
00:27:45.674Uh, but it, it's awfully useful for all those agents to exist on one consistent platform.
00:27:45.704 -->
00:27:50.355Uh, and we're focused on business use cases in the kind of go-to-market marketing and sales space.
00:27:50.355 -->
00:27:51.284But I think that idea.
00:27:51.763 -->
00:27:52.484Sort of holds.
00:27:52.484 -->
00:27:55.724So that's, uh, yeah, it's a really interesting thesis.
00:27:55.845 -->
00:28:05.714Um, I don't know if you've seen this, but there are like a few huge communities and YouTubers that are specifically focused on no-code agent builders.
00:28:05.714 -->
00:28:07.814Like there's one for N eight NN eight N. Yeah.
00:28:07.814 -->
00:28:10.273Think they're making like hundreds of thousands a month.
00:28:10.664 -->
00:28:10.753Yeah.
00:28:10.753 -->
00:28:11.595Teaching these workflows.
00:28:11.595 -->
00:28:17.924But agent AI is like kind of baking that community and those agents.
00:28:18.494 -->
00:28:19.694Into the platform itself.
00:28:20.023 -->
00:28:26.294Um, so like everyone can see and use and build on top of other agents.
00:28:26.294 -->
00:28:26.354Yeah.
00:28:26.384 -->
00:28:28.244So it's, it's a really interesting thesis.
00:28:28.394 -->
00:28:28.574Yeah.
00:28:28.604 -->
00:28:37.003Uh, and it sounds like you're headed into the area of like domain expert agents that can work amongst each other.
00:28:37.394 -->
00:28:37.724Yeah.
00:28:37.904 -->
00:28:40.214Uh, even if you start, you know, at the lowest level right now.
00:28:40.214 -->
00:28:41.474So I have huge amount of respect.
00:28:41.474 -->
00:28:43.449There's lots of, as you noted, uh, YouTubers.
00:28:44.309 -->
00:28:48.660Uh, kind of teaching the world how to use some of the existing low-code, no-code, uh, builders that are out there.
00:28:48.660 -->
00:28:49.680And I think that's awesome.
00:28:50.190 -->
00:28:55.289Uh, this is sort of an alter and I'm really competing with those in, in that particular way.
00:28:55.289 -->
00:29:03.089But what I'm suggesting is that maybe, um, you're gonna spend 90% of your calories on kind of teaching other people how to do that.
00:29:03.089 -->
00:29:05.609And they're usually within companies and that's awesome.
00:29:05.640 -->
00:29:07.170Um, and, and they'll make money.
00:29:07.680 -->
00:29:09.329But what if you actually built an agent?
00:29:09.329 -->
00:29:12.390So this is a little bit like a, a service business versus a product business, right?
00:29:12.390 -->
00:29:13.410It's like the service business is.
00:29:13.875 -->
00:29:16.634Oh, I'm gonna help people to do the thing or whatever and hold their hand or teach them.
00:29:16.694 -->
00:29:17.474Uh, that's great.
00:29:17.894 -->
00:29:26.265Uh, but the software business side of it says, well, maybe you have a few ideas, uh, in your industry or something you have expertise in that you could actually build an agent and sell it.
00:29:26.295 -->
00:29:35.295Similar to the app store model, it's like, uh, so this could be the kind of agent stores is the idea is to be able to monetize, uh, monetize the agents and, um, and put things out there.
00:29:35.295 -->
00:29:38.174So when we have, so there's two ends of the spectrum here.
00:29:38.174 -->
00:29:41.295We have the, um, kind of workflow, deterministic workflow builders.
00:29:41.669 -->
00:29:46.289Uh, like NNN make, uh, there, you know, some great ones out there, uh, of the other spectrum.
00:29:46.349 -->
00:29:50.099Um, in terms of non-deterministic things, you have custom gpt.
00:29:50.128 -->
00:29:58.829Uh, you have, uh, quad projects, you have, uh, Gemini Gems, uh, and, and they allow you to effectively do automation types of things.
00:29:58.829 -->
00:30:03.059And obviously the alls are getting better with reasoning and things like that so they can, uh, you know, do better things.
00:30:03.059 -->
00:30:07.559But something, it feels like there's something needs to be something sort of in the middle.
00:30:07.559 -->
00:30:09.059So I think custom gpt are great.
00:30:09.059 -->
00:30:10.289I use them, you know, all the time.
00:30:10.289 -->
00:30:10.829But, um.
00:30:11.625 -->
00:30:22.244But it's not really good at not, and and they've taken a couple of passes at this, but I don't think it's in their core, but it's like, oh, I actually wanna build a combination of things that straddles models.
00:30:22.275 -->
00:30:30.134Um, and, you know, it can bring the best of what AI has to offer in order, uh, for the manifestation of this idea that I have in this domain expertise that I have.
00:30:30.134 -->
00:30:30.795So, we'll see.
00:30:30.799 -->
00:30:32.565It's, uh, it's a fun, fun world.
00:30:33.224 -->
00:30:36.704I also did a lot of research of course, and I've tried agent AI many times.
00:30:36.704 -->
00:30:40.214And another thing that's interesting is you're making so much of the platform free.
00:30:40.680 -->
00:30:40.859Yep.
00:30:40.890 -->
00:30:43.829Is there like a, a mental model behind that?
00:30:43.829 -->
00:30:45.750Is there a plan for monetization?
00:30:46.170 -->
00:30:47.190Uh, how do you think about that?
00:30:47.460 -->
00:30:48.000Yeah.
00:30:48.029 -->
00:31:02.849Um, so we have the same mo that, uh, HubSpot has had, you know, we're HubSpot, the company is 19 years old now, and at, at the core of our, uh, approach to business is that you have to first focus on creating value, uh, before you look to extract value.
00:31:03.029 -->
00:31:05.305It's like, okay, like, you know, like.
00:31:05.878 -->
00:31:08.819So our idea here is like, okay, so we've got this thesis about the world.
00:31:08.819 -->
00:31:17.939We're gonna build out a product, um, and we want, and HubSpot, when it originally started, um, we had, you know, free tools, um, and, and a free version.
00:31:18.990 -->
00:31:21.750And the idea here is we wanna democratize this technology.
00:31:21.779 -->
00:31:34.023Um, you know, at the time we were talking about 19 years ago, uh, we're doing the same approach now that says we, so, uh, I'll, you know, brief history, when HubSpot started the, the fundamental thesis was, hey, there's this thing called the internet.
00:31:34.634 -->
00:31:38.055That millions of small businesses, uh, should be benefiting from, but they're not.
00:31:38.174 -->
00:31:40.454And the reason they're not is because it's scary.
00:31:40.755 -->
00:31:44.384It's not because there's not great tools out there for content management and web analytics.
00:31:44.384 -->
00:31:47.565And for SEO, there were lots of great tools and great companies behind those tools.
00:31:48.015 -->
00:31:55.845But for a mere mortal company, if you're a 25 person law firm, it was very daunting to say like, they don't even know what Woo Fu is.
00:31:55.845 -->
00:31:58.035Or we know all the great tools from those times, whatever.
00:31:58.035 -->
00:32:00.164So our thing was like, how do we.
00:32:01.380 -->
00:32:02.339Make it approachable.
00:32:02.339 -->
00:32:07.529How do we democratize this thing called the internet and make it accessible to millions of small businesses?
00:32:07.529 -->
00:32:10.589That is and still is, uh, wasn't, is, uh, our, our mission.
00:32:11.279 -->
00:32:18.720We're doing the same thing with AI that says, okay, there's this new thing called AI that millions of small businesses should be benefiting from, but not enough are.
00:32:18.750 -->
00:32:23.130Yes, they'll use Chatt PT and they'll use kind on the very, very fringes of it, but they're not really digging in yet.
00:32:23.700 -->
00:32:25.890What do we need to do in order to kind of get them into the fold?
00:32:25.890 -->
00:32:30.119And agents are kind of next layer of abstraction and I think people are used to the idea.
00:32:30.509 -->
00:32:32.519Of using AI for content generation.
00:32:32.640 -->
00:32:33.480Uh, simple things.
00:32:33.480 -->
00:32:36.869Write me a blog post, generate an image for me, kind of single step tasks.
00:32:36.900 -->
00:32:37.230Great.
00:32:37.259 -->
00:32:39.750Uh, I think consumers and businesses are comfortable with that.
00:32:40.230 -->
00:32:47.460But as we start thinking about agents that requires, um, more sophistication, how do I find the agents who maintains them?
00:32:47.819 -->
00:32:49.380What is training an agent?
00:32:49.380 -->
00:32:49.984What does that look like?
00:32:50.005 -->
00:32:54.690All of these things are new and so our hope is that, um, we sort of bring.
00:32:55.214 -->
00:32:59.023Got SMBs up and start with making it completely approachable.
00:32:59.023 -->
00:33:04.095And the easiest way to, to make something approachable is to remove all barriers and all friction, one of which is price.
00:33:04.365 -->
00:33:07.424Uh, and free is the easiest price to, to, you know, try things out.
00:33:07.664 -->
00:33:09.763And the second one is to make things easy.
00:33:10.154 -->
00:33:12.974It's like, okay, well don't try to bow the ocean to have this super sophisticated thing.
00:33:12.974 -->
00:33:17.054And there are lots of great sophisticated, uh, kind of agent, uh, and AI products out there.
00:33:17.054 -->
00:33:17.500We want to go.
00:33:18.359 -->
00:33:23.339Like the other end and do what we do best, which is educate, uh, simplify and make things approachable.
00:33:23.519 -->
00:33:24.390And we will monetize.
00:33:24.390 -->
00:33:25.740We are a for-profit company.
00:33:25.769 -->
00:33:30.180Uh, you know, we're red blooded capitalists, but warm hearted, uh, red blood capitalists.
00:33:30.660 -->
00:33:33.059Uh, and so agent AI will have monetization, right?
00:33:33.059 -->
00:33:36.000And that's the only way third parties are gonna be able to, to make money.
00:33:36.450 -->
00:33:40.349Uh, but we were, you know, and the original idea was around creating this kind of network, right?
00:33:40.349 -->
00:33:41.160A marketplace.
00:33:41.700 -->
00:33:44.849Um, and we would do this anyway, but it was particularly true.
00:33:44.849 -->
00:33:45.404It's like, okay, well.
00:33:46.769 -->
00:33:51.269Agent builders are gonna want to go to where the consumers are, like where their potential audience would be, right?
00:33:51.299 -->
00:33:56.849Um, and the, um, the consumers want to go to where the agents are.
00:33:56.849 -->
00:34:01.019It's the classic marketplace, um, kind of chicken egg problem.
00:34:01.470 -->
00:34:06.359And so now we have, you know, 2 million plus users on agent.ai, uh, still growing very, very quickly.
00:34:06.599 -->
00:34:10.889So when we go to market, um, with a kinda monetized offering, we'll say, Hey.
00:34:12.105 -->
00:34:21.974We have a bunch of free agents, uh, a lot of them community built, a lot of them we've built, we're gonna have a handful of agents that we think, um, are what we'll call premium agents, uh, that will have a price tag attached to them.
00:34:22.153 -->
00:34:24.673And we'll let third parties build pre premium agents as well.
00:34:24.673 -->
00:34:26.233And they can still put community agents out there.
00:34:26.233 -->
00:34:26.713Fine.
00:34:26.804 -->
00:34:28.364They can have commercial agents as well.
00:34:28.693 -->
00:34:29.684That's, that's where we're headed.
00:34:29.983 -->
00:34:33.793Uh, but we, we felt like, uh, the product wasn't good enough.
00:34:34.184 -->
00:34:37.063We hadn't learned enough and we hadn't built enough momentum.
00:34:37.829 -->
00:34:40.829To be able to kind of earn our seat at that particular table.
00:34:40.829 -->
00:34:45.360And I now, I think we're, we're, we're there, um, we have enough users, we've got the product.
00:34:45.389 -->
00:34:47.909Uh, and my apologies for the prior times you tried.
00:34:47.909 -->
00:34:48.900It's getting better now.
00:34:48.929 -->
00:34:50.969Um, so, but yeah, it's, it's, it's software.
00:34:51.059 -->
00:34:55.289So on, uh, making AI and agents accessible.
00:34:56.610 -->
00:35:00.780So you obviously see, you see 2 million users every day trying and using these agents.
00:35:01.139 -->
00:35:03.599Are there any like, patterns that you're seeing and.
00:35:04.199 -->
00:35:13.079Normal people that are succeeding with these AI and agents and using them for work effectively versus people who maybe try it once and don't try it again.
00:35:13.469 -->
00:35:28.349Yeah, I think the, the, what we've learned is that a lot of people, so we, you know, we have in any given month, like a hundred thousand people, new people that will show up on agent.ai and the patterns that we're seeing is that.
00:35:29.954 -->
00:35:34.844They're sort of the kind of early adopter types because they've heard of agents, um, mean it's not so mainstream.
00:35:34.844 -->
00:35:40.963Like we within the tech industry, talk about agents and things like that all the time for, for, uh, the rest of society.
00:35:40.994 -->
00:35:42.253Um, it's still very, very new.
00:35:42.673 -->
00:35:49.364And so the way we we approach it is, is very much around the education that says, oh, uh, you've used chat GPT and things like that.
00:35:49.364 -->
00:35:56.648Find areas of your bi like, here are things that people like you, if you're working in marketing, you're likely using it to write a blog post, but have you considered.
00:35:57.659 -->
00:35:59.190Taking it a couple of steps forward.
00:35:59.190 -->
00:36:06.900It's like, oh, I'm gonna write the blog post that I'm gonna auto generate, uh, the social media posts around it that takes a snippet from that or whatever, and, and carry that forward.
00:36:06.900 -->
00:36:17.369So if you can take something that's already familiar and say, and maybe they're doing pieces of it, but they're doing, like, I'm putting something in the chat GPT that I'm copying that, pasting that into something else to do or whatever.
00:36:17.699 -->
00:36:22.889It's like those are the kinds of things that are perfect for some sort of age agentic, you know, automated approach.
00:36:23.280 -->
00:36:23.985And that's where we start.
00:36:24.735 -->
00:36:39.885Uh, and the second level, um, is around, uh, what I, what we call, uh, knowledge agents that says, okay, you probably have, um, a set of PDFs, blog articles, webpages, videos, or things like that, that encapsulate the knowledge that you have in your organization.
00:36:40.394 -->
00:36:49.005Um, and you want to be able to kind of, um, make that easily and readily available either for yourself, for your own personal use, or for your, for your team, whatever that might be.
00:36:49.485 -->
00:36:50.744And there are ways to solve that, right?
00:36:50.744 -->
00:36:54.735There's custom gpt and things to be able to create, you know, rag based AI pipeline.
00:36:55.364 -->
00:36:59.835But just the fact that I said, uh, rag and, you know, AI pipeline already will throw people off.
00:36:59.835 -->
00:37:03.315So we just take it the other way that says, build a knowledge agent.
00:37:03.375 -->
00:37:04.454Here's all that's involved.
00:37:04.454 -->
00:37:05.775Here are five videos on how to do it.
00:37:05.835 -->
00:37:08.085Uh, we hear some examples that you can start with.
00:37:08.505 -->
00:37:14.414Um, and then they sort of see the power of a, of a kind of.
00:37:14.894 -->
00:37:16.333Somewhat trained model.
00:37:16.333 -->
00:37:20.083They trained, it's like, and the training that happens by virtue of knowledge they already have.
00:37:20.474 -->
00:37:28.483Um, and so we, we see it lot for like, marketers, like it for, uh, like something that will adhere to brand guidelines and things like that, that says, oh, here's our tone of voice.
00:37:28.483 -->
00:37:29.443Here's our brand guidelines.
00:37:29.744 -->
00:37:34.273Now some say, well, we could just give them a custom prompt or, you know, and, and they would do that.
00:37:34.273 -->
00:37:41.684But it's like, okay, well you want that, but then you also want the generation, but then you also want the brand guidelines around the social media side of it, and you wanna pull all of that together.
00:37:42.389 -->
00:37:49.980That's the kind of thing that um, you sort of have this, I'll call it the agent maturity curve, um, that you start simple.
00:37:50.010 -->
00:37:53.730Um, and my advice is always to kind of start deterministic.
00:37:53.760 -->
00:38:03.719Uh, because you know, the first thing and I was, uh, I'm still somewhat apologetic'cause I use the word agent, um, very liberally in terms of what's an agent.
00:38:03.719 -->
00:38:07.860And some would argue an agent has to have autonomy, has to have non-determinism.
00:38:07.860 -->
00:38:10.349It has to be actually making choices and doing things on its own.
00:38:11.010 -->
00:38:11.760I'm like, sure.
00:38:12.150 -->
00:38:16.769That's one end of the spectrum of agents, but there's another end of the spectrum of the agents.
00:38:16.769 -->
00:38:18.480It's just software that uses ai.
00:38:18.480 -->
00:38:24.780And it happens, uh, to be, even if it's like a single step or two steps or whatever, to me it's still an agent.
00:38:24.809 -->
00:38:32.639Uh, and the, the metaphor I like is, um, in just like organisms in, in biology, you can have, uh, very sophisticated organisms like humans.
00:38:33.360 -->
00:38:34.889But you can have single celled organisms.
00:38:34.920 -->
00:38:35.610Uh, right.
00:38:35.610 -->
00:38:37.349It's like they're very, very simple.
00:38:37.380 -->
00:38:39.239Um, they're still organisms, right?
00:38:39.239 -->
00:38:41.250So the, the kind of primitives are still there.
00:38:41.280 -->
00:38:43.500It still qualifies as a, as an organism.
00:38:43.500 -->
00:38:47.159It's just not, um, as sophisticated as, uh, as a mammal might be.
00:38:47.159 -->
00:38:57.659Anyway, so I guess my, my, my larger point here is that, uh, my advice to, to new folks, and this is embedded inside of agent ai, is like, don't overthink it.
00:38:57.659 -->
00:38:59.070This is not supposed to be scary.
00:38:59.369 -->
00:39:00.780It's supposed to be useful, right?
00:39:00.780 -->
00:39:02.085And, and start, it's like, oh.
00:39:03.014 -->
00:39:08.625Even if you have three deterministic steps and you know what the steps are, don't try to coerce a reasoning model.
00:39:08.625 -->
00:39:13.635However smart, to kind of figure out those, like I know the steps that I need to do to accomplish this particular goal.
00:39:13.965 -->
00:39:23.204And maybe one of those steps involves, or all of them involve calling an lm, that's fine, but don't try to hand over too much too soon on, on the agent definition parts.
00:39:23.324 -->
00:39:30.704Um, I think we're at a really interesting time too, because, you know, there are some agents that.
00:39:31.275 -->
00:39:35.054Work somewhat reliably that are, uh, non-deterministic.
00:39:35.264 -->
00:39:37.335Um, but they're not perfect yet.
00:39:37.605 -->
00:39:37.755Yep.
00:39:37.784 -->
00:39:38.235Right.
00:39:38.295 -->
00:39:48.344Um, so just in your opinion, or based on what you've seen, like, what do you think the current bottlenecks are that are just stopping these agents from going full blown autonomous?
00:39:48.344 -->
00:39:48.403Yeah.
00:39:48.465 -->
00:39:57.344And um, kind of work, like you said, like working together in the network and, you know, maybe working with multiple agents, like what's, what's stopping us from getting there?
00:39:57.690 -->
00:40:05.280Yeah, so I've, you know, a lot of thoughts on that because HubSpot has, so, you know, we have agent ai, obviously, um, very, very simple agents.
00:40:05.280 -->
00:40:07.110Um, it's on one end of the spectrum.
00:40:07.590 -->
00:40:15.539On the other end of the spectrum, we have, uh, a customer agent, which is for customer support that businesses use to be the frontline support, like customer facing.
00:40:15.539 -->
00:40:16.530It will answer questions.
00:40:16.980 -->
00:40:23.820Um, and, and we measure, you know, csat, you know, are, are the people happy when they get responses from, uh, you know, from the agent?
00:40:23.820 -->
00:40:24.719What's the resolution rate?
00:40:24.719 -->
00:40:25.349What percentage of.
00:40:25.769 -->
00:40:27.599Questions, is the agent able to answer?
00:40:27.599 -->
00:40:41.159And, and like anything else, it's like if it's kind of, uh, a big believer in iteration, uh, kind of measure iteration that says if you can go through a loop and you know, at the end of a loop whether you got better or worse, uh, you are going to win eventually.
00:40:41.188 -->
00:40:41.369Right?
00:40:41.369 -->
00:40:48.059It's like, it's like something, any system that can be improved iteratively over some period of time, uh, will be good enough by some, by some definition.
00:40:48.358 -->
00:40:51.599And so the customer support agents are effectively there.
00:40:51.599 -->
00:40:52.673So we have thousands of customers.
00:40:53.460 -->
00:40:56.969That put customer agent in, in front of their customers and answers questions.
00:40:57.389 -->
00:41:00.420Uh, usually the holdup, uh, is a couple things.
00:41:00.420 -->
00:41:11.369One is, um, they don't think about in the fullest sense of the term, uh, an agent as being, which is exactly what it is.
00:41:11.369 -->
00:41:20.375So if you think about it, you know, back to my earlier, uh, comments is that if you think of agents as, uh, teammates, which is they're going to become, but even if you, today, you thought of them as, as teammates.
00:41:21.269 -->
00:41:22.528You would not hire a teammate.
00:41:22.588 -->
00:41:28.318Let's say you were gonna hire an intern, graduated not from one of the top schools, but all the top schools that has a PhD in everything.
00:41:28.318 -->
00:41:30.269Great, brilliant, brilliant person.
00:41:31.048 -->
00:41:37.019You would not just hand them a computer and access to all the, the internals of the organization, say, here's what we want done.
00:41:37.019 -->
00:41:37.438Go do it.
00:41:37.500 -->
00:41:37.679Right?
00:41:37.679 -->
00:41:39.179It's like, that would be foolish.
00:41:39.179 -->
00:41:41.543We wouldn't do that regardless of how smart that person happened to be.
00:41:42.210 -->
00:41:43.559The same is true for agents.
00:41:43.559 -->
00:41:49.980It's like, okay, well agents need to be trained, uh, they need to be tested, they need to have the equivalent of performance reviews.
00:41:50.009 -->
00:41:55.440For instance, if you hired that intern, uh, that was brilliant, you would say, oh, maybe after week one or whatever.
00:41:55.440 -->
00:41:58.409It's like, okay, they write their first thing or produce their first video or do whatever.
00:41:58.409 -->
00:42:00.719It's like, let's sit down, let tell you the things you got right.
00:42:00.719 -->
00:42:03.360Lemme tell you the things you've got wrong because you want them to get better.
00:42:03.389 -->
00:42:08.068Um, and, and that's, and that's how you make agents actually work, is treat them.
00:42:08.474 -->
00:42:09.974Uh, as if they were teammates.
00:42:09.974 -->
00:42:13.634It's like you have to make the investment in training and customization and things like that.
00:42:14.085 -->
00:42:15.255Uh, and there's a balance here.
00:42:15.255 -->
00:42:20.684So if it's a relatively low risk use case, fine, you can throw things at it.
00:42:20.684 -->
00:42:29.864For instance, let's say you were gonna use a customer agent that was only available to your own internal support people as a additional resources and additions to documents and things like that, that they have.
00:42:29.864 -->
00:42:36.525It's like, oh, as you're answering customer questions or whatever, you have this tool over here that you can chat with and it can help you answer the customer's question.
00:42:36.525 -->
00:42:37.905But there's still a human in the loop.
00:42:38.804 -->
00:42:40.815Totally fine, uh, with taking that approach.
00:42:40.844 -->
00:42:51.614Uh, so then you can sort of build your comfort level, but over time you will find that, um, you sort of get out of agents, and this is true for AI more broadly, sort of what you put into it.
00:42:51.614 -->
00:42:58.934And we sort of want this immediate silver bullet kind of thing that says, oh, I, I bought this customer agent that's supposed to enhance all my customer queries.
00:42:59.355 -->
00:43:01.065It's like, well, it's not magic.
00:43:01.065 -->
00:43:02.175So it's, it knows.
00:43:02.639 -->
00:43:09.420Everything about everything, uh, Newtonian physics and Shakespearean sonnets or whatever, and it knows nothing about your business other than what's publicly accessible.
00:43:10.110 -->
00:43:14.519And if it can answer questions for customers based on the publicly available information, awesome.
00:43:14.550 -->
00:43:17.639You'll, you'll, you'll get your, your mileage will vary, but it can do that.
00:43:18.389 -->
00:43:20.760But it knows nothing about your internal business or whatever.
00:43:20.760 -->
00:43:27.239So it's, uh, kind of foolish to expect it to be able to do things, uh, because it was not trained on any of that, any of that knowledge.
00:43:27.239 -->
00:43:28.079So that's the, yeah.
00:43:29.324 -->
00:43:32.293By the way, a good, uh, uh, case in terms of agents working together.
00:43:32.293 -->
00:43:53.744We're seeing the, uh, beginnings of this on the, uh, agent side of this, uh, customer agent side of the spectrum is that, so we have a, uh, a knowledge base agent within HubSpot, and the idea there is like we have a knowledge base, um, that you write up articles or whatever, and historically that product, the original knowledge base product got created in order for businesses to have a knowledge base online that says, oh, here's we're a software company.
00:43:53.744 -->
00:43:57.523Here are the common things that, how to do x and do wire solve this issue.
00:43:57.900 -->
00:43:58.769Uh, super useful.
00:43:58.798 -->
00:44:08.519Um, and now we kind of feed that knowledge base to the, uh, customer agent that says, okay, well you've already invested all this time in, in producing 500 knowledge base articles.
00:44:08.730 -->
00:44:11.278That's perfect information to give to the customer agent.
00:44:11.369 -->
00:44:12.210Okay, that makes sense.
00:44:12.869 -->
00:44:15.838Now, what's cool though is that when the customer agent fails.
00:44:16.590 -->
00:44:18.480To, uh, answer a question.
00:44:18.480 -->
00:44:30.539It's like, oh, I don't know the answer to that within, with a sufficient degree of confidence, instead of just saying, I don't know that it does that, and we will kind of escalate to a human, but then it will kick off the knowledge base agent and say, Hey, here came this question.
00:44:31.199 -->
00:44:33.568It didn't have, uh, I didn't have an answer.
00:44:33.778 -->
00:44:39.269Here's a draft response to the question as a knowledge base article that goes into a human for review.
00:44:39.750 -->
00:44:44.009So we can kind of grow the knowledge base as a result of what we're learning, uh, from inbound data.
00:44:44.009 -->
00:44:44.789And the cool thing is that.
00:44:46.125 -->
00:44:56.264The way the knowledge base agent can kind of write its responses is that it has access to every customer support call, all the support emails that we've had or that customer has had over, you know, a long period of time.
00:44:56.264 -->
00:44:58.335So all the dots start to, to connect.
00:44:58.335 -->
00:45:02.715It's like, oh, well, you know, that investment you made in, in knowledge, that's great because now your customer agents informed.
00:45:02.715 -->
00:45:04.215It's like, oh, you know, it's like that.
00:45:04.215 -->
00:45:08.655Remember you we're logging all your emails and your, you know, call calls and everything like that.
00:45:08.655 -->
00:45:13.724We have like a textual transcript that can now be used to inform knowledge base, article authoring.
00:45:13.755 -->
00:45:14.744Um, anyway, so it's all.
00:45:15.329 -->
00:45:26.278So, so something I can't stop thinking about, this is like a, something I'm personally dealing with every day is like, how do you build a startup from the ground up in the AI era, right?
00:45:26.278 -->
00:45:32.219Like, when, when do you hire humans versus when do you hire an agent and use an AI for that?
00:45:32.278 -->
00:45:34.980Or how do you empower human with an agent?
00:45:35.219 -->
00:45:41.730And, you know, talking with you, it's, it's really interesting 'cause you're actually, I mean, you're very successful and you started HubSpot.
00:45:42.599 -->
00:45:45.989You're actually kind of building agent AI and you're building the team from the ground up.
00:45:46.380 -->
00:45:51.809Are you doing anything differently, uh, with a building, the agent, AI team versus when you started HubSpot?
00:45:51.840 -->
00:45:54.000Like, are, are you hiring slower?
00:45:54.000 -->
00:45:55.110Are you keeping it leaner?
00:45:55.110 -->
00:45:57.929Like what are, what are some tactical things you're doing with the team?
00:45:58.230 -->
00:45:58.500Yeah.
00:45:58.500 -->
00:46:06.000Um, I'm a believer in hiring slowly, even in the pre AI world, but the tactically, um, the thing we're sort of, um.
00:46:07.500 -->
00:46:13.320Things that I say are different is like we're solving more for generalists, um, than we would for specialists.
00:46:13.590 -->
00:46:23.519'cause over time, I think what's gonna happen is that the very, very deep, uh, kinda specialized skill that you couldn't, you needed but couldn't get to any other way other than hiring a human.
00:46:23.849 -->
00:46:26.670I think that's going to be, um, less needed.
00:46:26.730 -->
00:46:27.000Right?
00:46:27.150 -->
00:46:28.800Not, not needed, but less needed.
00:46:28.800 -->
00:46:30.869Especially in, in the world of a startup.
00:46:30.900 -->
00:46:31.454Um, where.
00:46:33.014 -->
00:46:35.293You know, resources are, you know, very, very scanned.
00:46:35.293 -->
00:46:58.903So what you want is you want people, um, that get us, get obsessed with the customer problem, uh, which is the thing why, you know, the, the company exists or the project exists, uh, but you want them to be, like, if I were to solve for like one kind of, I'll say trait, um, or, or skill, it would be curiosity, uh, is like, it's like, oh, okay, I, I've got this thing that I'm doing or I need to do, or whatever.
00:46:58.903 -->
00:46:59.083It's like.
00:46:59.744 -->
00:47:03.675You know, will they kind of lean towards like, okay, well I've heard AI is pretty cool.
00:47:03.675 -->
00:47:06.824I've heard agents are pretty cool and it can be our tools or other people's tools, we don't care.
00:47:07.275 -->
00:47:12.105Like they are sort of intrinsically bothered by inefficiency.
00:47:12.135 -->
00:47:15.494Uh, and they're curious enough to see if it's, it is possible to do something.
00:47:15.554 -->
00:47:23.324Um, and so we're looking for those, um, very customer focused, uh, generalists that are curious and have this, uh.
00:47:23.894 -->
00:47:27.795Kind of iterative approach, just like, um, we're gonna tr a try it, maybe it fails.
00:47:27.795 -->
00:47:39.855Uh, one of the pieces of tactical advice I give both to the team and the world at large is that, uh, anytime you're trying to do something, uh, whatever it might be, um, or solve a problem, do something, give AI a shot first.
00:47:39.855 -->
00:47:43.454Like, pick your favorite AI two oh choice, be it chat, GP PT Quad, or whatever.
00:47:43.454 -->
00:47:44.414Just, just try it.
00:47:44.835 -->
00:47:49.815Um, and if it fails, which it likely will based on what you're trying, um, fine.
00:47:51.313 -->
00:47:55.543Give yourself like a six month reminder from now to retry the same thing.
00:47:55.574 -->
00:47:59.503Because I think a mistake people make is like, oh, I tried AI for that.
00:48:00.014 -->
00:48:01.094It didn't work at all.
00:48:01.094 -->
00:48:03.403Like it failed miserably in crashing, burning flames.
00:48:03.824 -->
00:48:04.634It's like, okay, fine.
00:48:04.875 -->
00:48:17.114One of the things we sort of have a hard time understanding is humans, when we're on an exponential curve, uh, which we are with LMS and, and just their degree, just like we were with computers before that, but even more so, um.
00:48:17.849 -->
00:48:18.480No pun intended.
00:48:18.480 -->
00:48:22.260More so anyway, um, is things change just so fast.
00:48:22.260 -->
00:48:30.119So something that, you know, AI was not capable of today is entirely possible that it will be capable of, uh, you know, to a certain level of quality, you know, six months from now.
00:48:30.119 -->
00:48:30.989So don't just give up.
00:48:31.500 -->
00:48:37.079And it is like, if you're looking for reasons not to use ai, you will find a million reasons is not to use ai, right?
00:48:37.079 -->
00:48:44.409It's like that's, um, but kind of try it, have this kind of childlike curiosity looking, you know, my, uh, you know, my son is like.
00:48:45.119 -->
00:48:48.420Well, why wouldn't AI be able to do that like it thinks it's all powerful, all knowing.
00:48:48.480 -->
00:48:53.579Like, it's like, and, and you'll skin his knees, like, ah, it's not really good at that, and you'll move on to the next thing.
00:48:53.579 -->
00:48:54.719But at least try it.
00:48:54.900 -->
00:49:03.509That's, uh, anyway, so smaller teams, uh, more kind of generalist teams that we can kind of move around because we're not exactly sure, uh, what the thing looks like.
00:49:03.539 -->
00:49:08.548Um, because the product itself is changing, because the technology is trying to use is changing.
00:49:08.548 -->
00:49:12.989So we've got multiple axes of, uh, you know, really rapid, uh, rapid evolution.
00:49:14.235 -->
00:49:22.454At the end of the day though, it's um, like the way to, I think build a successful startup, um, hasn't changed all that much.
00:49:22.454 -->
00:49:28.695You still need to be, um, very customer obsessed, uh, you know, obsessed over the problem, not your particular solution.
00:49:28.695 -->
00:49:33.255All the things that were true five years ago, I think, um, most of them are still, still very true.
00:49:33.315 -->
00:49:35.534Um, but you have a new tool set just.
00:49:35.894 -->
00:49:39.463Like, you know, before the internet, um, you didn't have the internet, but then you did.
00:49:39.463 -->
00:49:43.605So you'd be an idiot not to use the internet, both in building your product and getting access to information.
00:49:44.144 -->
00:49:48.974AI is the same thing, like find people that will use the best available tools to solve the problem at hand.
00:49:49.934 -->
00:49:54.313Uh, before we move on, I just wanna like get that piece out again.
00:49:54.403 -->
00:49:54.554Yeah.
00:49:54.554 -->
00:49:55.394And I wanna break it down.
00:49:55.394 -->
00:49:55.483Sure.
00:49:55.514 -->
00:50:01.605So what you're saying is if someone, you know is curious, that's the, the first step is to be curious.
00:50:01.693 -->
00:50:04.125Um, and the second step is to just like.
00:50:04.634 -->
00:50:09.465Take your work problems, whatever it is to be a really hard problem.
00:50:09.614 -->
00:50:10.815It could be something simple.
00:50:10.965 -->
00:50:11.114Yep.
00:50:11.173 -->
00:50:14.025And then just ask chat, bt if you can do it or solve it.
00:50:14.173 -->
00:50:14.264Yep.
00:50:14.264 -->
00:50:19.454And then if it can't remind yourself again, is that like kind of the, the strategy there or, yeah.
00:50:19.485 -->
00:50:23.144So that, that's like the, you know, kind of single shot and, uh, time delayed.
00:50:23.653 -->
00:50:25.784Uh, so two takeaway messages here.
00:50:25.784 -->
00:50:27.585One is have the curiosity to want to know.
00:50:28.260 -->
00:50:35.309Um, have that experimenting experimental mindset to say, I'm gonna try AI to see if it can help me solve this problem.
00:50:35.789 -->
00:50:47.699Um, and then, and it can vary in terms of how often you try based on what the problem happens to be, but at least, you know, six months output, a calendar entry out there, because lots will have changed by then, um, is to be able to kind of try it iteratively.
00:50:47.820 -->
00:50:56.849But, uh, the other thing I, and I, I, I would recommend is, uh, and this is true for, uh, you know, software in general and problem solving is.
00:50:57.315 -->
00:51:07.034Is try to get, if I had to pick one skill, by the way, outside of curiosity, like a, a super tactical skill, it is useful, is what engineers would call, uh, functional decomposition.
00:51:07.094 -->
00:51:19.034And the idea here is whatever large problem you have, if you can articulate that problem as a, uh, broken down piece of, it's like, okay, if I could solve this, this, and this, then this larger problem would be solved.
00:51:20.054 -->
00:51:20.414And that's it.
00:51:20.414 -->
00:51:25.034That's like, that's the first thing, like first order, uh, past that, it's like, okay, well now treat each of those individual parts.
00:51:25.034 -->
00:51:26.835Like, oh, I have this problem to solve.
00:51:26.954 -->
00:51:29.295If I could solve these two problems, this particular problem would be solved.
00:51:29.295 -->
00:51:29.474Right?
00:51:29.474 -->
00:51:31.844And, and we, this applies to so much in life.
00:51:31.875 -->
00:51:37.664For instance, in, uh, when you're building a company, you have a, you have an org chart, uh, however big or small the company is.
00:51:37.664 -->
00:51:42.375However big or smaller team is like, oh, well we need a, a, a VP of marketing.
00:51:42.614 -->
00:51:42.974Why?
00:51:43.409 -->
00:51:48.059Okay, well we need a VP of marketing because we need to do lead gen. We need to do social media, and we need to do a EO.
00:51:48.119 -->
00:51:49.530Let's just say whatever list of things.
00:51:49.889 -->
00:51:56.159Okay, so you're saying if we hire someone like this and we hire someone like that and we hire someone like that, then our marketing thing would be solved.
00:51:56.190 -->
00:51:59.010Okay, now for the a EO, like what do you need to do a EO?
00:51:59.010 -->
00:52:02.730It's like, oh, we need to do this, this, this, and the other thing, whatever is you're constantly decomposing it.
00:52:02.969 -->
00:52:10.769And what you will find is that once you go deep enough into it, uh, problems will kind of fall into one of three buckets.
00:52:11.713 -->
00:52:12.434That's trivial.
00:52:12.463 -->
00:52:14.534It's like, that's not, that's not even a thing anymore, right?
00:52:14.534 -->
00:52:24.614It's like, oh, you have to do a Google search in order to solve this particular, based on how deep down the stack you go, another bucket of problems that's now new is like, oh, well that's actually solvable, uh, by ai.
00:52:24.614 -->
00:52:28.213Or we can build a workflow or an agent or an automation or something like that to do it.
00:52:28.603 -->
00:52:37.903But that idea of breaking large problems down into its constituent pieces and recursively going through that, it's like, okay, well then it's just, any problem is solvable.
00:52:37.903 -->
00:52:40.454If you have that ability to say, okay, well, what does this really mean?
00:52:40.994 -->
00:52:48.195To solve X, can I make it a, uh, decomposed function of this that involves these three things or whatever number of things.
00:52:48.644 -->
00:52:57.255Something else I noticed that you do differently than any other founder or billionaire I've seen is you still code yourself a lot.
00:52:57.675 -->
00:53:01.034Uh, like more than any other major founder.
00:53:01.153 -->
00:53:03.434It almost seems to me like it's.
00:53:04.275 -->
00:53:05.385It's gotta be a waste of your time.
00:53:05.385 -->
00:53:07.664You gotta be, you could hire so many people to do that.
00:53:07.905 -->
00:53:10.784Is there a reason why you're like deep in the trenches?
00:53:11.534 -->
00:53:12.525Um, yeah.
00:53:12.525 -->
00:53:14.235And I'm gonna go to a mini rant here.
00:53:14.264 -->
00:53:19.545Um, so, and this is not directed at you, just, uh, directed at the universe at large.
00:53:19.545 -->
00:53:20.179Um, okay.
00:53:20.625 -->
00:53:22.905Um, I'll, I'll say a couple things.
00:53:22.905 -->
00:53:29.954One is, um, and partly a life lesson is, uh, we as humans have a tendency.
00:53:30.375 -->
00:53:37.125To say, um, here are our weaknesses, here are our strengths, and I'm gonna kind of iterate on my weaknesses and make them less weaknesses.
00:53:37.125 -->
00:53:47.324And so I'm gonna kind of broaden, uh, broaden my scope, which is, uh, useful, uh, a much better return on calories, um, spent, is to actually amplify your strengths.
00:53:47.744 -->
00:53:55.423Uh, because if you amplify your strengths that you're already strong at, um, the return on that will far overshadow the weaknesses.
00:53:55.483 -->
00:53:56.594The weaknesses won't matter.
00:53:57.554 -->
00:53:58.259I'll like, I have like.
00:53:58.619 -->
00:54:00.659Uh, lots of quirks of weirdness.
00:54:00.659 -->
00:54:03.298For instance, I don't do phone calls, um, like at all.
00:54:03.298 -->
00:54:06.059I don't, and I have maybe four phone calls a year.
00:54:06.059 -->
00:54:06.239Right?
00:54:06.239 -->
00:54:10.588It's like, it's, um, and it's just always been, it's like, and I, and I, I don't manage anyone.
00:54:10.588 -->
00:54:12.210We have 9,000 people at HubSpot.
00:54:12.239 -->
00:54:19.798I've never managed anyone because I'm very, very bad at it because I've like, okay, there's a very limited number of things that I think I'm actually, um, you know, pretty good at.
00:54:20.338 -->
00:54:22.108Uh, I just wanna get better at those things.
00:54:22.139 -->
00:54:23.699Um, and if I get good enough at those things.
00:54:24.764 -->
00:54:25.963The rest of it won't matter.
00:54:25.963 -->
00:54:33.074Won't matter that I can't manage people won't matter that, you know, that I, I have all these faults and deficiencies, which brings me back to the, the kind of coding thing.
00:54:33.074 -->
00:54:36.943It just so happens that the thing I'm good at is producing software.
00:54:37.423 -->
00:54:38.713That, that's, that's my thing.
00:54:38.744 -->
00:54:43.454I, uh, I'm grateful for having, uh, I learned a little bit later in my life than I would've liked to.
00:54:43.454 -->
00:54:45.644I didn't, you know, start coding, uh, until in my twenties.
00:54:45.644 -->
00:54:50.954But, um, but, but I think this, this general rule actually applies.
00:54:51.014 -->
00:54:51.628Um, and I'll, I'll.
00:54:52.230 -->
00:54:53.608And I'll get to the ranty portion of it.
00:54:53.608 -->
00:55:02.699So the, the universe says, um, that you are good at, uh, you're an author, uh, you are a musician.
00:55:02.849 -->
00:55:03.599Uh, you're an artist.
00:55:03.599 -->
00:55:07.409You're an athlete in most disciplines, most disciplines.
00:55:07.710 -->
00:55:10.500We don't say, Hey, uh, you are a bestselling author.
00:55:10.500 -->
00:55:11.400You're Stephen King.
00:55:11.909 -->
00:55:16.528What we want you to do is we want you to stop writing and we want you to manage this team of writers over here.
00:55:18.224 -->
00:55:19.483Like that would be sort of silly.
00:55:19.514 -->
00:55:21.253Oh, you're a world class musician.
00:55:21.253 -->
00:55:25.125You've trained your entire life to be a world class musician.
00:55:25.543 -->
00:55:31.844We're gonna have you stop actually playing music and we're gonna start having you do like something else.
00:55:31.844 -->
00:55:34.634Because that like, it doesn't make any sense.
00:55:34.634 -->
00:55:42.855And I think part of why we do it in business is because we sort of associate businessy things like coding or whatever.
00:55:42.855 -->
00:55:43.664It's like, oh well.
00:55:44.340 -->
00:55:47.309Because the other question I get is like, well, why do you work so hard?
00:55:47.309 -->
00:55:49.829It's like, okay, well I'm not doing it for the money.
00:55:49.829 -->
00:55:51.059I, you know, I have enough of that.
00:55:51.059 -->
00:55:52.920I don't, uh, live an extravagant lifestyle.
00:55:52.920 -->
00:55:55.079It's, it's because AI love it.
00:55:55.110 -->
00:55:57.929Uh, and I want to get better and better at it.
00:55:57.929 -->
00:55:59.400And it's changing and it's fun.
00:55:59.460 -->
00:56:17.130Um, so that's why, and it's like, and it pays for me, it pays off both emotionally and uh, and financially, which is, I am not the smartest developer in the world, but I have a. I think a mix of skills that I've acquired over the fullness of time, of which that is, is one of the pillars.
00:56:17.159 -->
00:56:17.639Uh, so yeah.
00:56:17.730 -->
00:56:27.239So well, I think that's gonna be great advice for so many people listening and yeah, you've definitely done things uniquely, but you're wildly successful and, and it's brought you to where you are.
00:56:27.239 -->
00:56:28.409So, um, yeah.
00:56:28.409 -->
00:56:31.440That's, that's incredible advice and I think that's all the time we have.
00:56:31.440 -->
00:56:34.769So thanks so much for, for doing this, and I had a blast.
00:56:35.039 -->
00:56:35.639Oh, thanks for having me.
00:56:35.639 -->
00:56:35.909It was fun.
00:56:36.570 -->
00:56:36.929Do this.
