This Week in Startups - Zapier Co-Founder Mike Knoop on category creation, API evolution & AI architecture | E1769

Starting point is 00:00:00 It's really not that many great local Italian places. Oh, there is a good pizza joint called Centro Pizza on Broadway. Okay. So you know Broadway and Burlingame? There's like the Schitt Street and there's the Great Street. Yep. Broadway has a place called Centro Pizza and they make brick oven pizza and it's fucking amazing. It's the best pizza in the peninsula I found.

Starting point is 00:00:23 Centro. Does it look good. Yeah. It's pretty great. Okay. There's your cold open, everybody. This week in startups is broad. brought to you by Mbroker's startup insurance program helps startup secure the most important types

Starting point is 00:00:35 of insurance at a lower cost and with less hassle. Save up to 20% off of traditional insurance today at a broker.com slash twist. While you're there, get an extra 10% off using offer code twist. Lemon.io. Need to speed up your product development without draining your budget? Hire vetted engineers from Europe at lemon.io. Go to lemon.com. Go to lemon.iow. slash twist to get 15% off for the first four weeks. And 8Sleep. Good Sleep is the ultimate game changer. Now you can add the pod cover to any mattress.

Starting point is 00:01:12 Go to 8Sleep.com slash twist to check out the pod cover and get $150 off at checkout. All right, everybody, welcome back to this week and startups. My guest today is Mike Mnup. He's the co-founder president and head of Labs. Had Zapier. If you don't know Zapier, I'm about. to make you happier. Zapier is an amazing pool. I discovered, God, it's close to a decade ago, that helped me do really interesting automations between Google Docs, my email,

Starting point is 00:01:43 you know, basic stuff. If somebody signs up for my newsletter, put them into this Google, this Google sheet. If it's somebody's in this Google sheet, pipe it into my Slack room when somebody signs up for Launch Fund 4 as an LP. And I've been doing these automations over the years, and I train everybody on my team to learn how to use Zapier, Notion, Coda, the Google Doc Suite, and Zapier,

Starting point is 00:02:11 and all these products because you can automate so many tasks. Your partner and your co-founder, Wade's been on the pod, I think twice in the past. But, Mike, is this your first time on the pod? I think so. Yes.

Starting point is 00:02:23 First time. Thanks for having me. Well, I just wanted to say also, congrats. I mean, when people saw Zappier, And I guess your contemporary, if this than that was a, there was like a couple of companies trying to do this. And everybody was like, yeah, that's a niche business. It is not a niche business.

Starting point is 00:02:41 Explain to everybody when you started the company. Then we'll get into all this AI stuff, which is why I wanted to have you on, because AI changes everything with what you're doing. And we're going to do a bunch of interesting demos and talk about how startups and everybody can be using AI and Zapier to plug everything together. but when did the company start and then when did you realize that you were onto something big and then what's the footprint of the company now? Because I care all kinds of numbers. Somebody told me you're making over 100 million in revenue. I don't know if that's true. But where's the company at today and where did you start? Yeah. Well, I think it surprised me as well in terms of like how big the business could get

Starting point is 00:03:19 when we started it. Brian Wade, my two co-founders and we got started back in Columbia, of Missouri, small college town, University of Missouri. We got started a startup weekend. So that was kind of what brought the three of us together. And we were all working with like APIs and our day jobs and side jobs. I was like one of the early moderators and like big users of the Facebook API when it came out in like 2009, 2010. So you're all using these APIs like in contract work.

Starting point is 00:03:46 And we're just doing the same things over and over again with them. And I think Brian was the one who pitched the idea at startup weekend. And the idea was like, hey, there's this huge, you know, wave of API. They're really cool. Like, wouldn't be even cooler, though, if more people could actually use them because, you know, I still have a very technical event

Starting point is 00:04:02 to be able to take advantage of them. And that was kind of the thesis. And as you started looking online, and, you know, if you sort of go online and search around for using the easy APIs or more commonly, what you'd search for, how do I connect these two services together, all you would find on the internet

Starting point is 00:04:17 back in the 2010, 2011, 2011, 2012 era was basically developer documentation. You know, you'd find a stack overflow link of like, oh, yeah, great. Here's a bunch of code you can use to connect, you know, Salesforce with Gmail, for example, or high-rise with, you know, Basecamp or something popular tools back in that. High-rise and BaseCamp, yeah. Really old school there.

Starting point is 00:04:36 You do that same search today in like the sort of landscape of the results. It's totally different. But that was sort of the sort of landscape that it looked like back then. And I think our observation was, you know, okay, well, you know, you see all these forums where folks are almost begging the vendors for integrations. You know, you'd go to, say, the high-rise forums and just, you know, to see these forum threads with like hundreds of their users begging for like, hey, can you add this like random XYZ integration?

Starting point is 00:05:01 And it never really made sense to them to make add more than one, two, three, or four, you know, the top requested ones just because the long tail, it's a bit of an N-squared problem, right? Every new app that gets added, there's an integration that wants to get integrated with it. And we realized, well, okay, we're probably never going to like capture, you know, the direct native integration experience. Like the vendors are going to build those directly themselves. But we could provide this sort of ubiquitous platform, you know, maybe we can

Starting point is 00:05:24 get 5, 10% of all of the integration market out there because we'll be able to service the set of users that just the vendors themselves are never going to be willing to service. That was kind of the vet and that was the original thesis that like, hey, this could be more than just like a small niche, a small niche SaaS company. Over the first few users, we got kind of started building, you know, I think one of the things that really changed like my perspective of what the business looked. I always thought for a long time, I was actually not a personal user of Zapier for the first a couple years, like I was building for our customers. And I always sort of saw it as kind of boring

Starting point is 00:05:58 productivity software. And that was like, that's how I viewed the software. You know, it's cool, great business, like boring VDB software. And what sort of started to change my mind about it was several years in, we started going to a lot of these like conferences with our users and with partners. And we started having a lot of people coming up to us and like sort of like shouting our name, like shouting at giving us huge high fives and just being. so a few, like, thankful to we existed.

Starting point is 00:06:24 There was like passion from the user base. Yeah, it was weird. It didn't match with my mental. Right. Of the business class. When you double clicked on that, because this is really the key.

Starting point is 00:06:33 You had what we call in the industry, market pull. Not only where people were looking for this product, and this is beyond product market fit. You had people searching the internet. How do I integrate these two things? How do I create some glue? How do I solve this problem?

Starting point is 00:06:48 And then they're so delighted. They would scream your name at a conference at you. Yeah. Yeah, they'd see the big orange t-shirt and like that would, they'd run out to us. This is a great feeling. Yeah.

Starting point is 00:06:59 And what it really was was like these, these users were not, and certainly the software can be used this way. They were not using the software for pure like optimization, time optimization use cases. Not to like, hey, save me five minutes a week or save me an hour a week. These users were like doing something that was like transformational for themselves or for their team or for their business.

Starting point is 00:07:19 It was like, it was almost like a skill in line. Like, hey, I thought I couldn't do this. And because Zapier existed, I was able to do it. So, you know, you think of like a solopreneur or like a one or two small person business that thinks like, hey, it's out of reach for me able to build a business. And because I actually have access to these tools, and I can build an inbound lead generation through, you know, like a Google form and a lead scoring mechanism and an outbound email thing with MailChimp.

Starting point is 00:07:45 Like, I can actually do it now. Maybe I was budget constrained to be able to hire a developer. And I didn't have the skills at the time to go learn how to be an engineer to kind of stitch together all these tools myself. So it'll unlock a lot of folks that think to be able to do things with the software that, yeah, just previously felt out of reach. And I think that, like, that feeling was what drove the, drove the passion. And I don't know, it got me way more excited about really trying to grow the business as

Starting point is 00:08:05 as a bunch as we could. Yeah. And the company's now worth $5 billion, yada, yada, you've raised a ton of money. You've got how many customers, how many employees, ballpark? Several hundred thousand paying customers, over 10 million folks have checked out and tried Zappier over the last decade. We've been around quite a while at this point. over 5,000 apps at this point as well.

Starting point is 00:08:24 And you've blown past 100 million in revenue. That rumor's true. The last number we shared was like 150 million. Wow. That's just mind-blowing. It took a decade or just over, I guess, right? You're kind of on your 10-year, past your 10-year anniversary. Yep.

Starting point is 00:08:38 But it really took, if you look at that 10-year-plus journey, at what point did you have that inflection point where, hey, this is really starting to ramp up? Because I think some people get discouraged during those first. couple of years when maybe you have light product market fit. And like you said, you didn't think it was a big deal. If you could pinpoint, you know, that moment when you said, hey, you went to the conference, people started yelling out your name. They see the orange shirt.

Starting point is 00:09:05 What moment in time was that? And then when did the business actually start to crank and make revenue? Yeah. I think the 2014 probably was around the year where we started just to get enough like recognition in the market from like users and customers and partners to get that like passion and hear the excitement. That was also the year that we got profitable. So, you know, one of the other unusual things about our business is we've raised very little, venture capital, only a million dollars back in 2012 once the balance sheet. Since then, we've basically run the business on

Starting point is 00:09:31 cash, uh, from customers. Um, so that any of those fundraising you've done is just secondary or something since then. Yeah. Yeah, we, we've sort of offered. We wanted to, um, offer an equity program for everyone in the organization a couple years ago. So yeah, we went out to the market to get a, we've never raised money because we didn't know the share price. It's actually worth. So we went out and actually, we went out and actually, got us your price and said, okay, now we're going to start. We can build our compensation models around that. And actually offer that to everyone now going forward in the organization.

Starting point is 00:09:56 And I remember Salesforce Ventures was one of the early investors. Obviously, you went to Y Combinator, another great hit by YC. And then you just set up a secondary plan for your employees. How do you, everybody has a lot of questions about that. How do you look at executing it? You know, this employee stock option plan, equitably, fairly, keep people motivated, yada, yada. Did you have a process there?

Starting point is 00:10:25 There was definitely a history, too. It was pretty interesting for us. So when we first started the business, we were three dudes from Missouri. So we really had more of that, I guess, ethos and how we kind of ran the business, which was like, you sell product, you make money, you scale the business basis of money you make. You know, we just didn't have the like Silicon Valley, like, raise $100 million. That wasn't sort of our default operating model coming into the business.

Starting point is 00:10:46 And because we were able to get profitable really early, you know, one of the things we thought to do, we actually did offer equity to early employees. Like we kind of, you know, we went through YC,

Starting point is 00:10:55 so we got the like, kind of traditional startup advice. Like, oh, we set up an option pool and, you know, offer equity.

Starting point is 00:11:00 So we did. And the reality, those, all those early employees because we were hiring out of our networks when we've been remote since 2012 as well,

Starting point is 00:11:07 we were hiring out of the Midwest. We're hiring internationally in Europe. And like, none of those early folks really valued the equity part of the opposition package at all. No, they've never seen anybody make money off equity. In fact, they've seen,

Starting point is 00:11:18 people get lied to with equity in some of those places that would ever be worth something. So they just are like, hey, give me cash. And if you want to give me a little extra cash. We switched. We switched to profit sharing really early on. It was probably around 2014 when we got profitable. I think when we sort of switched over that model and said, you know, this is what our sort of teams are telling us they want. Our employees are telling us they want. So like, let's talk about that instead. And it was way less overhead too for offering it because it's, you know, we had a global sort of employee base, just like the logistics of offering and props sharing work, where we're a lot simpler.

Starting point is 00:11:50 So we actually bring in that model for a really long time. And up until, you know, closer to 2019, going into 2020, where, you know, Zappar wasn't a lottery card anymore like it was in the early days. Like, okay, we built a real business. You know, north of $100 million recurring revenue, this is not something that's, like, going to go away. So we said, all right, we want to start offering equity for everybody and get everyone a chance to sort of, like, participate in the upside of the business at that point.

Starting point is 00:12:14 And that's where we kind of kicked off the logistics to like, okay, let's actually go try to figure out how we're going to create a secondary market. Can we get us to your price for this asset, figure out what it's worth, build that into sort of our compensation models? So now we still do have like a bonus program that looks more traditional. Like we kind of pivoted our profit sharing to more of a bonus program. But we added in the sort of equity company. Melchim famously, you know, did this.

Starting point is 00:12:35 So we call companies like this internally at our firm Alicorns. You know, it's like a unicorn and a Pegasus. And my joke was they fly over traditional funding rounds. com.com we invested in that company when it was like a four and a half million dollar company and nobody would invest in it 40 VC said no we said yes and then Alex and Michael came to me and they're like oh we're raising a little bit of money and doing a little secondary you cool with that I'm like yeah whatever and they're like yeah it's at 250 million and then I think the next round after that was one point X billion and they didn't need the money like you they just did it off money and 37 signals was similar

Starting point is 00:13:09 um what's the uh oh survey monkey was another several so survey monkey and MailChimp both did it this way. It is possible. You raised under two million and you got to over 150 million in revenue. Just let that sink in. That is the definition of market. I'm not like dogmatic about not raising money. I tend to like Zapier's done some weird stuff, right? We got profitable early. We've been a remote company since the very beginning of the business back in 2012. You know, I like to think of Zapier a bit of as an existence proof of like alternative ways to grow and scale companies. Now, I'll also be clear. I think I'll a lot of those things got pulled out of us rather than us, like, expressing them intentionally

Starting point is 00:13:50 or proactively, like, hey, we found a niche in the market that was like underserved and we're able to actually, like, do this thing. You know, one of the things I think that's under realized about Zapier is, I actually think it's one of its fundamental innovations is a bit of a business model innovation more than anything else. Explain the business model. Yeah. Yeah.

Starting point is 00:14:07 Well, so like in the, you know, in the 90s and 2000s, like integration was, is a thing. Like, APIs existed. You just had to, yeah, middleware. Microsoft BizTalk, if you remember. Yeah, that tech, that tech, that product. But it's just like costed millions of dollars to have like huge fleets of, you know, integrators basically custom engineers and IT to come into your organization and like stitch all this software up together.

Starting point is 00:14:30 And our sort of, I think, innovation was, hey, we found a way to actually deliver the software at a, in a usable fashion, and we found a way to reach customers through search, which cost us sort of zero dollars, right? And so every time we're adding integrations to the platform, which, by the way, are also built by a majority of by our partners for free, there's, you know, not a, there's no money that changes hands there. So like partners are building integrations sort of for free to get access and deliberate integration of their customers. That opens up and adds new search landing pages to Zapier, which we get new customers then from Google for $0,000 effectively. So we're able to sort of find this flywheel that meant we were able to acquire customers for very, very low cost, which means we can deliver the software at $10 a month, $15 a month, $20,000. a month starting price.

Starting point is 00:15:13 Yeah. And that let us reach a sort of set of customers and users in the world that otherwise just weren't being served historically. And so that's just how we make money still through today. We have software as a service, you know, we have starting price plans around 20 bucks, 50 bucks, 100 bucks goes up from there. You know, we add on layer on features around teams and companies and organizations and things like that.

Starting point is 00:15:34 We're starting to add more of a traditional sort of sales and go to market plan as well for like market customers and mid-market customers. by and large, we make money directly by selling software on users and charging for the amount of the amount of like tasks and usage they have on Sabir. Listen, I work with super early stage companies at launch, like literally year zero. They haven't even incorporated yet. And then we hit the Series A. People have thousands of dollars in MRR.

Starting point is 00:15:58 And maybe they've only raised a couple of hundred thousand before that Series A. And they don't have their insurance set up. And in fact, we recently had a great startup that didn't have DNO. And we had to really stop everything because they were having board meetings. They were making massive decisions. There were legal issues. And they didn't have the basic D&O insurance that protects directors and officers. So we send them right to Embroker. Embroker is business insurance built specifically for startups. A single application will help your startup get four quotes for four lines of coverage in 15 minutes. Think about that. Four quotes,

Starting point is 00:16:31 four lines, 15 minutes. And they're going to connect you with one of their expert brokers for unmatched service that goes beyond your policy. We use it at launch. It's easy, peasy. It's easy, lemon squeasy. It's easy, breezy. What more do I need to tell you? I use it. I love it. A lot of our startups use it. They love it. Try and broker today with the code twist and you'll get 10% off their startup package in broker.com slash twist. That's eMBR-O-K-E-R.com slash twist. And use the code twist for 10% off. Okay. Let's get back to this amazing episode. Something interesting happened recently. You mentioned that you were playing with the Facebook API, I believe. They quickly deprecated that Jamoff talked about it on All In, I think recently when they realized like, hey, wait a second, this is like the core to the business. We don't want you having access to this. Zuckerberg being savvy. And then famously now Elon and Reddit, I just had Steve Huffman on a couple weeks ago. Or last Friday, I think actually was. And he talked about them. I don't know if you saw the episode where we talked about turning off access to the API or I'm not turning it off charging a reasonable fee for it. Monetizing the API. So I'm wondering just.

Starting point is 00:17:41 With this collection of examples, they're all happened to be social sites, which is interesting. Yeah, that's my observation as well. Okay. LinkedIn is another one from earlier that you didn't mention that Craigslist has no API,

Starting point is 00:17:53 never did, and will in fact sue you if you use any of their data through a scraper. Yeah. So maybe you can talk about how. If you want to pick out a business example, of MailChimp and Shopify also went through sort of an interesting API breakup. I think when they started competing with each other. Just play on that.

Starting point is 00:18:09 You can go look at their like public blog, plus around this. But yeah, effectively, you know, Mailchip, I think, was starting to introduce products in the market in order to compete. And like, some of those were going head-to-head with Shopify. So, you know, both of them sort of mutually said, well, it's not great for my business, just be giving away value to my competitor. So, like, we're going to start, like, disallowing our use case or not, not providing the same native integration that they previously had. And one funny outcome from that was, like, we had customers for both of them basically coming to us and say, hey, my like, vendor of choice is going to stop, like, supporting this native integration.

Starting point is 00:18:40 can I just use Zapier? And that led to both of Shopify which also like basically just sending us a lot of other users who depended on that navigation because we have a lot of a step in as a bit of a neutral middle ground. While you were speaking, I typed in MailChimp, Spotify API and you're the number one result.

Starting point is 00:18:56 I think on Zafir because you're neutral. You're Sweden. But what do you think of this charging for the API because obviously that changes your business? You now have to, I guess, ask people to put in their tokens to do this. And then I guess with AI and Open AI specifically,

Starting point is 00:19:16 you know, they have calls and stuff like that. Does that dramatically change your business or do people just have to fill up their, you know, actually, like, first party vendors, most software providers is actually the preferred way to go. Like Open AI is kind of introducing a bit of a new way to do product monetization where like, hey, you have a direct billing relationship with Open AI. And if you want to use a platform product like Zapier to plug in that intelligence

Starting point is 00:19:40 layer into a sort of a workflow, you bring your own key, right? You bring your API key to Zapier. My sense is this is actually like the smart, savvy way to go about it for a lot of these products. I kind of actually wish that things like, you know, Twitter and all them would actually adopt more of this model where it's like, okay, if I have a, I'm going to establish my direct billing relationship with my sort of, you know, first party vendor and then allow that user to bring their token to other tools. And you just charge for the user to have access to those tools.

Starting point is 00:20:08 You can, you know, you can say, okay, well, allow access to sort of the API for, say, Twitter in this example, a lot of Twitter apps, as long as that customer is, you know, Twitter blue and already paying for it, for example. I think that's sort of like, I don't think anybody wants to get disarminated. It's like, why would you ever let like a third-party company charge on your behalf anyway? I think it's probably not the best place to me.

Starting point is 00:20:28 Yeah, it got kind of weird. I think if you look at the time period where you started your company, it started right at the kind of end, tail end of Web 2.0. And the Web 2.0 movement, really where API started in 2005, 6. seven, eight. People were just looking, people were under-resourced. Twitter was under-resourced as a company.

Starting point is 00:20:47 They couldn't raise a lot of money. They couldn't afford to have iOS developers. And, you know, when apps came out, so they're, hey, you all have at it. Reddit didn't have enough money. It was kind of free outsource development. It was how the community looked at it. And you're like, hey, you make some value for yourself in the world.

Starting point is 00:21:05 Don't do anything stupid. And, you know, have at it. And then the problem, I guess, of course, becomes then when you have to go public like Reddit does, or Twitter has to turn a profit eventually, you need to tighten the screws here. And then you find out, whoa, these people were really abusing the API. They were taking our data or users and selling it to people and get all these kind of gray markets, et cetera. So I think it's actually kind of cool for the idea that you could just fill up your card. And I have a couple of startups who are doing this at OpenAI. They fill up their card.

Starting point is 00:21:36 And yeah, then they run it down. And it's like, okay, I need to get more, what do they call that a card that you get when you are in college and you go to the cafeteria? Whatever that card is called. Meal card. Yeah, your meal card. Yeah, your meal card. Yeah, points on the card. Get points on the card.

Starting point is 00:21:51 So, all right. The reality, I think for a lot of like B2B companies is APIs are actually in their interest, right? Where social companies have a bigger downside that they have to protect against, which is disintermediation. Like, you know, I think in Twitter's case, the stories that I had read on the online was, you know, folks, basically there was like a bunch of, of first-party platform apps that were using the Twitter API that we're starting to get consolidated under one owner. And Twitter sort of got spooked and said, well, shoot, we don't want like our business front end to get disremeded with our users through one owner. So like, we're going to tighten things down. So if you can figure out a clever way to like protect against that outcome

Starting point is 00:22:27 from happening, then I think that's all upside from a sort of monetization standpoint. I think users at this point in time around subscriptions are like used to the idea of like, okay, if you're going to provide an ongoing service, there's, you know, there's an expectation that there can be a cost associated with that. Whereas on the BDB side and the pro-sumer side, integrations are like purely upside. And there's no disintermediation risk. In fact, these integrations are usually really good.

Starting point is 00:22:49 We actually ran a bunch of studies, one with Typeform that showed integrated users churn like 10% less. Yeah, I mean, we use... The fact is, we use Notio, we use Typeform. And Typeform, we love, and SurveyMonkey we love, we love all these products. If they didn't have integrations, yeah, we might actually use...

Starting point is 00:23:08 a product like, you know, Google Sheets allows you to do forms. We might use a product like Google Sheets instead and just be like, ah, it's not as good, but it has integration. So we'll go with that, right? It's almost like you would pick, we would not use certain products,

Starting point is 00:23:25 certain SaaS products if they didn't have. I can have from our users too. They will select. We have a, quite a few users lately. We'll actually say, go to our app directory page on Zapierreter and use that to choose software because they can like trust. So I at least I know what integrates with.

Starting point is 00:23:38 things, like they have a sort of the right mindset on that. Okay, listen, you got an idea for a tech startup. Great. You think you want to change the world. You think you got this. This is the one. Well, you've got that same problem that we all do. You don't have an engineer or you don't have enough engineers to make this happen.

Starting point is 00:23:53 And you need product velocity. You need to go fast. And how are you going to go fast? And how are you going to control your burn rate? If you got no engineers, well, what if you had a partner who could provide you with more than a thousand on-demand developers? And those developers were all vetted experience, result-oriented. and passionate about startups and building great products.

Starting point is 00:24:11 Well, what if they also charge competitive rates? Does this all sound too good to be true? I know it does. Well, then you need to head to Lemon.io because your dreams have come true right now, startups. Choose Lemon. com, because they only offer handpicked developers with three or more years of experience and ones that have strong portfolios.

Starting point is 00:24:27 Only 1% of candidates who apply to work at Lemon.io actually get accepted. And if anything goes wrong, Lemon will get you a replacement ASAP. A couple of great launch founders have worked with Lemon.com. people in our portfolio, and they have had great experiences. So I want you to learn more at lemon.io slash twist. And when you go there, you're going to find your perfect developer or an entire tech team. And you're going to do that in 48 hours or less.

Starting point is 00:24:50 And Twist listeners get 15% off the first for a week. I want you to stop burning money. I want you to hire developers smarter and faster. Visit lemon.com.i.o slash twist and give it a shot. All right. I want to talk to you about, I want to actually talk to because you have some big thoughts on AI regulation. We'll get to that in the tail end of the show.

Starting point is 00:25:07 and the third act. But for the second act here, let's go over some of the cool stuff people are doing on Zapier with integrations because my core, and I'll let you fire up your, and share your screen while we're talking,

Starting point is 00:25:21 my core premise here is there's going to be like a permanent hiring freeze at companies because everybody's getting 30% more efficient a year using these tools, at least, I think, and they will continue. So why add more people if writing a job rec

Starting point is 00:25:35 takes more time than writing a script and automating something. So that's the core tenant I come to with this. You think that resonates with most businesses. When faced with a hiring wreck versus making things more efficient with tools like yours, what should you do? I think it's what a lot of users want. Like, users want automation technology to do work while they sleep. Like, that's, that is what buyers want.

Starting point is 00:26:01 That's the dream. I don't think we're there yet. And all the use cases, even internally, earlier this year in March, we held a company-wide hackathon. We actually told everyone the company, our pencils down. We're going to take an entire week. And the company needs to re-educate themselves around like, what is possible with this technology, what's not possible, and figure out ways to work into your workflows. And we've gotten up 20% of every individual person at Zapier.

Starting point is 00:26:26 20% of all of employees worked in some sort of AI into a Zapper workflow that they use. As far as I actually have not talked with a company that was a higher person. presenters than that yet. So I actually think there's perhaps interesting things to learn there. But it's existential for us. We had to do this. Like, AI and automation are essentially synonymous, I think, going forward.

Starting point is 00:26:46 And, you know, Zapier's business is basically, hey, software that works while you sleep. Right? Yeah. And you're of course referring to auto-GPT or baby GPTs, I guess people call these, where you give a set of instructions to an AI and it performs them

Starting point is 00:27:02 over time, perhaps even getting better at the task with some scripting or instructions. So let's do some examples. Well, it's not too far apart from, you know, even how you think about what Zapier does today. It's just hard to use for most people. Like Zapier, if you set up a Zap, it's a workflow, it is going to do something forever and without you using the keyboard to come interact with it. But it's really limited.

Starting point is 00:27:26 It's constraining, right? It's like rigid. And it's also hard to set out. You mentioned at the top of this podcast. So, like, hey, I have to educate my new employees. How do you use this technology? It's just not easy enough to actually use right out of the box. The penetration rate of this tech isn't very deep yet.

Starting point is 00:27:41 So I think exactly the opportunity. But it's still too hard to use. And I think that's where the language and technology really has a chance to shine. But yeah, the demos that actually have are the first one is actually they, so they're mostly centered around a chat GPU plugin that we've launched back in, we're one of the launch partners with open AI back in March. And I'll show this off in this order. We'll go through maybe a couple examples, and we can end on one of the APIs,

Starting point is 00:28:08 how this actually works on the hood. So this is an example. I think how we've seen most of our users, like, even in some internal employees that's after adopting this stuff is, you know, they'll still, I've heard a lot of anecdotes actually internally where basically folks want two tabs open all day. They'll have a chat chip tip T open at one tab and Zapier open the other tab. And the reason is because the models of how the software works are like completely different.

Starting point is 00:28:30 Right. Chat GpT is a piece of software you have to interact with in order to get value with it. So, you know, the next example is one I've used myself is, you know, grab an email from my inbox that matches a certain format and, you know, draft a automatically draft a sort of reply to that. Okay. So you're in chat chip ET4. You're using the plugins. You pick Zapier and you say, hey, I want to check for an email in my Gmail account. And you've already authorized it to go to Gmail. And now it finds the latest email and does a reply for you.

Starting point is 00:29:00 Yeah, this one summarizes the reply. And then I think if I kind of just zoom forward here, one of the actual downsides with sort of the plugin architecture on how chatchptu works right now is everything sort of has to go through these like confirm flows, which is I understand why they do it. You know, sort of the safety argument around it. However, I do think that there's probably some edge cases where we actually take a stronger safe stance than than their platform does. And like it kind of creates a weird system where like two safety systems are trying to be in the middle and it creates a really awkward user experience. So I think there is sort of more stuff we can do there. But here's an example where the plugin's going to come back and grab the email response and summarize it back in.

Starting point is 00:29:39 Yeah, so this is an example where we actually opened up a tab on Zapier to give a preview of what the plugin action is about to do. I think this is an example where we've actually inserted on safety things to let the user explicitly know what actions are going to do on their behalf instead of just letting them roam free on Zapier on your account. And now we're back inside chatchpti after the user's confirmed and it's pulled in the email and summarizing the email. And I think then in this demo, we actually even follow up and ask chat chpd, hey, can you sign that with my name and rewrite it my tone and then send it through Gmail as well? And you can actually fire off now from the chat chpt interface using the Zapier plugin.

Starting point is 00:30:19 If you've authorized it, it makes you go to that step. It will actually do the send from the chat chpity interface. It will. In this case, we're creating drafts. This is kind of another one of those, like, probably tips I would have for most folks that are adopting this tech is like, you know, the technology is really, really good at drafting things. So you almost want to lean into use cases where you get to get a preview of it and you can add it and mark it up and have sort of control over it before you send button yourself. You absolutely can hook this directly up to like sending an email directly. But the one we found most of folks inside Zap here adopting is, you know, these flows where it goes through creating a draft free, being able to review it and approve it before hitting send.

Starting point is 00:30:54 And basically the vision for what we want to try and get this to is it feels like an off-flow. You know, wherever, whatever product you're in, if you're in chat, GPT or you're in any other sort of ad product, you need to plug in sort of an action library into it. You know, you click a Zapier button. You say that, that vendor says, hey, I'd like to get access to, you know, your Gmail account, your Salesforce account and, you know, your platform account. And the user says, you know, looks like an awful, a little pop-up.

Starting point is 00:31:17 The user says, yep, that sounds good. And now you're back inside the sort of first-party product, and you can go from there. In the initial version that we released, it's like one, there's one extra step, which is in addition to having to sort of approve it and allow you also explicitly today have to choose which actions from those apps you want to allow the sort of chat Chappet have access to. So today, for example, with the Gmail when you saw, when we set that up, you had to like say, yeah, okay, I want to allow chat Chapti access to Gmail, but I also want to have it allowed to send a draft email. So there's one extra sort of step that we have to choose those. And that's somewhat of a limitation of sort of the language model technology and somewhat of a limitation of the API experience overall right now. Yeah, it's, it's, there is a, it's a little cloogee. You have to log in, I remember doing this.

Starting point is 00:32:05 You have to log into Zapier. And then there are some links that you have to go to to make sure you can search your Gmail, make sure you can send from Gmail. And that I'm sure will be abstracted in the coming weeks and months, yeah? Yeah, I don't disagree with you at all, by the way. And I think if you, and I have, I've looked at it and sort of obsessed over some of the usage and some of the numbers from this stuff. You know, I do think that a lot of the plugins and folks I've talked to, retention is a problem right now with them. You know, I think if you kind of look at like, it's almost like a bit of a numbers game. You know, if you're sort of able to spread your user base over enough users, you can sort of find a percentage of them.

Starting point is 00:32:43 They're going to find you sticky use cases to build this like chat and plug-in thing into their workflow. the reality is most people in the world haven't even worked Chach Chibouti into their workflows yet. So, like, asking them to then add on a plugin that is also like, you know, an active development. Like, it's, I think we're still quite a point of ways away before you're going to see like kind of the refinement you needed from a lot of these, this like plugin ecosystem and how even just getting the penetration of Chachshapit into like legitimate like sticky use cases. I think it's still still a search for most most companies. Well, you know, it's, it's going to be. a slow process and then it's going to be a really fast one because once you know the first 10 people in an organization figure this out and they become

Starting point is 00:33:28 bionic and they're able to do really interesting things with their Gmail box like say who are the people that I was you know emailing with back in 2012 to 2020 that I'm no longer emailing with summarize my conversations with and then suggest which one you know some emails to catch up with them and like, whoa, that's going to be super powerful or like, you know, you're a venture capitalist. Hey, what founders was I talking to 10 years ago? What are they doing now?

Starting point is 00:33:58 And it's like, whoop. Like, this is going to lead to being able to do things that would take so much time. Nobody would ever even consider doing them. Like, you would have to hire a full-time person to, hey, go through my emails from, you know, the 2010 to 2020 period. find every founder, put them into a Google sheet, and then look up on their LinkedIn and see which ones are still at the same company. Boom. It's like, whoa.

Starting point is 00:34:26 And it works. And it works. So, it's super powerful. We sort of see that exact same thing. That's the message we delivered internally, which is how we got so many folks that have started it out in real use cases is, you know, hey, this is like a chance to go learn the technology.

Starting point is 00:34:41 The future is not going to be like, and I can't, I really start able to think of any sort of technological disruption that like, you know, displace jobs. jobs in a quarter. But over the course of two, three years, five years, like, you're going to see folks that accelerate ahead because they understand and use the technology. You're going to have, like, entrances in the job market who natively know this stuff, especially if you look at the adoption rates of, like, chat chip to be in education coming. Like, all those folks graduating

Starting point is 00:35:05 through college and entering the job market, like, they're going to have a skill set that I think a lot of other folks probably won't have. I think that's going to be a leg out for a lot of them. Any other language models now built into Zapier that have integrations, Google Bard? Yeah, we have, well, none like Chat Chbett where we have a plug-in launched yet. There's nothing to announce at this point. Yeah. We do have all of, we have way more language models actually built into Zapier sort of first party, though. So this is kind of the brain burner about it, right?

Starting point is 00:35:35 When we actually went to go build the Chat Chatsby-T plugin, one of the reasons we built that was, you know, we saw this a huge influence. of AI apps launching on Zapier. We had like, Hugging Face and human looper all sort of getting built on Zapier. And we sort of realized like, oh, wow, there's this huge explosion of AI products that's happening in the market that are not going to get on Zapier. And we wanted to offer an API to them to be able to bring Zapier's integration platform into their products. Just felt like, first thing we've ever actually launched a public API, it's a bit of an

Starting point is 00:36:06 almost embarrassing point that we're 10 years in. It's like we're finally now just launching a public API that other sort of vendors can pull in. But we felt like it was sort of what needed to happen at this point, given the pace of products that were getting released. But in the more traditional way where you go to Zapry.com and you're building workflow, we have Anthropic now with Claude is on Zapier. We have the Google, the Bard version. We've got opening X integration as well. So you can build those into more traditional workflows.

Starting point is 00:36:35 But I do think some of the more exciting, interesting ones are like the paradigm shifts, where you have like a completely different, you know, front end interface for how you build and use this stuff. If you want to get ahead in your career, you need to sleep well. And if you want to be a stallion, if you want to be a workhorse, you need to get a great night's sleep. It's that easy. So do what I do. And that's getting an eight sleep, right? Especially over the summer, right? Temperatures are going to rise. You want to stay cool. And eight sleep mattresses allow you to set the temperature. And they're selling more than just mattresses right now. They have something called the pod cover by eight sleep. And this fits on any bed. It's just like a fitted sheet. So if you don't want to get a new mattress,

Starting point is 00:37:12 but you do want eight sleep, you just get the pod cover. It's going to keep you cool all night, all the way down to 55 degrees. And it's going to improve your sleep by automatically adjusting to the temperature on your side of the bed. And, you know, your partner or spouse, they get to set their own temperature. So no more battle with your partner or your spouse over the thermostat. Nope, it's going to adjust based on you and your partner's preferences. And you get personalized sleep reports. Eight sleep, I love it.

Starting point is 00:37:37 In the winter, oh, I keep it nice and toasty. But in the summer, I like those cool. crisp sheets. And I don't want to have to blast the AC in my house and burn all that extra fuel. No, I just use my 8 sleep. I dial it to exactly where I like. And, you know, I get a better night's sleep. Temperature really does correlate with a good night sleep. So go to 8Sleep.com slash twist to get $150 bucks off the pod cover. Stay cool this summer with 8Sleep. Please. And they're now shipping, not only in the U.S., but Canada and the UK and some other countries in the EU and Australia. So 8Sleep.com slash twist for $150 off the pod cover.

Starting point is 00:38:13 Can't go wrong. Talk to me about LinkedIn. They do not allow people to use their API. Is LinkedIn very protective of it? Because it seems to me like nine out of ten times somebody gives me examples of where this is going. LinkedIn comes into play. How do you think about LinkedIn and how do they operate with Zapier at this time?

Starting point is 00:38:34 Yeah, I'm trying to remember. It was five plus years ago. I think when they went through their API sort of tightening phase, they had a pretty opening API at the time, and then they sort of tightened it down and got rid of a lot of their like, like automatic message sending stuff, the like contact scraping stuff. They sort of segmented out into their,

Starting point is 00:38:49 I think it was around the time where they really wanted to go after recruiters as their, like, kind of key customer and buyer. And so they kind of shaped all of their API usage around that persona and kind of just said, you know, all other uses were just going to delete and get rid of if we don't care about them. Yeah. The one the most of our customers and users wanted was the things like lead hydration, where I could take an email address and go, like, get information about that lead,

Starting point is 00:39:15 particularly for like marketing flows. You know, you talk about, hey, you've got to, you know, say a founder of contact form that you have on that form and you get an email address. And you want to like automatically pull in a bunch of information. Say, they don't have to ask the founder to like type in the resume or whatever. So folks, we're using things like that for those use cases. And the reality is the market now is kind of like address the gap. You know, there's a billion different sort of lead hydration. companies that all sorts of scrape public information and, you know, wouldn't be

Starting point is 00:39:38 surprised with some of that originally did come from LinkedIn, but, you know, they had that big public scraping case. They had a big public scraping case with, I think a company based in Israel that they lost. It was like back and forth. They won the first one. They lost an appeal. I'm not actually sure what the current like status at all is. Well, I mean, the great irony of this is they tightened their grip and said, hey, you can't do certain things. And so what that does is since there's a need, like we saw with Napster back in the day, if people want something,

Starting point is 00:40:07 you know, whether it's a TV show, music, or to enrich a lead, enrich an email, and hydration, I've never heard that. It's a great term.

Starting point is 00:40:17 You know, get an email and then get the person's title and where they worked previously. It's really smart. Then some gray market company, uh, doing gray hat stuff is going to do it. And they're going to scrape all of LinkedIn.

Starting point is 00:40:31 and then they're going to back into it. And I know this because I've had so many companies do this with Facebook data, LinkedIn data, and then legal letters get sent. And the only people who get really impacted are the good actors who want to play by the rules. And then the people who get who benefit, they get punished. And the people who get rewarded are the gray hats and the black hats

Starting point is 00:40:57 who are going to just scrape the information offshore. and they don't answer to anybody. So, yeah. Yeah. Yeah. I look at sort of the, you know, the mass core of our users and customers, too. I would,

Starting point is 00:41:08 almost every time there's an API defecation or terms of the service change or something big, like those are the folks getting affected. And by large amounts, Zapier's customers are very small businesses. Thinking like one to four sized teams and companies. Talk to me about like the large enterprises. Is there anybody who's taken Zapier to like a large organization and coordinated it? And do you have that feature?

Starting point is 00:41:29 because right now, I have my team using it, but I don't know if we have a central relationship between. We do. We're like building some offerings here. We're trying to figure how to do this basically. So that's sort of the TLDR. We do have some like examples. Netflix is probably one of the larger ones where,

Starting point is 00:41:45 you know, they'd have a very IT forward perspective where, you know, their IT organization basically has an identity of saying, our job is to make you are more efficient, more productive. Like we get, that's why we get paid.

Starting point is 00:41:54 That's why we're sort of here. Versus, you know, a lot of more traditional old, school mid-market plus companies where, you know, hey, they're like technology side of the house might be looking at the rest of the business. Hey, I'm a cross-center. I'm about protection. I'm about control.

Starting point is 00:42:07 I'm a risk management. And those are just very fundamentally different perspectives. I think when you start talking about like introducing new technology, like language models and like AI out of mission to business process. So we are seeing more and more of like IT workers thinking this way. And we're starting to put together some like packages of software and services around actually basically doing what we did internally, which is. Like, I actually think we have some expertise now at figuring out,

Starting point is 00:42:32 how do we actually adopt AI use cases into real workflows that do allow folks to get legitimate time back and allow them to move on higher value activities and use cases and workflows and jobs and actually make, like, deploy that into the organization. And that's like the common thing that I get asked whenever folks come up to me and talk to me these days around the whole to say I stuff is like, how are you guys actually using it? What are you've seen your users actually use it for? Because I think a lot of folks are like, it's still the time, there's broad awareness. for what language models and AI can do at this point,

Starting point is 00:43:01 I don't think there's broad penetration for it, actually into real use cases yet. Yeah, so talk to me about how you look at, again, I guess everybody's got an opinion on this, but AI regulation. I saw some of your tweets. Do you think that this is moving so fast that there's going to be significant negative consequences for humanity?

Starting point is 00:43:25 Do you think we need to slow it down, or do you think we need some thoughtful regulation? How do you look at this? Because Open AI, you know. Here's my, I'll speak like most to what I know. Certainly, like, I think there's some interesting philosophical, like, things we could jam on. I don't feel equipped to, like, have that argument or do you even debate at this point?

Starting point is 00:43:45 But I can speak to what I know, which is Zapier, we have like millions of legitimate, useful workflows and automation that we've, users have set up over the last decade. I know that. I also know that it is way too hard for most users to use Zapier, even today. I know you gave us lots of, a few praise in the top period and said, hey, I love that it's easy to use. The reality, though, for most users, it's not.

Starting point is 00:44:10 We fought for a decade on trying to make Zapier, using enough to use for the sort of traditional, you know, professional. It does not know how to code. Does not know. It's not technical. However, we still have a long way to go. And we've fought on that problem for so long that I think we are reaching some limits of the paradigm of traditional software

Starting point is 00:44:28 to actually put like workflow automation into the hands of like end users. And I think this AI language technology is the first thing I've really seen that I think offers a step function, not just like an acceleration of a smooth curve, but actually like a step function and an adoption rate of how many business users and users can actually set up

Starting point is 00:44:48 and use more technical concepts, things like automation. Most of the people use app here, even though they might not call themselves technical, well, they're still builders at heart, right? They have that sort of like identity or like, oh, I'm going to go create something. And I think this is where this like language and all technology that helps, it drives down the barrier to creation by just a ton. So, you know, I get excited first and foremost.

Starting point is 00:45:12 I get really excited about the idea of like, oh, wow, well, we have like 10 million people who've like tried Zapier. Maybe this could get us to 100 million folks who've like tried and used automation successfully. And I think that's like a really positive thing. And especially if we model all the use cases on like, what people are using is happier for. That's all great. You know, I just want to make more of those people. And, you know, I think there's a chance to at this point.

Starting point is 00:45:35 So first and foremost, that's kind of where my head goes first is like, I think the technology of transformation, particularly in sort of the prosumer B2B business workflow automation space. And it's not just time saving. You know, these are legitimate. Like, it is like an unlocking technology for a lot of users around what they can actually use language models and automation to do. It's not things that they weren't. otherwise doing. On the open source side, I do think, like, so we haven't talked about this much.

Starting point is 00:46:00 Basically, I actually, I know you introduced me as sort of the president of the company. I gave up my exact title last year, last summer in July. I quit the exec team. I went to Brian away, my co-founders, and I said, I think we have, I got to go on in this language model, MLAI stuff. We've got to learn what this is going to do. So Brian said he was going going to do the same thing around the same time. So both of him and I basically said, we're going to get rid of our exact team roles. And we're just going to go. full-time and focus on research and engineering for AI, particularly in the context of SORDAPier. So back to the laboratory.

Starting point is 00:46:33 Yeah, basically. Out of the exec suite, no more fancy bathrooms, back down to the garage, right? Back to the garage, quite literally. So like, and I do think like you really do have to go hands on to learn what's like possible with this tech. If I look around, I think one of the coolest things is when I like look around at like, you know, my peer group of founders, you know, folks have been around for 10 plus years, all of them are doing something similar, which is like going back and actually.

Starting point is 00:46:55 hands on the technology. And I think you have to to learn what's possible. Yeah. And I actually think Zapier plays a role here too. Because I think this point about you have to go hands-on to learn what it can do and what it can't do is well beyond like, hey, I'm going to open up a GitHub repo and download code locally and run it. Things like Zapier can be at gateways for a lot of users to discover what is possible and what's not. Same way that the chat GPS is offering a view of like what's possible, what's not. You know, we have users that are like basically experimenting, a trial through trial and error with workflows and zaps and plugging,

Starting point is 00:47:24 a reasoning engine, you know, a language model into the middle of a workflow to say lead score or, you know, draft or reply to, you know, a message that I received or draft a pull request that I received or, you know, score the GER tickets or summarize customer feedback and, like, dump it out into a Slack channel. But there's a lot of trial and experimentation regarding it. And I think those users are figuring out what it can't do at the same time, right? They're figuring out, oh, I shouldn't just automatically send an email. No.

Starting point is 00:47:51 Not ready for that. Yeah, be careful. I shouldn't insert this into an HR hiring decision where I. I'm not going to review the decision. Definitely not. So, like, I have a lot of trust when I look at our users of how they do their own experimentations to find what it's good and bad for. Like, I think we got to put that experimentation mindset in more hands.

Starting point is 00:48:06 So that's why I get really excited about the open source thing about, hey, the more we sort can get this technology diffuse in a more individual hands at the end of the day. I think it's going to allow more people in the world to understand what it's good at, what it's bad at and like calibrate. And I don't know, I have a large degree of trust, I think, and sort of folks ability to figure out that and navigate that chart, deal with the antidotes of the technology, as long as you give them enough time to.

Starting point is 00:48:31 That's where my, maybe the philosophical act comes on. It's like, okay, if you really can like sort of drop a, like, I don't know, some sort of step function technology change in a month.

Starting point is 00:48:39 And like, okay, maybe there's like a moment where there's like enough disruption there. It's worth asking the question around. But like, based on everything I've seen of the last 12 months of language models, that is not what we're dealing with.

Starting point is 00:48:50 What we're dealing with is more, hey, there's something that would, take you 100 hours might take you one hour or something that took you 10 hours might take you one hour or you're still going to learn how to do it which is really important I think but you do agree that this is going to make companies massively more efficient and you're going to need much much fewer people to do much more I mean it's hard for me not to agree with that statement in the limited condition based on where folks want this technology to go like as soon as it exists yeah

Starting point is 00:49:22 there's a huge demand for it. Yeah. And so then the question becomes, you know, we, as technologists, looking at society, are left to wonder, are there still problems to solve? Because if a 10-person team can do the work of a 20-person team, they can solve twice as many problems. It's not like there are not a long list of problems to still be solved in humanity. And so that's where I look at and go. Or the new ones.

Starting point is 00:49:47 I mean, the sort of classic use case, probably even Wade might have showed this when it was on the show is like the typist example, right? even literally the word computer. It used to be a profession. It had itself back in the 50s and 60s. Typhus was a profession. Like, it's silly. I got trained in middle school how to type on a keyboard, right? I was just having this.

Starting point is 00:50:02 I wrote in my, I started doing some email newsletters again. And I was like, you know, there used to be, I started my career as a PC support specialist. What a PC support specialist did was they set up your computer. They upgraded the memory. They put in a larger hard drive. They set up your Ethernet card.

Starting point is 00:50:19 And then they sat there for two days, installing software on your, computer using CD-ROMs. All those Apple engineers took your job, Jason, where they made the beautiful iOS onboarding flow when you get a new phone. Exactly. And now it's like, yeah, you don't need a PC support specialist to come and set up your computer and your Microsoft office for you.

Starting point is 00:50:36 You can simply, oh, here's my substack. Thank you. I guess there's a section. And then when I saw startups and when you were setting up your startup, you were right at the point of cloud computing. So did you rack your own service for Zapier? and have a COLO? No, we did not.

Starting point is 00:50:54 Linode, if you remember that name. Linode, yeah, great. I still have a box on there somewhere, but yeah, that was the very first cloud service we used. Yeah,

Starting point is 00:51:03 they were the pioneers, right? And so you, you were the first generation of startup founders to not have to go order PCs and build a rack and find a co-location facility, rent space,

Starting point is 00:51:15 go down to the COLO facility, have a SIS admin, and you might not have even had a SIS admin or somebody at your co-location facility. The generation right before you, if you were working at Flickr or Facebook, you were racking servers, and you had two or three people on your team

Starting point is 00:51:31 who were managing that for you. Yeah, Linode.com slash twist and get a $500 credit. I forgot. I was a sponsor. There's a plug. Thanks for the plug. If I can add one other point of the regulations side,

Starting point is 00:51:44 I do think one side thing that I think is, I'm fearful of this is how it's going to play out. I'm not sure. we'll have to check back in in six months, 12 months, and see if this is true. But I am a little disappointed that I think the way that most of the research around AI and LMS's heading is towards more closed companies, like they're not being as forthright and like forthright and sharing of basically the technology, the progress, the architecture, like they're, we're kind of getting into the space where people are realizing how much value is in. I think the research and they're just going a lot more closed.

Starting point is 00:52:21 Pulling up the ladder behind them. I do think is going to, that that's my like, that's almost my like sort of anti-acceleration viewpoint. Like, I think there is a, there is a path where actually progress slows down for a little bit of time right now because we're either sort of, one, approaching some of the assentote limits of what we're going to exploit out of transformers and language, large language model, the kind of current architectures we have. And to, like, more research is getting sort of closed up. So there's not as much open sharing, not as much progress. And as you know, like, the reason open access basically is because of,

Starting point is 00:52:49 of some of the progress that came out of sort of public sharing from another competitive company with Google. So I do think that's a bummer. That is such a weird move that they went from Open AI to Closed AI. They literally took the reason they existed and reversed it. They're like, this technology is too powerful for everybody to not have a say in it and for it not to be transparent. Then they got a couple years in.

Starting point is 00:53:13 They're like, you know what? This technology is so powerful. It's too powerful for everybody to know how it works. And I am 100% agree with you. Do you think that leads to the developer community? I mean, like, fundamental level to, I completely trust like Sam and Greg. Like, sure. As stewards of technology, I can't think of two of the better people that I would like try

Starting point is 00:53:33 to put in charge of that problem. But like if I look at the second order effects of what that then leads to, like, it does, I am a little worried that it does lead to like more closed up nature. We're going to see less progress. We're going to see less sharing. We're going to see less like, you know, technology getting pushed to the edges. The good news and all of that is it seems like the open source community and the open source

Starting point is 00:53:54 models are advancing much faster. I don't know if you saw that Google memo, but there was a Google engineer who's like, listen, at the pace that open source, the open source community is rocking on this. They're going to just beat us and we don't have a moat and either is open AI. So it's almost like they're squeezing too hard and that's leading to people saying, want, I don't want to build on a closed system, which, you know, you may call it open AI, but I don't want to have the risk factor of working with open AI. So I'll look at some alternatives. And the reality is open AI more than anyone has pushed for the technology to be

Starting point is 00:54:34 developed in public, though. So I will make the argument in their favor. I know that I'm sort of, you know, poking at a few things and decisions I've made. But like, I also think we wouldn't be sitting around this conversation. Zaffrey would not have shifted its viewpoint. No, but they've had I mean, they've sent mixed messages to the market. Yeah, I think that's fair. Yeah. So it's like the largest message is we have a product that you can use for free. And I do think that that has changed a lot of folks' opinions around what's possible.

Starting point is 00:54:59 And catalyves a lot of energy that wouldn't be. Every six months, they seem to just drop the API price 90%. Right? They've done that twice, I think. I think you can bet on like cost going down. I think you can bet on context windows going up. I don't think you can. there's not like a smooth ramp on architecture improvements, though.

Starting point is 00:55:17 I do think that that is like one thing. I've been personally spending more and more time on lately is... Explain that to a layperson, yeah. So we understand what your point is there. I don't proclaim to be an expert in this either. Sure. But I'll try it from my best interpretation of what I understand about transformers at a fundamental level is this is basically an architecture

Starting point is 00:55:34 that was not necessarily, it was published from a paper at Google back in 2017. That paper was the continuation of actually quite a long journey of research as well around what I think originally started around translation, like literally translating a sentence from one language, say English to French, right? And the first sort of deep neural networks that did this were sort of constrained, they kind of fixed the amount of characters that you could translate from X to Y. And that led to the invention of this technique called attention, which allowed you to have variable length inputs and outputs. So you could, you know, the word in English is a different length than the word in French. So you could actually kind of deal with that problem.

Starting point is 00:56:10 And this led that attention mechanism was like its own neural net at one point. And the sort of infamous vapor attention is all you need was the dropping of one of these like ancillary recurrent neural networks because it wasn't like an important part of building a transformer. It really simplified the architecture stack. And that architecture stack also happened to be one that really ran in parallel, which fits sort of the GPU scaling curves that we've sort of seen over the last decade. So those kind of two things kind of in parallel, allowed sort of to invent, you know, progress around GPT. you know, one, two, three, and now four, and sort of so on from that. So, but we have, like, there has not been at least a, like, well-established alternative to architecture. All of the, like, AI products, progress, research papers you're seeing.

Starting point is 00:56:55 Like, a lot of the momentum and sort of attention is really shifted into what can you build on top of language models, right? Like, what can you do with a language model that has this, like, it's seemingly capability of reasoning, right? this capability of tool use, this generality around being able to generate content. Like, what can you do with that? You know, I would have actually, up until last summer, I would have said there's like, I'm like 99.9% sure large language models are not on the critical path to something like AGI that reaches human level generality of intelligence. I've decayed that prediction to call it 80, 85%.

Starting point is 00:57:32 And the reason for that is there are some things that you can do with like GPD4 right now. that no one's productizing. And because it's too slow for the language model to generate tokens, you have to let these language models think out loud and they improve their performance. The classic example is the thing that actually inspired me to go all in on AI last summer, which was the let's think step-by-step paper that came out last January. And this was a technique that some researchers found where if you put, let's think a step-by-step at the top of your prompt and then ask the exact same question again,

Starting point is 00:58:02 the language model actually boosts its performance because it gives it time to generate tokens almost like an internal monologue where you're thinking about, you're letting the model think out loud for what it should do, and then letting it reflect over those tokens that sped out to generate its actual next action. There are some performance evals that went from like 30, 35% of like 80%, it's crazy step function increases. It's really weird.

Starting point is 00:58:24 If you literally say to the chat GPT, let's try that again. And can you try to find me three more? And you just keep doing that. You get to like the six or seven back and forth, then it's like, yeah, I got you your answer. And I did it right. And it's like, another thing too.

Starting point is 00:58:40 Yeah, some partial right correctness is like another big challenge. You've even thought that internally. Anyway, so there's these use cases that like can demonstrate greatness in certain scenarios, oftentimes they're unen reliable. Like the example, you just mentioned where you had to ask it six times and it got one out of six right. Okay, the questions on how do we like figure out which one is right more consistently? Yeah.

Starting point is 00:59:01 Or the second one is these really deep reasoning chains where you can actually let the models continue. It's like thought process. This is the like auto-GPT style stuff where you can let the model reason through like a tree-based search of reasoning for almost 30 minutes, 45 minutes, an hour in cases where we had demos run in last fall. And it can get it to the right answer. But it's so slow and so expensive. Like literally it costs probably $1,000 stress to run that like reasoning search. You're only going to do it for use cases where the product experience is like okay if it's offline and completely asynchronous and like there's huge ROI attached because like the models are expensive. So. Well, I mean, find me a stock to short and explain your thesis. There are examples. Yeah. Okay. Now. An offline ETL job would be another one.

Starting point is 00:59:42 Like, hey, I need to process, you know, million records of data. And I'm okay if it takes three days, that's fine. Just like that was another good use case. So there are examples of like cases where the model just can do better than how they're getting practiced because generally with like consumer facing, prosumer facing products, latency matters a lot. Reliability matters a lot. So like all these product builders are self included or chopping off.

Starting point is 01:00:04 use cases that are slower expensive. And as the sort of cost curve comes down, as the context window goes up, there could be some interesting techniques around reasoning that are just out of reach from a sort of capability, from a, not a capability, from a cost and like performance

Starting point is 01:00:20 standpoint that might be coming to reach and maybe there's a way to build an system on top of this. Having these auto GPs talk to multiple language models and having dueling language models where they analyze each other's data and they start talking to each other, that kind of feels, it's a different type of singularity, but it certainly feels promising if you were to say, hey, I'm looking for

Starting point is 01:00:43 stocks to short. Please go to five language models and ask them about, you know, the stocks that are most shorted right now and then put together a thesis based on their five. And you start having the check all different GPT models around the world working on the same problem. And then some of them asking reinforcement questions to each other. I mean, this is, do you know what the rumors of like the GPT4 architecture? Have you read about those? No, tell me.

Starting point is 01:01:10 What you're describing is effectively the, the rumor unconfirmed as far as I know, but around how GPT4's architecture works, which is essentially they have eight different, um, you know, attention heads or, sort of model heads that are all trained on different subsets of their e-val.

Starting point is 01:01:23 And, uh, they're all 200 billion parameter individual models. And they essentially like do a mixing mechanism where they like, run the input through each one and they mix the output together from the log probes and use that to generate the final tokens. So not too far away from what you're describing where you have like, it's like a mixture of experts, I think is how they would describe it.

Starting point is 01:01:42 Each head is an expert of a different type of reasoning or a different type of input problem and you try to figure out expert. Right. Yeah. And then I don't even know if those are verticalized experts. I mean, who knows how they chop those up is like one expert because of Wikipedia. Oh, man, it sucks that this is closed, right? That seems so interesting.

Starting point is 01:01:59 I would love to read about that. I think that could probably accelerate. progress in some way. The fact that there's only a couple hundred people in the world, they probably know the real answer is. Probably because they also have some exposure based on who trained that data. So let's say one of them is like this is Wikipedia, Reddit, Quora, and Twitter data. And it's, you know, consumer, it's a crowdsourced information.

Starting point is 01:02:19 This is the SEC, academia, you know, the New York Times, Wall Street Journal and like a professional, you know, quote unquote professional. This is a journalist answering the question from the journalist framework. based on the Wall Street Journal, Washington Post, personas and data sets. So, or informed by those. And if they were to actually say that, then you would be able to make the case of,

Starting point is 01:02:41 well, hey, you're literally picking your expertise based on datasets. And that's probably why they circled. That would be an interesting reason to circle the wagons and not share. Pretty secretive about what data they use. I think they'll consider that pretty proprietary. I actually think this is a problem that goes away

Starting point is 01:02:58 over the long run, not because of like some cultural acceptance of this fact, but more about I think the amount of data did you actually need to do to train these systems just gets dramatically lower through model architecture innovation. Ah, fascinating. I mean, there's some existence

Starting point is 01:03:12 proofs here. I actually think there's a lot of really good reason to go all the way back down to like fundamental art and do like an architecture search, essentially. Why as you can? We now have two existence proofs in the world of emergent reasoning intelligence behavior. One is humans, right?

Starting point is 01:03:28 Discovered through sort of genetic evolution. over billions of years and large language models, which were invented over our own cord, running on our own silicon, on algorithms that sort of invented, the fact that n equals two there suggests that, like, oh wow, there's so many more. If you run the probability, like,

Starting point is 01:03:44 you would be shocked with like, it stopped at n equals two around architectures that led in, like the transfer architecture is nowhere near what you see anywhere modeled in sort of the human brain, completely wildly different in how they work. So like, it suggests that there's probably more. And the thing that's interesting by humans is how, comparatively little

Starting point is 01:04:02 examples they need and training data they need in order to be sort of generally intelligent. We seem, humans seem to be born with some of neat amount of like, capabilities or abilities. Very strange. Like, past following, pattern recognition. There's some, like,

Starting point is 01:04:18 things that show really, really in toddlerhood that like, they never get trained on. Fear of, like, fear of predators. Like, we actually understand predators in some way natively. Like, you were grown, if you were, and I think they've done this with studies, like, you, you don't need to have seen a shark coming at you to know you're about to fucking die.

Starting point is 01:04:39 I don't pretend to be a scientist, but like, there's just like some pretty compelling, like I said, existence proof examples where, oh, okay, yeah, they're likely more architectures out here that we should go search for. And like, I think a really good constraint function to go do an architecture search would be to say, let's pay in the amount of sample data that we put into this architecture search so that we can find like architectures that are just way more cost efficient and performant. I could talk to you for hours and we've talked for an hour. Mike, you got to promise me you'll come back. Maybe like six months from now. I think this is moving so fast. I'm going to, I'm going to make an executive decision here. And two things, since you and I are, you know, in close proximity to each other.

Starting point is 01:05:22 Number one, we got to get some, got to get some ramen and then number, or a lobster sandwich. And then number two, you got to come back on the show in six months. Yeah. Thanks so much for- Your office recordings. We can do one in person, too. I know. I'm literally looking for, I'm selling our office in the city and I'm setting up our

Starting point is 01:05:39 incubator and accelerator somewhere in like San Mateo area. And when I get that, we're going to have an in-person studio again. And we'll do a live version of this where we get like audience questions and stuff like that. So I'm trying to find like a theater. Like, I want to get like a theater or like a warehouse space where I could have like 50 people come and just kind of customize yourself. I like raw.

Starting point is 01:05:57 Yeah. I hate these fancy space. A lot of people have been emailing, oh, I got a fancy space for you over here. I got a fancy space over here and El Camino come to this office park. You want to find a restaurant that just shut down basically as I'm hearing. That's what I'm looking for is I'm looking for like a shutdown restaurant, like an old Mexican joint with a parking lot or something where I can have founder Fridays,

Starting point is 01:06:14 but we have drinks and then I can have you and I just sit and wrap out about stuff. So those two things will be on the docket. And thanks for making Zapier. It's just such a great product and it's made life easier. warm welcome, a nice intro. You had the Zapparer makes you happier, which I love as well. I came up with that. You guys advertised on the pod years ago, and I'm like, how do you actually pronounce this?

Starting point is 01:06:38 And I think I was the one who came up with Zappear makes you happier. And so. For a one time, it ended up in the footer of the website, too. I think it still might be somewhere on the about page. I might have been the origin story of that. I'm not sure if I was or your marketing team. Thank you. Because I use it still at the time, you know, everyone was

Starting point is 01:06:55 pronounce it. Zapier, Zapier. Zapier is what we used to joke Zapier. Anyway, go to Zapier. I don't care what you call it. Yeah, exactly. Go to Zapier.com slash twist. I think the landing page is still up and we'll see you all next time. Bye bye.

Starting point is 01:07:10 On behalf of the producers and the partnership team, thank you for listening to episode 1769. We'd like to take one more time to thank our partners and broker. Use code Twist to get an extra 10% off insurance atembroker.com slash twist. Lemon.io, get 15% off your first four weeks of developer time at lemon.com slash twist and 8Sleep. Go to 8Sleep.com slash twist to check out the pod cover and get $150 off at checkout. If you are looking to become a partner of this week in startups, you can email Hannah at hana atlaunch.co. That's Hannah atlaunch.com. Thanks for listening.

This Week in Startups - Zapier Co-Founder Mike Knoop on category creation, API evolution & AI architecture | E1769

There aren't comments yet for this episode. Click on any sentence in the transcript to leave a comment.