It is true that people are getting isolated and that people aren't leaving their homes. I think it's a double whammy of the addictive nature of this technology and COVID just making people into homebodies. What are your thoughts, Mikey?
Most of this isolation does not come from AI. It predates it. It's a lot of social media. You can say there are pros and cons of social media. Imagine you can give the entire world a median competent therapist with AI. That's probably good for some of the mental health crises that we face in this country. To me, it's just somewhat incongruent to say that the correct solution to that problem is to ban AI. You're probably going to make things worse. Human fulfillment is really important. Probably working is an important part of being fulfilled, but it's actually kind of amazing when you see some of the stories and you know how many people you make smile every day.
How are you thinking about raising kids into this crazy addictive soup?
Well, I'm fortunate my kids are young enough that this is not yet a problem. But I am starting to think about the kind of environment we want them to be raised in. And I do think it's important that they are involved in the community. So we want them to be in person at school. We want them to be in dance and gymnastics and sports and music and all this other stuff because I think they get to interact with other people and they get to learn those skills. We will probably find ways to keep encouraging more of that and limit time with the phone where you're just on social media and some of these other tools.
This week in Startups is brought to you by Interpret. Interpret turns feedback noise into customer intelligence so your team knows exactly what to fix and build next. Head to interpret.com/twist to book a demo and see it in action. Squarespace. Turn your idea into a beautiful website. Go to squarespace.com/twist for a free trial. When you're ready to launch, use offer code twist to save 10% off your first purchase of a website or domain and LinkedIn ads. Start converting your B2B audience into high-quality leads today. Launch your first campaign and get $250 free when you spend at least $250. Go to linkedin.com/thisweekinstartups to claim your credit.
All right, everybody. Welcome back to This Week in Startups. And this is a new podcast we're launching in 2026. It's going to be called This Week in AI. Yes, that's right. A new podcast is coming in 2026. We'll be doing This Week in Startups three days a week. One day a week, we're going to do our round table only with experts, only with builders in the AI space. So think of this as a weekly round table in the spirit of what I did at All-In or when we do the venture capital roundtable here. And I am stoked for our first round table because we found some of the great guests we've had here on the program and people who are building in AI. So no journalists, no analysts, just people who are actually building. And we've got a full docket today. Wade Foster is here. He's the co-founder and CEO of Zapier, who I have known for a decade.
I think that's right. We confirmed prior that was the last time I was on your show. Nine years ago.
Nine years ago, which when you've been doing podcast for 15 years, you're overdue. Zapier makes you happier as everybody knows. And you started in the integration space where if people wanted to, silly example, but Gary Tan at YC started with this with Posterous back in the day. Hey, you're posting something to Instagram. What if that could automatically go to your Twitter handle? Or you posted something to Twitter. What if that could go into a Google spreadsheet and you would have a record of all your posts? But since then, a little thing called AI happened. Yeah, Wade.
That's right. We got a lot of workflows and agents and all the above inside of Zapier these days.
And we're going to get into that in the program because the year of the agent was supposed to be 2025. And we're going to talk about why it didn't happen exactly and what's going to happen in 2026. Also with us from my portfolio, one of my breakout founders, Aliari is here. He's from Micro One, which we invested in three years ago. Olly, I think.
That's right. Yeah, three years ago in the pre-seed. One of the first investors.
One of the first investors, and I know one of the most cherished, I assume, on the cap table. But the company started, if my memory serves me correct, using machine learning and AI to find the world's greatest developers. Yeah. And then something happened during that three-year journey and you pivoted. Tell everybody about that briefly.
Yeah, exactly. We started as building an AI screener specifically to vet engineers and it kind of was born out of a need of my previous company which was a software development agency. Basically, we were vetting a bunch of engineers, assigning them to projects and I developed this tool to make that easier for me. And then yeah, Micro One at first was this end-to-end recruitment engine. And then we quickly realized that the human data market actually has a bottleneck which is recruitment, which was the pivot that we ended up doing.
So now you're helping major language models with their training data. We're going to talk about training data as well. And Mikey Shulman is with us. He makes an incredible product called Suno which allows you, and we did a demo on this on This Week in Startups a couple weeks ago. It is absolutely extraordinary that civilians can now, in the same way Canva enabled people to make beautiful invites to their birthday party, now is allowing people to make incredible music. Welcome to the program, Mikey.
Thank you. Excited to be here.
Yeah. And we have a lot to talk about as well with training data. Working with the music industry which is always fun. So let's start with the talent wars. All three of you are dealing with the fact that we have an investment cycle that is unprecedented in our industry, maybe only comparable to the fiber buildout. And as part of that, not just startups, but also the major companies, the Metas of the world have been on a tear in terms of hiring people. All three of you have to compete for talent. And we had an insane moment in 2025 in the summer, which seems like 10 years ago at this point, when OpenAI was losing developers to Meta's $250 million offers. And then just this week, OpenAI decided to end vesting cliffs for new employees, which is a pretty aggressive move. For people who don't know what a vesting cliff is, just briefly, when you join a startup, your first year of equity vests, you basically get it at the end of month 12. Why do we do this? Well, you don't want to give equity to people who don't even make it to year two. That would seem profoundly unfair. So maybe Mikey, you could explain to us what it's like on the front lines right now hiring people and what impact Meta over the summer had and now maybe what impact OpenAI's new offer is because this seems like a pretty seismic shift or maybe I'm reading it too. What do you think, Mikey?
I think it's a little too early to know what the impact of the OpenAI thing is. That's very fresh but certainly people are talking about it. I think for the Meta thing over the summer, it certainly skews how everybody thinks about compensation and even if you're not directly thinking about going to work for Meta, it certainly gets everybody's head in a totally different ballpark for what normal compensation looks like. And so certainly if you don't have that kind of money, we're a small company, we're 120 people, it makes recruiting a little bit harder. For the OpenAI thing, I just think about it like this. If I'm talking to a candidate who has an offer and trying to convince them to join and I make an offer with a vesting cliff and they say I'll only accept it without a vesting cliff, I just think what does that say about that candidate? Maybe they don't actually intend to stick around for all that long. I think this worries me just about creating this really mercenary, really transient culture around building companies and something I think certainly smaller companies need to try to resist.
What is the pitch you give them to come work at Suno as opposed to say taking a crazy OpenAI or Meta offer? How do you counter that in the marketplace? Practically speaking.
For us it's a lot about the mission. If you really love music, if you really love to work at the nexus of AI and creativity, where there are no right answers, where things are messy but beautiful, I think we're kind of the best place for that. But yeah, I'll just say the uncomfortable thing out loud. If that means taking a significant pay cut compared to what Meta or OpenAI is going to pay you, it is a less compelling pitch.
Yeah. And so what is the international pipeline looking like? Because we're now in year two or three of this very vibrant competitive landscape. It has to have changed a bit in terms of developers becoming more attuned to being attractive. Ali, I'll hand it to you since you were recruiting developers. Is the pipeline of people who are qualified to work at an AI startup increasing dramatically? Are all developers basically retraining, reframing, getting focused on AI now? And is there hope that this bottleneck will be released or is it going to continue for some time?
Yeah, I think the barrier to entry for developers has reduced by a lot which in some ways makes it a bit easier. Obviously the supply kind of increases to hire developers but in some ways it makes it actually harder because now you have to really deeply vet folks. Most people can kind of do the take-home exercises pretty easily using AI and it's hard to distinguish whether they have actually done anything on the exercise themselves or if they've relied entirely on AI. So we've doubled down on our vetting process in terms of developers and the bottleneck is still there. I think the hardest bottleneck is researchers, like AI researchers in SF specifically. I think the way we think about it is other roles like finance and operations and so forth, we actually have not had a hugely difficult time hiring for, but specifically AI researchers in SF is the role that kind of rejects the offers the most and is hardest to attract, particularly because of the change in perception in total comps because of the Meta pushings that have happened. So we've been mitigating a bit by being open to hiring globally. Of course in terms of AI researchers the best talent is in SF but we are hiring a bit globally and we're improving our office space, upgrading the office and so forth to solve that bottleneck.
If you're selling a business-to-business product, you're looking to reach professional customers who are doing business and the question you have to ask yourself is where do those business leaders spend their time and the answer is obvious: LinkedIn. And these are the real decision makers. So when you advertise on LinkedIn, you're not wasting your money. You're going to target your ads based on job title, based on industry, based on skills. So maybe you want a CFO in the technology space who has been in their job for five years, right? You want to zero in on those demographics and behaviors. LinkedIn has over a billion members and 10 million of them are C-suite executives. They're going to help you reach exactly the right person with the authority to pay for your product. Right. Go right to the top. Use LinkedIn. And we've got a special offer just for Twist listeners. Launch your first campaign and get $250 free when you spend at least $250. Go to linkedin.com/thisweekinstartups to claim your credit.
All right. Wade, you've been running Zapier for how many years is this? Year 12, 13, or 14 for Zapier?
14. And in the beginning, people thought this was like a very niche idea. I remember one of your investors when you hit some crazy milestone, I don't know if it was $250 million valuation or $500 million valuation asked me, should I sell? And I was like, not if Wade's running the company because this is a tool people are going to need for some time to come. But maybe you as the elder statesman here, sort of a twofold question. One, how do you frame the company as an AI company now having started as a tools company, a SaaS company? Not sure how it was originally framed. And then how do you keep people engaged in year 14 as a startup? So there's two questions. Take them however you like.
Yeah. So I think the first one is Zapier in a lot of ways we're really fortunate. We have 8,000 integrations that connect to all these different tools. And it turns out if you want to build agents, if you want to build AI workflows, they benefit a lot from having access to data. And so a product like Zapier is maybe easier to adapt in the age of AI compared to some of these other companies that were built on pre-AI era concepts. And so I think a lot of our folks see the opportunity. They feel like it's well positioned. In terms of how we retain and compete for talent, I think Ali was kind of hitting it on the head. There's a handful of axes here where I see the talent battles competing on. There's AI researchers versus everyone else, and obviously AI researchers are highly in demand whereas in the everywhere else camp, you can still find great developers across the globe. I think you've got SF versus everywhere else. Obviously the talent wars in SF are very intense. Not everyone wants to live in SF. You moved to Austin. I moved to Missouri. There's a lot of people spread out across the United States, across the globe that are very talented and you can tap into that. I think the other axis we see a lot of challenges on are execs or leaders who are actually using the technology themselves. There's a lot of people who rose through the ranks in the 2010s who have kind of stepped into more of a purely management role. And those folks tend to run 2015 playbooks. And those 2015 playbooks just don't work particularly well right now. And so you have to find leaders who have some of that leadership management experience, but are in the trenches playing with the tools, playing with the technology, understanding how it changes the game on the field. And that's another axis that I think is pretty tough to find right now. So if you look at these different axes and think about where you want to recruit from, you got to find the pocket where you have an advantage. For Zapier, one of the things we benefited from the longest time is that we said we'll hire remotely, we'll hire distributed. And that maybe became less of an advantage during COVID, but now that everybody's doing return to office, it's kind of become a bit more of an advantage again because we're happy to take you wherever you're at.
Tell me about the pipeline of developers. Do we think, and we'll go around the horn here, the co-pilots and how easy it is to onboard and become good at UX using these different tools and the onboarding at Replit and just being able to sort of become a developer. Are we seeing an increase, Wade, in the number of developers and number of people who are developer curious? And is that pipeline increasing? Because that was the bottleneck previously is we just didn't have enough people going for computer science degrees and becoming developers. Is that increasing radically because of AI or less so than you would think?
Yeah, you're seeing a lot of folks who are able to ship and build things. That is definitely coming true. Now, are all of these folks capable of deploying production applications? And that's where I'm like, there is still a hurdle there where you need to understand foundational skills to do a good job at building and scaling tools that you would make commercially available. But even though there is a gap there, it is still a good thing that more people are able to build stuff because it shortens the overall software development life cycle. You can have a sales rep come in and say, 'Hey, I prototyped this feature concept that I was talking to this customer about. Do you think this idea would be good?' And that is a way tighter communication loop with a product manager or an engineer than in the past where you might have a sales rep come in and just say, 'Hey, someone asked for blah blah blah feature.' And it's like, I don't quite understand what you're saying. I can't quite see it in your head. And so it is really beneficial that we have these tools that enable everybody else in a company to get their ideas out of their head and express them in a way that more people can quickly understand, 'Oh, I see what you're getting at here.'
And so it's in a way these tools replace the whiteboard session at a company where some product manager or some frontline salesperson is like, 'Hey, we should have this feature. Let me draw for you on the whiteboard.' It's like, 'Well, let me just make my own wireframes.'
Yeah, I think that's right. I think it replaces the whiteboard session. I think you see a lot more internal tools. I think you see a lot more throwaway code where it's like, 'Hey, I have a one-off task and I'm just going to do it' instead of in the past you might choose not to even do that thing. You just be like, 'I can't do it. Whatever. I'm not going to worry about it.' And you sort of go along with your day. And so there's a lot more of these micro problems in organizations that you can actually solve now because you can throw it to Claude Code or something like that to do the work for you.
Mike, what are you seeing in terms of the developer pipeline? Is it increasing or is it not a major issue for you because you've intentionally kept the company small? How many total employees do you have, Wade, and how many do you have, Mikey?
We're at 120 but that is not intentionally small. We're trying to grow as quickly as humanly possible.
750. Yeah. So Mikey, tell me how do you think about the pipeline training and maybe people who are developer curious being able to digitally with a vibe coding tool or Claude Code make these little solutions that maybe aren't production ready, but they certainly around the table make the conversation move faster. Are you seeing that as well inside of Suno?
100%. I'll echo everything that Wade said. Maybe just a fun data point. I think the last line of code I wrote was a little more than 12 months ago, maybe 18 months ago. And then at the time maybe 15% of people here were using these things and now literally 100% of developers and many non-developers are using these tools. And I think the thing that's super interesting is that when you start to be able to solve these micro problems quickly, when you start to be able to make these prototypes quickly, you shorten a lot of the iteration cycles, it changes how you work. I actually think it makes more of the interpersonal stuff more important. It reminds you how much of a team sport even engineering is. People have some platonic ideal of an engineer sitting in a room by themselves just hacking away, but actually when a lot of that work happens much more quickly, you have to talk a lot more. And I think that's a good thing. Talking to people is fun. But it has totally changed how we hire, how we develop code, how we develop products. It's kind of remarkable.
It's time to talk about one of my favorite companies in the world. Not just a sponsor, but a true friend of this podcast. Of course, I mean the one, the only, Squarespace. Squarespace is the number one best way to start a website for your company. And that's going to help you stand out online. Whether you're just getting started with your website or you're upgrading and scaling your existing business like I'm constantly doing, Squarespace is going to give you every tool you need: a great domain name, the ability to book meetings, sell content online, and it's going to be gorgeous because they have a new AI-powered blueprint website builder you can use, or you can just use a template. It's up to you. And with a couple of simple prompts with the blueprint tool, you can design an entirely customizable website that perfectly suits your aesthetic, your voice, your personality. Support Twist's longest-running sponsor by going to squarespace.com/twist for a free trial. And when you're ready to launch, go to squarespace.com/twist for 10% off your first website or domain purchase.
Olly, what are you seeing game on the field in terms of rank and file employees participating in the creation of prototypes? And what are your customers telling you?
Yeah, I think very similar to Wade and Mikey here, we're accelerating the cycles by a lot. I think in a kind of paradoxical way, design has actually sort of become the bottleneck instead of engineering. And so we actually for some of our product buildouts are experimenting with not designing first and just having engineers create the prototype and jumping to the iterations that are required on the engineering end. And then once it comes to the state that they believe is functional and looks good in terms of UI/UX, we then kind of backtrack and create the design files then and fix up some UI aspects afterwards. So this is a shift that may happen across organizations where we actually don't get bottlenecked by Figma designs and Figma designs come after the engineer has actually developed the product. I think the second thing is very similar to what Wade was saying: we're seeing a lot of admin tools, like admin portal tools and kind of internal software being built very quickly in a very lightweight and purpose-built manner where a lot of software subscriptions become unnecessary and frankly the flow of using that very lightweight internal software is actually superior to any subscription that you're going to make to a much heavier weight software. So I think those are two things that we're seeing. And then the last thing I would say is this results in, as we said earlier, higher supply of engineers because the barrier to entry is lower. So you have to kind of rethink the vetting process. So what we've done is we've moved away from these kind of DSA-type coding problems in our interview process to a much more 'use AI, build the thing with AI actually and then just explain your thought process' and then we take that thought process and analyze it and see if they're actually an engineer.
Does this mean we're going to a singularity in terms of startup employees who are valuable, Ali, where the person who can identify a problem, vibe code a solution, Claude code a solution whatever it is, and then the design is perfunctory and done at the end. It's kind of like we're moving towards one type of person who is the identifier and builder of the solution and just good enough so I don't have to hire some verticalized SaaS product.
I think we're going to see a similar trend in a lot of different domains. So I wouldn't say the singularity is for engineers only. I think engineers operating these machines and being way better engineers is the same analogy that you can make for recruiters operating an AI recruiter machine and not doing interviews anymore and being much faster at their job and embracing the more creative parts of their job versus continuous interviews back and forth all day. So I think if you apply this same kind of structure of 'let me be an operator of an AI agent as the main part of my job and let me try to get to that as fast as possible,' then each of those domains will have the singularity of this idea of running an AI agent basically.
Wait, what happens to Gen Xers who had like a more verticalized approach where like 'I'm a designer, I am a product manager'? When you have these, I'll assume 30-something Gen Z's, elder millennials or younger millennials even who can just kind of wrap the whole thing up in a bow and bring it to management.
Well, I think the skill that is really valuable these days is not how much you know, it's how adaptable you are. And an interesting thing that I've seen is there's this debate like who benefits most from AI. Is it the folks who are senior, the best at their skills, or is it the junior folks who can now look work fastest? And I think clearly both are benefiting a lot. But what has been interesting for me to see is the folks who are really sharp at their domain and who are adaptable and interested in learning are just doing some insane stuff right now. And they're able to collapse the skill stack in a way that I think was just harder before. It's faster than ever to go from 0 to 80 on a particular domain. And so now if you were like a really good engineer who was interested in design or interested in marketing or interested in some of these other things and you had good ideas, the types of things you can do now are very impressive. And so I think the Gen Xers who have spent their years in the trenches learning a lot, honing their craft and really know how to do that who are also embracing AI, I think you're probably very well positioned to build some incredible stuff these days.
Best advice for a young person graduating school who can't get that entry-level job. What's your best advice to them to break in? Because we are seeing this even in big tech, we're seeing in traditional companies where there's somebody on the team who's an elder statesman as you're talking about. And talking about Gen X versus elder statesman is just crazy for me because we were the punk rock kids 15, 20 years ago when Web 2.0 came out. What's the best advice then for young people entering the workforce who just don't have a job offer from Google or Meta or Apple today and they're kind of concerned that nobody is handholding them into a career?
I think that curiosity is great. The thing I remember, I graduated from college in the financial crisis and no one was hiring then even if you were a good student. But this was kind of the age of the internet. So around when I was graduating you saw like Stripe and Twilio and Dropbox and Airbnb, these companies were coming online. And what you could do was you could just go make things, you could build things. You could learn to code. You could figure out how to ship things. And you could email founders and say, 'Hey, I made this cool thing. I think your company could do this, that, or the other.' And that sort of high agency creativity, I don't think that ever goes out of style. And in fact, I think it stands out even more so today. And so if I was in that same situation again, I would be using all these tools, I would be experimenting with them. I would be building them. I would be finding products I'd love. I'd be hitting Mikey up and saying, 'I love Suno. This is incredible. I think you all should do this, that, and the other.' And here I made a suggested feature. And I bet good money if I did that, Mikey would email me back and say, 'Hey, let's talk.'
Mikey, has that happened? Is that actually happening that some number of young folks realize sending resumes and doing an AI interview with a hundred AI agents and trying to get into a company just not as effective as just doing some spec work, building a prototype and emailing it to you and telling you, 'Hey, Suno doesn't have saxophones that are really great and I vibe coded and I got my friend who plays sax to do a bunch of riffs of Mark Knopfler songs and we fixed it for you here. Here's where I think you should be going.' Do you actually have people doing that or not?
Every day. Our first Android engineer sent us a little app that he made that was like we used to have this homepage with this big wheel of all the different weird mashups of genres and it's on the web and it's really made for mobile because it's interactive and he just built it and sent it to us and that was our first Android engineer.
We interviewed him. But yes, then we hired him.
This is just so obvious, but young people are like, 'Well, if I'm not getting paid for it, why would I do that?' The game on the field, Ali, now is you're in a global competition. And you have your hooks into all the global markets. People in other markets are doing this because they really want the work.
Yeah, exactly. I was actually going to say, adding on to Mikey's point, we actually hired our last two AI engineers also in a very similar fashion. One of them built a prototype of our end-to-end interviewer and said, 'Look, I reduced the latency in this one specific part. We should chat.' And we interviewed him and hired him. Another example was someone built a very similar model to our proctoring model that exists in the interviews and he did a great job and we ended up interviewing him and hiring him as well. So I would certainly echo Wade and Mikey's points here, which is because it's easier to build, you should use that ability and build stuff very quickly and just email the exact team of these companies that you're interested in and it is very likely that you will get a response back. And I think the other thing I would say is we recently actually did a study on the impacts of AI as well as AI interviews specifically in the hiring market. And what we saw interestingly is that younger folks and less experienced folks are actually benefiting from AI interviews. And the reason for it is simple. If you don't go through an interview process, the only signal you have is one's resume. And of course, if someone is more experienced, the signals in one's resume is by far greater. And so there's this positive feedback loop that exists that you can't really exit unless you don't only rely on folks' resumes. So what we saw is that inexperienced folks that are willing to grind and learn and so forth actually did better in interviews and ended up subsequently getting hired at a higher rate.
A very simple way to say this is do the job that you want and somebody will hire you. But you just start doing the job every day and somebody will hire you. We took our hiring process and I moved it because I saw Naval did something really interesting. He has Naval Ravikant, founder and co-founder of AngelList, in his feed where he has this little podcast feed where he just does three minute thought pieces. He just said, 'I'm looking for a chief of staff and I'm looking for a producer. We are high agency independent workers. We work remote. If you're interested, here's the email address. Tell us what you can do. Show us what you can do.' That's it. That was his job description. That was his process. So I was like, 'Well, that's pretty smart. We need a community person here for our founders.' And we just popped up the email
[email protected] and said, 'Email us what community experience, what you've done and what you would do for communities around startups and angel investors and seed investors.' And we did the same thing. We wanted a CapCut editor. So I said, 'Hey, if we want them to do CapCut, just create
[email protected] and have them email us clips of the existing show.' And we paid them for it. But my lord, the hiring process went so much faster when we just simplified it and said, 'Show us you doing this task.' The best way to tell if somebody can do a task is not to talk to them. It's to look at their work. I think that's something all CEOs know inherently and all high agency people know, but a lot of young people in the workforce are trying to figure out. All right, let's go to our next topic. Training data. Ali, you do this for a living at Micro One. And Mikey, you're very focused on the music space. So I got a lot of questions about this. How much training data is left, Ali, in the world that hasn't been scraped, open crawled, etc. or licensed now because we're entering this age of licensing where Reddit, Quora are building really nice businesses? Obviously Elon bought Twitter in order to have proprietary training data and that's why Grok is so good at real time queries. Is there any data left and how are these large language models or verticalized models creating data? A bit of a softball question since this is what Micro One does for a living.
Yeah, so there is no data left. The internet has obviously been trained on and I think there's a large scale kind of equivalent data sets that have similar quality. So training on them actually doesn't really meaningfully improve models. So the way we look at it is we're in the phase of post-training where net new data is required and that net new data is created by humans, oftentimes human experts and subject matter experts in specific domains that are creating very complex high-fidelity data sets to continue improving model capabilities. And I think the route that foundational model companies are taking is they're trying to improve in the highest economically valuable domains and they try to improve with data structures that are emergent, which means improving not only in that domain but also improving the model generally.
If you're not listening to your customers and iterating based on their feedback, your startup will fail. So if you're drowning in feedback from support, reviews, calls, it's time for Interpret. Interpret is the customer intelligence platform that helps you turn random comments and feedback into actionable insights that help grow your company. Their AI system is already trained on your business and your product. It then reads all your support tickets. It reads all your reviews, call transcripts, and surveys. Then their agents actually update your team in Jira, Linear, Zendesk, and Slack. So this feedback doesn't just sit in some email or messaging thread, but it actually helps keep your team moving the ball forward. Find out why Canva, Notion, and Perplexity are already using Interpret to stay on top of what they need to fix or build next. So if you're ready to turn feedback chaos into customer intelligence, head to interpret.com/twist to book a demo and see it in action. That's interpret.com/twist.
Give some examples of the top high-value verticals. Obviously people have watched, we have an investment in a company called TaxGPT that's doing tax law specifically which is granular, which is detailed, which you need expertise in, which is always changing. Obviously tax laws change every day globally in hundreds of markets and thousands of different little niches. So I'm assuming accounting and legal are obvious ones in there and maybe coding. What else is in this list? What are your clients asking for generally speaking?
Yeah, the three main categories are finance, medical and legal, and coding is a very close fourth. Those are the ones that we sort of focus on based on the demand that we see. I think the current state of the market is very much still in this phase of getting the models to be really good at question and answer pairs, really complex questions and very complex domains. But we're now starting to enter the phase of models actually doing things in those domains. And I think tax is actually a really good example. If you think about what is a question answer pair in tax, it may be giving your income of the year and then asking for tax advice on purchases you can make before end of the year. And the models do a pretty good job at answering these singular questions. But where models fail very much so currently is when they have to answer multiple questions in a row, which is equivalent to saying they have to do things. And the simplest way to think about this is if you assume that any given domain has 90% accuracy in answering any arbitrary question, if you have to answer multiple in a row, let's say five questions in a row to actually do some actions, and you do 0.9 to the power of five, you're going to very quickly go to zero. The accuracy is going to converge towards zero very quickly, which is why training data for tasks that have these signature actions is much harder. So one of the environments that we're creating is a tax environment specifically for California W2 taxes. And as you can imagine, there's lots of steps to filing someone's taxes, such as collecting input data from them first, making sure you actually have the right input data. There's some partial rewards for that. And then there's a conversation that happens around tax advice and a bunch of steps in between. And then ultimately the tax is filed. And each of those steps depends on each of the previous states. And the previous states need to be accurate for the following states to be accurate as well.
What do you call this career where a tax expert who's 10 years into their career can say, 'You know what? I don't like clients anymore or customers. I just want to answer questions and be heads down talking to an LLM and making it better.' What is the career path that this is now, Ali?
I don't think it's mutually exclusive to not liking clients and wanting to just train models. I think the way we see it is these experts do both. The category is called AI training, you know, AI training experts or human data experts. In fact, one of our competitors actually just came out with a study that there's over 200,000 people doing these type of jobs now. I believe just in the US, it might be internationally, but I think it was just in the US. And so there's a lot of folks that are spending 15, 20 hours a week of their time training models. And I think the way to frame it is they are getting paid to do things they're already really good at, and they're getting paid often more than their day job in almost every case actually, to train models that will help them do their job in the future. And I think that's a nice way to look at it.
HDEs, human data experts, I love it. And they could be doing this freelance as a side hustle, doing it very quietly. They maybe work for a big four accounting firm or whatever and then on the weekends they get $200 an hour, $300 an hour work to do this kind of stuff.
And I don't think most people are aware of this fact that this is what training data has moved to. Wade, you have, I'm assuming, hundreds of thousands of people every month architecting workflows which you can then have as proprietary training for the Zapier god operational. I don't know what you call this person, but there's some air traffic controller in the Zapier brain. Tell us about that. Are you building the Zapier brain of operations based on all this ticky-tacky stuff that my team's doing to take data from one repository, get it into Slack, and then bring it back over to Notion and move it out of email into structured data? Tell us about your thoughts on training data.
Yeah, I think any company that has proprietary data is trying to figure out, 'Hey, can we use this to provide a better product experience, a better application than the foundation models can directly provide themselves?' And so in the case of us, we're sitting on 14 years of workflow data. And as a result, we know how to make these things reliable and when things break we know how to fix them. And we can use that data because very few other people have access to the corpus of information we have. And so we can use that to improve the experience. And that's kind of what most companies that have these large corpus of bodies are trying to figure out: how do we build a better product experience because we have access to this and no one else does? To Ali's point, all the public information has been soaked up at this point in time, but the private stuff is where a lot of the alpha is today.
Have you been approached by the Sam Altmans of the world, the Claudes of the world to take your data set and they license it from you for $50 million a year? And how do you think about that? Because I have had startups who've had this kind of internal discussion in board meetings. 'Hey, should we keep this proprietary or should we license it?' Obviously Quora and Stack Overflow, they've gone the licensing route. Do you think, 'Hey, maybe I should just license this to an LLM?'
Well, I think it's no secret that the foundational models have an insatiable appetite for training data and so they're going out and spending large amounts of money to get access to these proprietary tools. I think our approach is we'd rather keep that for ourselves. But I do think there's benefit to releasing benchmarks and eval tools that can help you get a sense for how well these models perform against these tools. And if you make those public, you'll likely find that the foundation models will start training on those things and that can be beneficial to you because you're getting some cooperation with everybody to improve on a certain thing that you actually wish the models were better at because if they were, you could create a better experience for your customers. And so I do think a lot of companies are looking at how they release benchmarks.
And are you integrating yourself into OpenAI's? I don't know what they call them now. I think they were originally called apps or integrations.
Yeah, connectors. How do you think about OpenAI's connectors as a competitor or a collaborator for Zapier?
I think it's a pretty important collaborator. So we're a connector inside of ChatGPT. You can use us as an MCP provider inside their agent builder.
Explain what that is. MCP for people who don't know yet.
So MCP is Model Context Protocol. You can almost think of it like APIs for agents to talk to each other, that's probably a simple way if you're unfamiliar. It's really important that these models have access to tools and data. If they don't have access to them, you're not going to get a great answer. So for example, imagine you're running an insurance shop in the middle of the country and you want to ask ChatGPT, 'Can you generate a marketing plan for me?' Well, you'll probably get a marketing plan back that you'll go, 'Well, it's decent, but it knows nothing about my business and so I'm not actually going to be willing to use it.' However, if you say, 'Hey, ChatGPT, let me give you access to the last 10 campaigns I ran. Let me give you access to the types of customers that I like and the types of customers that I dislike. Let me give you access to my CRM data that has all that closed won data, the closed loss data. Let me give you access to this, that, and the other. Now, how about you go try again and generate that marketing campaign for me?' And you'll find that you get a much better output after giving access to your data, your tools, and things like that. And so Zapier plus ChatGPT or Claude or Gemini is a pretty powerful combination.
Yeah. And that means they have to have an account and if they get value from it, they put their credit card in and you get a new customer. So it's a win-win. Mikey, you have the most difficult, I think challenging task which is the music industry. They're pretty protective of their IP and very litigious obviously and then it's super valuable. They also have to contend with their artists and does Mark Knopfler want to be part of this and have people, you know, the first thing I do whenever I use one of your tools or one of them is hey make me a Mark Knopfler song because Dire Straits hasn't been making albums since On Every Street in the early 90s. So I'm desperate for a Dire Straits album. How do you think about training data? Are you hiring musicians and experts to make riffs and play the blues and then training your data with proprietary data like that, or is it really you just have to make relationships with all the labels and the artists and that's the mode in your business? Because there are these sampling websites, right? There's a number of sampling websites that let you buy samples and build songs. So explain to us the game on the field right now. Splice, right? That's the name of it. Splice, because that's all open, that's all royalty-free.
I don't want to say one way or the other because I don't know exactly how Splice does it, but yeah, I mean actually the funny thing is everybody here is actually a musician, or almost everybody here is actually a musician. That's what I was talking about earlier, that the mission alignment is ultimately what gets people to want to come work here instead of getting a lot more at OpenAI. I think there's a lot to say. I have always said though that yes, it is an organized and sometimes litigious industry. It's not for people without a thick stomach lining, but at the same time there's a lot more to do together than fighting one another. I've thought that since the day we started this. And I understand that music and technology don't have the most straightforward past. And so there's a lot of you need to show and not tell, but when you do actually, people get it. And AI can be a tool to make the music industry bigger, to deepen the artist-fan engagement, to make amazing tools to let people make music that they have never been able to make before. And once people see that, they actually kind of get on board. This was recently put to me by someone as, we're the GLP-1s of the music industry. Everybody's on them, but nobody wants to talk about it. And I think slowly eventually people will start to admit that they're using and enjoying the stuff that we make.
I can tell you I am on GLP-1s. I have for four years since my friend Kevin Rose and Tim Ferriss talked about the awful use of them and I lost 43 pounds. And you can go to Ro.co, which is where I get my GLP-1s, and I'm now a spokesperson. Did you guys know I'm sponsored by Ro.co?
I'm an official spokesperson. Me, Serena Williams, and Charles Barkley. Three exceptional individuals at their craft. So I've done two of these. It's not an influencer deal. It's like being a brand. It's like being a celebrity. So those are my new celebrity endorsements. I'm an endorser. It's my endorsement deal. But they are using it and they're making great music. And Mikey, they don't need to be precious about this tool, do they? Because a lot of the great music of this century has been based on sampling some incredible David Bowie song, some incredible hook. I mean, Kanye West, Diddy, no Diddy, you know, whatever. I was in jail and this Diddy documentary is pretty brutal, but you know, he was famous for sampling the Sting song 'Every Breath You Take' in his tribute to Biggie Smalls. This is not new to the industry and they actually are particularly good at figuring out how to create compensation already for sampled music.
