Workday's AI Strategy, AI ‘Co-workers’, IPOs Are Back & The Future of Consumer Crypto | Sep 17, 2025

0:13Welcome everyone to the information's TITV. My name is Akash Pasricha. It is Wednesday, September 17th, and we have got a full tour of the tech sector today, folks. We have got the CTO of Workday coming on to talk about the company's latest acquisition. We are also going to talk about a big story we published this week about how Anthropic and OpenAI are headed to the gym and what that has to do with reinforcement learning. We've then got our friends Ty Haney and Dave Morin coming on the show to talk all about consumer crypto. And last but not least, we are going to get some inside analysis of what we are seeing in the IPO window right now.

0:51It's going to be a busy show. Let's get right on into our first segment. Workday made news yesterday with a$1.1 billion acquisition of an AI agents company called SANA. The company is based in Sweden. It makes tools for law firms, mining companies, and fintech companies. I want to bring on Workday CTO Peter Bayless to tell us more about what exactly he saw in the company. Peter, it's great to see you. Thanks for having me. So look, I mean, gosh, Sweden. How did you even find Sana in the first place? I didn't know Sweden was a hotspot for AI agents. Turns out, Sweden's tech is pretty good. And we got to know Sana actually originally through Workday Ventures.

1:33So Sana started in 2016 as a learning company. So Joel Hellermack, the founder, never actually went to school after age 16, but, you know, self-taught, you know, programmer and basically built a learning company where you use AI for learning. And that morphed into what Sana is today. Got it. And so tell me, how is it that Sana's AI agents are different from all of the other AI agent companies that we hear about? Yeah. So one, they're really consumer grade experience. I mean, it's a beautiful experience that Joel and team have built. I think starting from learning. Learning is a space that's pretty boring.

2:06I don't think people like to take learning classes. And somehow what they built was a ton of learning from their DNA. So they were going to make this an awesome experience to actually do corporate learning. And wait, sorry, when you say learn, I just want to make sure I understand. So AI agents and learning, how does it work? The agent can't learn for you, can it? So the company started as a learning company. So think you're going to take your information onboarding course. You guys probably have something like that, right? So, would be an option to go and actually do that coursework and that course learning.

2:36And what they did is they made it a core part of their company to basically make this a great experience, like a beautiful experience for working with that content. So, in the last couple of years, they moved over more to the agent side, which is really search, workflows, and then sort of lower code agent building. That DNA transferred over and is really clear in the product. Got it. So this is not the first AI agents related acquisition that Workday has made. You also acquired a company a couple months ago. What exactly is the customer experience that you are trying to build here at Workday with this technology?

3:10Yeah. So people have a lot of opinions about the Workday customer experience, present company included. The vision with Sauna is really simple. you know, they have probably what we think is the best user experience for working with organizational data. So doing enterprise search, getting your questions asked, creating content, consuming content. So I want to know, you know, what's going on with my deal with Acme company, you know, SANA will go and do that deep research against my enterprise data. So obviously, public data, which all the foundation models are trained on, you've got great web search APIs.

3:46Sana has built the back ends to go and pull into Salesforce, to go and federate into Google Drive, to go search my Outlook, to go search my Gmail, like to go pull all that together. And so with Sana, you get kind of two things on top of Workday. One, for the stuff you do in Workday today, you're going to much better user experience, much more intuitive, much more beautiful. Two, you typically come to Workday to take time off or, you know, you have a couple times a month max as an end user. Here it becomes a system where you think you want to come on a daily basis. the customers that have adopted the SANA agent platform, it's basically their new tab in the browser.

4:20Right. You know, we've had a lot of companies on the show that have been in a similar position to you, which is that they're trying to build their own technology. They go to the market, they find something that they want to buy. I mean, it seems like every big tech company is buying now. And I guess the question in this environment, is it just easier to buy rather than build? I think it's a mix. So, you know, we look for the best things that let us accelerate our roadmap. For some of the stuff that Workday does, it's pretty deep in the platform. Like, we do a trillion transactions a year across people's people and money data.

4:57And that has really, really, in the guts, you've got a ton of data model, object model transactions. It's a lot of stuff we've been working on, stuff we announced, you know, at our annual conference this week, where you can go and, say, apply Gen.AI to go and optimize an internal business process. like no one really in the market's probably gonna go and build that tech, but it saves, you know, hundreds of thousands of hours for customers. When it comes to something like an experience layer or something like Sana, which is, I think, you know, kind of a new, as Joel would call it, UI for AI. You know, that's something where we feel like, yes, we could go and build this, but what's the opportunity cost?

5:31And for us, I think the companies, especially with, you know, recent leadership, we're extremely aggressive. Like we think there's a right to win in this, you know, AI for work. I mean, who else is playing in this space? We could ship pretty quickly. And going to SANA is basically bringing in, you know, more founder-level energy. And also, frankly, like a complementary piece where probably could build up. We accelerate for, you know, 18 to 24 months. Peter, one of the questions that I wanted to ask you was Workday is sort of involved in the business of recruiting. And, you know, recruiting is such an interesting space right now because AI, it is certainly helping to accelerate the process of recruiting.

6:11But then you hear all these stories about, well, you know, AI is generating resumes and also AI can have some bias as it relates to, you know, scanning to these resumes. And I wondered if you could talk a little bit about how at Workday you are building your technology with those problems in mind. Yeah, it's a great question. I mean, this is, I think, true of a lot of places where people are applying AI, especially at work, right? How do you go and take advantage of these models that are really incredible at so many things, you know, especially today, things like coding, but increasing more on back office tasks without actually, you know, screwing it up in the process.

6:43So we spend a lot of time with a whole dedicated, responsible AI team, technically separate from our engineering teams to, you know, keep the separation of church and state. And frankly, I think what we're seeing with customers is that they're starting to realize that if you keep the human loop at the right places, there's a massive accelerant to go and actually, you know, really make hiring ultimately more equitable because you're able to look at more resumes, look at more candidates. I mean, this paradox acquisition we mentioned, it's sort of incredible. They're doing one interview on average per second.

7:13One interview per second. Oh, wow. On their platform of the year, every single second, do 31 million interviews, right? And the scale at which they're able to go and bring candidates in is kind of insane. And you've got people like Chipotle, right? So Chipotle reference customer, they went from 12 days on average to hire down to four days using Paradox. And then at the same time, they doubled their applicant volume. So while there's definitely, you know, you got to put the right guardrails in place. I think that AI is coming for every industry. You still leave it in the hands of a hiring manager to make the final determination.

7:44But you ultimately can do so much more. Another example I'll give you, we rolled out this performance review agent. No one likes writing performance reviews. No one really likes doing them in Workday, right? How do I make that better to go and do that? And the reality is when you write a performance review, if you're in a rush, you're just thinking about, what did I hear from this person in the last week, right? So what we do with the performance review agent is kind of like we do with Asana. You're going to suck the information about that employee over the past year, all of your one-on-one notes, all of your OKRs, all of the documentation that you've provided, and you get a much better perspective.

8:17So I think AI done right actually is really net positive for work. It's one reason why I joined the company four months ago. And we basically just need to make sure we do it in a smart way where one, it's responsible and maybe more importantly, it actually aligns to what's most important for work. Great. Well, Peter, like I said, you guys have been very acquisitive and it's an exciting spot to be right now. So I imagine there are going to be more acquisitions. And so when you continue your spree of shopping, I guess, come back on the show and talk more about some of the companies you're buying.

8:48That is Peter Bayless, the CTO of Workday. Okay, well, one of the most interesting new dynamics in AI is the way that companies are now trying to find more creative ways of training models. And in fact, we are now seeing companies use what are called reinforcement learning gyms to fortify their products. My colleague Stephanie Palazzola wrote a story about that this week. I want to bring her on. And I also want to bring on Jonathan Siddharth, whose company Turing also makes some of these Reinforcement Learning Gyms to help us explain what this is all about. Jonathan and Stephanie, it is great to have you.

9:23Great to be here. Great to be here. Thank you, Akash. Okay, so Stephanie, tell us what was your story about? There's a lot going on. We're talking about gyms, not the gyms that we know about typically. Tell us about what's happening. Totally. So basically, there's two things going on here. The first is that labs are running out of new sources of kind of large-scale, high-quality data. So at this point, all the labs have pretty much trained on the entire web, basically. The second thing that's going on is that labs are increasingly coming out with these agent products, which they're saying will go on your computer or device and take actions on behalf of users, whether that's actions in your personal life, like ordering food for you on DoorDash, or in your kind of work life, So helping you make an Excel spreadsheet, for instance.

10:10And in response to both of these things, labs are getting more creative when it comes to getting training data. So part of that is, like you've mentioned, asking companies like Turning to make basically fake copies of applications. So a fake version of Salesforce, a fake version of Excel, so they can let their AI models run around in them and kind of figure out how to complete tasks in them. And those kind of fake copies of the apps are what is known as RL environments or RL gyms. And they're also getting even more advanced experts with even more higher level degrees and working experience to label data for them to improve the models on.

10:49Got it. So, Jonathan, your company builds one of these gyms or these simulations? That's right, Akash. And just like in a regular gym, humans go to train. In an RL gym, these enterprise agents go to train. And once they train, you get to super intelligence. And as Steph mentioned, this is part of a broader shift to where we, in addition to learning from humans, we're getting to this paradigm where the agents learn from experience. I think that's the interesting, That's one interesting aspect of this. And these RL gyms are like these mini world models for business, where the gym itself is a collection of prompts and environment and verifiers.

11:42And you need expert humans to create these prompts, create these verifiers, and create that environment. and Steph's story did a really good job of making this very concrete where Steph, you and I discussed an example of let's say an investment analyst having to work through doing a DCF for let's say a particular investment opportunity. To do that, you would kind of need the agent to be able to use different tools. We need a way to verify whether the answer was correct. So we are creating thousands of these gyms across different enterprise consumer use cases. It's been a crazy, crazy build-out.

12:33Gyms, Akash, are of two types. Sometimes they are designed for computer use agents. I mean, the model is learning how to use a keyboard and a mouse to understand what's on a graphical user interface. Or it could be a gem that's focused on function calling or tool use. In some ways, it's even cooler, which is the model calling different functions and figuring out a trajectory that results in it completing the time. And then presumably the human in the loop would correct the agent, I guess, if they click the wrong button or if the DCF goes, if the cells get unlinked or something like that, then the human would correct them in the loop.

13:17That is correct. So the past trajectories unfold where the agent is trying out different paths through the space. The human can correct the trajectories, sometimes label the trajectories, sometimes nudge the model in the right way. And that data is also super valuable, this trajectory data. So a big picture here, I just want to zoom out here. So you have this new way of training. Stephanie, I mean, I presume, so all of this is really coming into focus because it's getting harder to train these models, right? I mean, this is really addressing this question of, is model efficiency, is it starting to plateau?

13:55We're trying to get ways to bring it back up again. And is that kind of the core challenge that we're trying to address here? I think that's definitely a big part of it. You know, as I mentioned, a lot of the labs, pretty much all the labs have trained on, you know, all the websites out there, all the books out there. And so I think they're just coming up with new ways to continue improving models. And this is just, you know, in the story, for instance, we mentioned that Anthropic could potentially spend up to a billion dollars on these RL environments over the next year. So they are really, you know, doubling down on this as a very promising way to continue kind of the AI improvements that we've been seeing.

14:35Right. Jonathan, you know, your business also sort of is in the, you know, recruiting experts to sort of help train these models. How do you think about sort of recruiting these experts and actually incentivizing them to come work for your company? So, as Steph mentioned, these models today are fundamentally constrained by data and compute. And Akash, you might have read that story about that MIT report that said like 95 % of GNI. Everyone was going crazy. about. Everyone is going crazy. Yeah. My hypothesis that one big reason that happens is that these models haven't seen enough enterprise data.

15:21And to do that, you have to hire enterprise domain experts from every field imaginable. But Turing, what we're doing is imagine every industry in the world, healthcare, retail, life sciences, manufacturing, imagine every function in that industry, software engineering, sales, marketing, imagine every role in the org chart with that function under marketing, performance marketing, SEO, product marketing. Imagine every workflow that that human in that role goes through. We're hiring all of them. I truly imagine evaluating models and creating data inside even something like these origins, as well as other types of data capture will be the most popular job on earth.

16:03So we might be one of the largest employers of talent in the world in a few years. So this is a massive, massive build out. You're seeing the compute side of this where companies like OpenAI and others, like with Stargate, you're talking about hundreds of billions of dollars in compute, right? Companies building at gigantic scale. So on the data side as well, companies need data to keep those data centers humming. There's this joke, Akash and Steph, among AI circles where somebody talks about how A researcher talks about how these models don't do a good job of reasoning when the data is out of distribution.

16:39Like, are they truly generalizing or are they just in patterns of the training set? And researcher two says, then we bring all data into distribution. Like all data, right? And to Steph's point, what has happened is the internet well has run dry. So right now, there's a huge amount of interest in number one, enterprise data. And for that, RL gyms are really, really helpful. In addition to other types of imitation learning. Number two is advanced STEM data. And the cool thing about a lot of these enterprise domains, as well as coding and STEM domains, is it's verifiable. So it lends itself very well for reinforcement learning.

17:22Right, right. Which is kind of cool. Like, I know everybody's famous. No, no, finish your thought quickly. We... No, everyone remembers the Move 37 from AlphaGo, where again, in AlphaZero from DeepMind, the model improved by playing against itself. RLGEMs are a way for enterprise agents to play against themselves and get rapidly better. Great. Well, Jonathan, it's a fascinating business, and I think we could probably spend hours more talking about it. But Stephanie, you had a great column this morning as well, which we didn't get time to get to. But I suggest everyone goes and reads it. That's in our AI Agenda newsletter.

18:05Stephanie and Jonathan, thank you for coming on and telling us everything we need to know about the new gyms that are taking shape here in AI. That is Stephanie Palazzolo and Jonathan Siddharth. Okay. Well, our next guests have spent a lot of time in the land of consumer crypto. It is not an area that we have talked too much about on this show, mostly because it hasn't got as much attention as it did a few years ago. But if no one else is talking about it, well, that in some ways is the best time to start talking about it. I want to bring on Ty Haney, the CEO of Try Your Best, and Dave Morin from Offline Ventures.

18:38It is their first time on the show. Ty and Dave, welcome to the show. Hey. Do you guys coordinate your – do you decide we're both going to wear hats on shows? We're very blockchain, you know, this is the look. Yeah, no, we're the consumer people, man. So Ty is wearing a TYB hat, and Dave, are you wearing an Outdoor Voices hat? I mean, is that how you get the best of both brands? No, I should be, though. Okay, well, look, I have to say, I want to talk about consumer crypto, and that's what I sort of reached out to you guys saying, let's talk about it. Ty, I was living under a rock. I didn't know that you were back at OV as of, like, what, two months ago?

19:18Yes, new news. OV, the first company, and then into TYV, which I'm still running, and then most recently rejoined OV, so very full circle and very focused on world building at the center of this community commerce opportunity. And just, I mean, share with us here very briefly, I mean, what is your sort of 18-month vision here for OV, broadly speaking? It's the same as when I started build the number one recreation brand, and that has opportunity to be well over a billion-dollar business. Dave was part of that in chapter one has been part of tyb and now again part of it in chapter two and so uh tyb is fundamentally the tool that drives this model we call community commerce which is on chain we can get into details there right that's that's a big catalyst for how we build uh this billion dollar recreation brand and and talk about so let's let's talk about tyb now what exactly is the core problem that you see the company solving here yeah 100 so consumer has a relationship issue Essentially, we, for the last 10 years, have spent a lot of money on Instagram and Facebook to acquire customers that don't stick with us.

20:23CAC has exploded. There's been no default tool for driving LTV and really longevity of these relationships and value of a customer over time. And so that's where TYB comes in. It's a community engagement and rewards platform that acts in a game in a way to allow brands to nurture relationships with their fans and customers and ultimately make them more valuable over time. And so tactically, I mean, so just paint the customer experience here for me. So I go to, I saw on your website, you've got like Urban Outfitters, for example. So I go shop at Urban Outfitters. And, you know, maybe I want to buy the cool records that they have on display or something like that.

21:02And then so what happens with the rewards? I sign up for a reward. How does it work? Yeah, so I'm a fan of Glossier. I'm prompted to join Glossier on TYB. I start to engage in these various challenges, things that as a consumer are very natural and I already do, but I'm not rewarded for them. And so really TYB becomes this one portal in on the consumer side to rewards and relationships from brands. so multiple brands all in one experience, versus traditional loyalty programs live on a brand site. And so for me to think about my parade coins and then when to use them, it's not all that relevant.

21:39So one portal into rewards and relationships across various brands is really what this phase one of TYB turns on. And our belief is the future of loyalty is all about identity, status, and your proof of fan. And so this is like posting, like if I post about the brand, then I can get rewarded for it in crypto, essentially. Well, essentially, like I'm entering this game, I'm engaging, I'm earning, and I'm progressively leveling up with this brand's tiers. And then that tier within Glossier can mean something within Nike. And so status becomes interoperable. But at the end of the day, really what from a TYB perspective, we're helping brands and consumers create is this idea of an on-chain consumer identity.

22:22Got it. Got it. That's an identity that I take with me and means something unlocks various perks, access to things, et cetera. Got it. And Dave, I'm coming to you in just a second to talk about the industry more broadly. But last question, Ty, on this topic here, what is the business model for TYB? Yeah, first phase has been B2B. So annual contracts will quickly move into this more community consumer monetized opportunity as we turn on customer as affiliate. it. And so that's a phase we're stepping into right now as we've gotten well over a million members and high value members on platform. Got it.

22:55So Dave, you know, we heard a lot about consumer crypto a couple of years ago with like NFTs, for example, you know, and Web3, it felt like people stopped talking about it, but you have continued to invest in the categories. So give us a little bit of the lay of the land here, but what are you seeing in terms of what startups are most attractive to you? And like, what should we be paying attention to as it relates to consumer crypto more broadly. Yeah, on the consumer crypto side, we've seen a real focus on real world use cases. You know, how do you take something, put it on chain that is connected to the real world?

23:33And in the case of what Ty is building at TYB, like she said, CAC has exploded. It costs more than ever to acquire a customer. And generating loyalty amongst those customers has also been largely elusive to most of the great brands in the world. And so the idea that you can convert this internet phenomenon of, you know, people becoming fans of things, of people spreading memes about what they love the most into loyalty to a brand and a higher LTV is a really powerful, I think, use of crypto. And so, you know, more broadly, I think what we're seeing is this era of mimetic finance where financial assets are spreading like memes.

24:24Financial assets have become culture, you know. We see that. Sorry. Actually, we see that. It's, you know, it's happening everywhere. Yeah. And so memes are market signals. You know, culture is the leading indicator. What people want to, you know, become a fan of, what they want to spread through social networking systems. That is their identity more than ever. You know, we've been saying that since the beginning of the consumer internet. You know, when I was first working on the earliest social networks, like we always said that identity is everything. But what you see today is this is spread not just from the internet, but into financial assets and all matter of things.

25:02And crypto is right at the dead center of this. Over the last year, we've seen a massive explosion in usage of Solana and meme coins. And a lot of people think it's silly, just like the stuff that's going on was going on with NFTs. But what I see is that consumers are engaging in cultural actions in a way that is incredibly powerful across the Internet and financial ecosystem. And so when you look at what we're doing with TYB, it's really bringing all of that together to just like radically increase the LTV for brands. So, Ty, how do you think about this idea that, you know, some percentage of the world has used crypto now?

25:47I don't know what the number is, but, you know, I imagine it's some percent. It might not be a lot, but it's growing. But my question for you is how you think about adoption about crypto broadly? Because, you know, you could make the argument that, hey, if people haven't used crypto in the first place, you know, why would they want rewards in crypto? Why would they, you know, want to sign up for a program like this? you could also make the flip side of the argument that, well, this is how we're getting them into crypto. How do you think about adoption and the chicken and the egg in terms of where you sit in that and getting people to use the cryptocurrencies?