He Built Siri at Apple. Now His Private AI Runs Inside the FBI, Google, and Netflix.

Ivan Lee: And so with all the new kind of security concerns or the fact that they're... They need more data to train their next generation models, our clients can kind of rest assured that this data is never reaching them.Jordan Metzner: Built This Week, breaking it down. Built This Week, we show you how. A fresh idea, aSam Nadler: clever tweak you locked in. You built this week. Hey, everyone, and welcome to Built This Week, the podcast where we share what we're building, how we're building it, and what it means for the world of AI and startups. I'm Sam Nadler, cofounder here at Rise Labs. And each and every week, I'm joined by my friend, cohost, and business partner, Jordan Metzner.How are doing today, Jordan?Jordan Metzner: Hey, Sam. How's it going? Yeah. Exciting week in AI. Lots of news from all the major labs, lots of new cool stuff, like, we'll talk about later, but these agent bots are are making a shine, and we're seeing some some new change in who's the market leader in them as well, obviously, kind of a little shift in the in the power kingdom.But I'm even more excited about today's guest, so maybe we should get this thing started.Sam Nadler: With that, I'd love to introduce Ivan with DataSore. And, Ivan, if you don't mind, please tell us a little bit about yourself, and then we'll jump into what you're building.Ivan Lee: Yeah. Absolutely. My name is Ivan. I'm the cofounder and CEO here at Datasore. I've been working on this for eight years now.And my own background, I studied computer science. I was an engineer by training. I've been a product manager for my entire career up through starting this company. Built AI products at Yahoo and Apple. This is my second startup.And, yeah, excited to talk about the news and what's been happening in AI. It really has been a really fun year.Sam Nadler: Awesome. So I think we're gonna get we're gonna get into the details. But if you don't mind, just tell us a little bit about what you're building. You know, I think it, targets a really interesting use case of companies really trying to protect their data. You know, most of us are just sending our data to OpenAI and Anthropic at every moment.So give us a little insight of of of why you decided to build this and and the types of companies using it.Ivan Lee: So Datasore fundamentally builds private AI for enterprise. And what that means is, we set it up so that the AI is deployed within your servers, within your infrastructure. So none of the sensitive data, whether that be clinical patient data or, social security numbers, none of that is being sent to Anthropic or OpenAI. And so with all the new kind of security concerns or the fact that they're... They need more data to train their next generation models, our clients can kind of rest assured that this data is never reaching them.The way we do this is by deploying open weights models, directly onto the servers along with the harness, the the context, everything else needed to make these agents work. And we largely work with, health care, finance, legal, public sector, folks who would care about this kind of privacy and security.Sam Nadler: Yeah. I think it's, I think there's gonna be a huge shift to doing this over, you know, I don't know exact time period. And, obviously, there's the immediate use case with, like, HIPAA and health care. Is there any, like, surprising... And legal, I'm assuming, but is there any, like, surprising industries that, you know, have been quick to adopt a a kind of a private solution like the one you're offering?Ivan Lee: Well, surprising in a couple of ways. I think, first, historically, health care and and legal, these industries are notorious for moving more conservatively, more slowly. But the use case is just so compelling that they really do... They've been moving a lot more quickly to adopt this technology than perhaps any other technology in history, cloud, mobile included. Right?And so there is just this immense appetite to to be adopting LLMs and AI agents. But at the same time, we still... They still have to take into consideration exactly what you said, HIPAA, SOC two compliance, just increasingly... It started with, you know, cost consideration and compliance consideration, but I think one thing that's really shifted this year is a lot of companies realizing that these are... Their subject matter expertise, their domain expertise is their moat.That is their kind of... The thing that has kept them competitive over the last few decades. And so there's more and more examples of Anthropic partnering with a company and then launching a product of their own in that same category. Right? There's Claude Design.There's Claude Legal. And so increasingly, these companies are realizing, wait. Am I comfortable sharing all these domain secrets or know how in, you know, these really complex processes and just partnering with them to teach them exactly how this is done? Pair that alongside the fact that there's been a whole new wave of open weight models this year, and suddenly, there is both motive as well as opportunity to build your own AI and turn that into a strategic asset.Sam Nadler: Awesome. And is there any... And, obviously, I'm I'm guessing the answer is no. They're not comfortable sharing their, you know, proprietary knowledge with Anthropic and OpenAI. It...But is there... Are... Is there a performance gap, or or are you seeing that performance gap close?Ivan Lee: That's a great question. So that is actually, the most... That's the the number one question that a lot of these companies ask. I think, you know, it's been a whirlwind of four years. It's only been four years since this technology was invented.And in the first couple of years, there was an immense performance gap. If you recall, just a couple of years ago, OpenAI was so far ahead of the pack, it was inconceivable that anyone would catch up. Right? Fast forward to March of this year, and if you weren't using Anthropix Cloud here in Silicon Valley, people are questioning your sanity. Right?And like you said at the beginning of the podcast, you know, the the market leader is is constantly shifting now. So we've gone to this point where the models are largely commoditized even amongst the proprietary models. And then on top of that, we've had this, like I said, wave of open weight model releases. Some of those come from China, but recently, Meta reentered the open weights game with a really powerful model of their own. We've been big fans of Google's Gemma four here, at Datasore.So there's a lot of options, actually. And when it comes to the performance gap, absolutely. Like, if I were working on solving cancer, I would hope... Or, like, if our clients were working on solving cancer, I hope they're using Astral and Fable. But if you're using this to summarize your emails, if you're using this to recap your meeting notes, you don't need Fable for that.That is absolutely overkill. And that's where a lot of the low hanging fruit, a lot of the enterprise agents that are necessary out there, these open weights models are more than good enough.Jordan Metzner: Okay. I think that's like a super interesting point because that basically means we're reaching the asymptote of like log... Rig myth curve. Right? Essentially, where they're kind of all starting to meet at the same place and the marginal gains are are diminutive.And what I think is proof of that is if... I think Claude announced Duet yesterday or the day before, and they announced that they built it on top of Opus five and a half. And I think, as you mentioned, now we've got, you know, a few other models that are, I would say, pretty close to comparable to OPUS 5.5 on the market, especially in the open source world. And so that line just keeps getting closer and closer and closer together, right, and flatter and flatter and flatter to your point as well. So maybe you can talk to a little bit about kind of how that has an impact on hardware and demands and kind of...I mean, I know I I think they're talking about AWS being out of hardware, and everybody wants kind of the same hardware to run these machines. So tell us a little bit about kind of, like, how hardware has had an impact on your ability to to create these kind of private private clouds, for lack of a better term.Ivan Lee: Yeah. Absolutely. Well, I think I would point to Meta's most recent announcement. They have this incredibly powerful 100,000,000,000 parameter model that they just released. And alongside that announcement, they mentioned that it can...It's built and tested to run on two m five Mac studios. Now for maybe consumer you and me, that's a little expensive for us to bring into our homes. But for the average company, that's, you know, low 5 figures, to have your own 100,000,000,000 parameter model. That is incredible. So I think, yes, there is a factor of, like, the hardware.Like, we've had clients who have bought millions of dollars of new hardware just to run their private AI solutions, but I think that cost just keeps plummeting over the next couple of years as you get more powerful models that can run on increasingly commoditized hardware. Let me take a moment. I wanna actually share a a slide that kind of speaks to the, the gap in model performance. This slide is a snapshot of a very powerful kind of benchmarking site called artificial analysis, but I wanna draw your attention to the two black lines. First of all, the lower black line, this is GPT four o.This is a model that was deprecated at the beginning of this year, but a lot of people lamented because they actually really liked it. It was that first model that really started feeling kinda human and really started being able to answer questions legitimately. Legitimately. There are so many tasks that can be solved with the four o model, and you can see how low that is sitting here in amongst its peers. Right?The second model is... Here in Silicon Valley, we've been selling agents for a couple of years. But realistically, pragmatically, agents didn't start working until Claude 4.5, until December 2025, and that's the second bar. Everything above that is already at 4.5 or smarter. Right?So here's another view of the same picture, but every single blue model is open weights. That means that DeepSeek 4.1, MIMO... I don't think this, when I took the screenshot, Meta's MUSE hadn't been tested yet, but every single one of these clears that by a significant margin. So, again, there will be 5%, 10% of tasks that should be sent to 5.5 max, fable, astral, whatever, mythos. But there's so many things for a lot of these clients, a lot of low and middle hanging fruit.We could pause all model development now for the next decade, and we would still be gathering a lot of improvements across the enterprise just using these open weights models alone.Jordan Metzner: Okay. Then tell us the Jeviance paradox here because then what happens? Right?Ivan Lee: So that's where I think there is... I personally feel like there is a lot more to... There's a lot more tokens to be consumed on the enterprise front. There are two things that are happening simultaneously. One, tokens right now, if you...Your average chatbot query is 200 to 2,000 tokens. Your average agentic task is 3,000,000 tokens. It's over three orders of magnitude more than your chat GPT query. Right? That is why Jensen and Sam and Dario love h s.Right? But they really are that much more capable. And so you have significantly more token consumption alongside token maxing. That's why we had this kinda like... We actually got into a GPU glut earlier this year where Anthropic had to slow down its services.OpenAI had to stop working with, you know, offering on on Gemini, on GCP. We got into this glut because people were overspending. So on the one hand, as more agentic adoption occurs, we're going to need a lot more tokens. And then in parallel to that, we're still... Are you familiar with the crossing the chasm paradigm?Jordan Metzner: Maybe not.Ivan Lee: Crossing the chasm is this kind of, like, MBA chart. I'll pull up a a version of it here. There's this MBA chart that everybody alludes to where every technology has to have this adoption curve where, essentially, every technology, whether you want to talk about the Internet, mobile, the microwave, the television, everything goes through innovators, early adopters. Then there's a chasm, and some companies... And some technologies never cross this chasm.There's a early majority, late majority, then think of your, you know, friends who were the last to get smartphones. There's the laggards. So you can actually chart every single technology adoption across this same curve. Now I would argue that chatbots are here. We're we're at a billion monthly active users for ChatGPT and Gemini, and I think we're already in that early majority.It's already crossed that chasm. Contrast this with something like VR where it's still at the early adopter stage and may never cross that chasm in its current state. Right? Now I think agents, on the other hand, agents are at the early adopter stage and we haven't crossed that chasm yet. But when we do, this is absolutely going to skyrocket token usage even further because a lot more companies are going to start adopting this over their Microsoft Copilot ChatGPT tool right now, and then they're gonna use way more tokens, and it's going...And that same shock that hit Silicon Valley earlier this year is going to hit the rest of the world, really.Jordan Metzner: I totally agree with you. I think... I mean, Sam and I have both been using Rockbot a lot. I've gone... Kinda gone through some of the agents, but what I realized is it also collapses some of my software spend.Right? And so that's like a shift of, like, CPU to GPU. Right? Essentially, the same thing, like, kind of to your mentioning. Like, I don't need to pay for some software that was doing some specific thing for me in, like, some outbound sales marketing tool, and I can have my agent do it for me.And so it's just like CPU away, CPU away, all shift to tokens. But again, like, does that make the token... As the tokens get cheaper to your point running on these Macs or what what... You know, whatever the next iteration of that is. But I guess gross token consumption of mine is...Will significantly go up. Right? And same with the enterprise.Ivan Lee: It's it's a really interesting kind of macroeconomic situation with all these variables shifting. Right? Everybody's looking to make the models not only more intelligent but more efficient. There's obviously new GPU and hardware, research as well. There's so much money being thrown at this that, like, the scales are being tipped on both sides.Sam Nadler: I have one last question, if you don't mind, before we jump into the news. And and maybe it's obvious, but what is your typical client engagement look like? A client approaches you. They said, we're we're interested in the solution. Is it almost like a consulting agreement where you help them get set up?You help them choose the the the the model that's most appropriate? You help them, you know, set up the the deploy deploy it appropriately across their organization, ensure it goes well, and then is it kind of hands off after that, or just kinda walk me through what it typically looks like?Ivan Lee: Yeah. Great question. So there's a new term to describe what we do here in Silicon Valley. It's AI native services. And so to the client, it feels like white glove consulting.Right? They will describe the workflow that they want or the workflow that they currently operate. And, yeah, we'll do everything from recommend the model. We will talk through how complicated this is because sometimes it's as simple as deploy this, connect this to our internal Dropbox or OneDrive, and suddenly it works. Sometimes it's a more complicated agent.So there's a little bit of education in there on, like, hey. How much, you know, work is this going to take? How long is it going to take? How powerful does the model need to be? But once we have all of that, internally on the data source side, we have this proprietary model factory, essentially.We have this thing that automatically takes the kind of requirements and suggests which model is going to be best. We take the kind of workflow and the benchmarks and the historic data, and we throw it all in there, and it's... And it kinda fine tunes and does all the harness engineering and everything else you need to make that model actually work. Right? So to the client, it feels like consulting services, really speedy services.From our side, it's actually this kind of internal productized platform that builds all of this from a privacy perspective first.Jordan Metzner: That's awesome. Yeah. Awesome episode. Great interview. To your point of that time frame, it wasn't even...This... None of this was possible previously, and so that means, like, this is new... A new frontier and a new ability for, you know, ChatGPT is awesome, so everyone's throwing their data out. And I think Sam said at beginning of episode, now you can you can actually, like, you know, host your own models and and get high high level performance. And it seems like that's just the beginning of where, you know, large companies will take this.Rise Labs builds production grade AI for the world's most ambitious companies.Ivan Lee: Support engineers. Data scientists. AI architects. All working together. On your problem.Jordan Metzner: While others talk about AI. RiceLabs is already implementing it. Okay. Cool. Shall we jump into the news?Sam Nadler: Yeah. So, you know, I think we've talked about it a little bit, but, you know, Meta reporting over 10,000,000 Muse users. I know, Jordan, you've used Muse a little bit. Also, Anthropic, I hope you guys can see this, announced Duet, if I'm saying that right, another agent to take on Muse and Dots. We've been using Grokbot, recently.We have some, like, I think, pretty successful recent use cases in the last couple weeks, getting our teammates to leverage, or deploy some of these in their their routine workflows. I've been curious, have you used it? I'm assuming you have. Which tool is your favorite? What do you like about it?And Jordan, we'll we'll shift to you after that.Ivan Lee: Yeah. I love these. I didn't mention earlier when I was at Apple. I work on Siri, so I've always been a big fan of personal agents, and I think that this technology enables it more than ever. I think what's most interesting here is that enterprise agents have really charged ahead.Right? It started with coding agents. Every engineer was picking up coding agents at the beginning of the year. That was kind of the extreme product market fit that was first found here. Now we're seeing agents for customer support, service, marketing, sales, etcetera.And then Muse came in because even as all of this was happening in the first nine months of the year, there were no personal agents. We've been promised these travel agents. We've been promised these things that, like, book weddings for you and do all this party planning and everything else, and they weren't really capable. And I think Muse kinda, like, figured out a UI, has all this context direct from you... From your Instagram, from your Messenger conversations, and it was able to kind of take the first pass at it.And I think it's been fascinating to me, again, as a product manager to see the first wave of consumer agents come about. I am currently on the wait list for Anthropic. Anthropic is historically more enterprise coded, and so I'm curious what their take on it is going to be like. I didn't see a cute mascot associated with the duet launch. So I think this is this is very interesting.We benefited as a company a lot from just consumer adoption. Like, a lot of our clients may not have been using AI at work, but they were starting to use ChatGPT and Gemini at home, and they were wondering how they could bring it to work. I think this kind of thing is only beneficial for us because now they're going to start using agents at home and wondering how they can bring similar agents to work.Jordan Metzner: Yeah. That's interesting. I've tried a few of them already. Yeah. I've liked Rockbot the best.Just maybe this is moving really fast, and I feel like what I like today might not like tomorrow, and it's gonna move really, really quickly, and then integrate it to data. I think we haven't seen Google's solution here, and I think we're highly underestimating Google because it's already connected to all these data sources like your email in our case, you know. So we're on Google stack. Right? So for us, that's very helpful.I think, obviously, we haven't seen a Microsoft solution, which obviously we'll probably see, like, an updated copilot here that can do this kind of stuff. And to your point, like, then who will be the winner we've talked about at the beginning and take all the tokens away. Right? And so it's kinda looks like this, you know, arms race between different players. It is interesting how Facebook has, like, kind of gotten...Been able to get ahead in this space, and even Elon too. I mean, it looks like, you know, Dots from OpenAI is not, like, as popular as these other as these other platforms are. And I think in the enterprise, yeah, it's doing so much work for me. I think it it... It's gonna start to redefine work, to your point at the beginning, you know, and it's a new definition.But, yeah, super fun and exciting.Ivan Lee: Yeah. And kinda final hot take there, I think agents are just universally a better fit for the enterprise than they are for the consumer world at this point. Like, so much of white collar work is just repetitive tasks with data inputs and data outputs. Right? And it's...Like, we work with clients that have existing operational playbooks. They didn't build it for agents. They built it to train the next human employee. And so that... Those playbooks are already in place.Whereas for consumer, you know, you and I will have different preferences on which airport, what seat we want on the airplane, and all these other things. So I think there... We're gonna see a lot more market fit for enterprise agents for the next few years before consumer really begins to take off.Sam Nadler: Yeah. Have you discovered in just your own use of agents any, like, favorite use cases?Ivan Lee: The thing that I most recently built on Claude even without Duet, I have this, you know, RuP bracelet, and I created an MCP server for it, and I connected Claude to it so it knows how well I slept yesterday, and it tracks my activity every day and gives me a dynamic training plan for the day. I... You need to go to the gym. You need to run. You need to take a rest day.You didn't get enough sleep, so you shouldn't be doing these things. So I really enjoyed that because it it literally adapts daily to, you know, what happened. And, like, if I was too busy or I was at a conference, it understands that.Jordan Metzner: That's pretty awesome. Now that, like, everybody can hack into all these devices and make their own version of it, it seems like maybe that's also a new opportunity for, like, new open source hardware where people will be able to, like, kinda pick out your own version of how you wanna connect it. I mean, like, an alternative to Eight Sleep that's open source with... You kinda pick out your app, your favorite app, etcetera, and kinda connect it all. But, yeah, it's super cool.And obviously, just have your agent do it for you, you know, is even better.Sam Nadler: Ivan, where can people find you?Ivan Lee: Best place to connect would be on LinkedIn. I believe my tag is just I y lee, so shoot me a message there and happy to chat anytime. And, of course, our website, datasource.ai.Sam Nadler: Amazing. Thank you, everyone.Jordan Metzner: Yeah. Thanks, Ivan. Thanks for joining us.Ivan Lee: Absolutely.Sam Nadler: Alright. See you, everyone.