Speaker 2
We started with a meeting bot. Everyone was doing zoom meetings. It was a bot that would join your meeting, capture, record it, transcribe it and make that shareable. When we started, that was novel. And then every week a new startup came and did it. And now there's dozens of companies doing that.
Speaker 1
You just described my whole like news writing
Speaker 2
life of 2020 is that that startup over and over again. Yeah, exactly. So that became very crowded. And what we realized from our users is, and this is where any of our best ideas came from, the seed in a problem that our users described, they love the idea of comprehensiveness. They really wanted to capture more of their lives, not just these Zoom meetings they're having. And that led us to ideate around, how do you capture more than just meetings? And around that time, the enabling technology of Apple Silicon came around. Apple M1 came, a chip that allowed you to do a lot more things locally to offload a lot of the things that would otherwise have to be done in the cloud. And that's when we moved from scribe with MeetingBot to Rewind, the Mac app. That was again, long before chat GPT, GPT 3 .0, 3 .5, which is the model that powered it. And I actually say we got more lucky than good there where the data that we were actually already capturing with Rewind just really lended itself well to RAG, Retrieval Augmented Generation. Like the ability to use that data, you know, to ask a model like GPT 3 .5 or 4 .0 or 4 .0 now, if you asked it to draft me an email to Sam Altman, if you just ask it today without any context, it does an okay job and knows who Sam is, it certainly doesn't know who I am and doesn't know our relationship. But if you augment it with the context of our relationship, the fact that he invested four years ago and that we've been working tirelessly in this migration revolution to Limitless, all of that context, you provide a large language model. It drafts a perfect email. In fact, an email I sent to Sam that an AI could have drafted worked great and we met and we connected. So these things, otherwise you have a blank piece of paper can easily be made easier by having AI augmented with your context of your past.
Speaker 1
What would you have done had this not happened? Like in this sort of parallel universe where like chat GPT doesn't happen, and we don't get this kind of incredible run of stuff we've been on the last few years. It doesn't sound like you were counting on that. You weren't saying like, we're gonna bet on this foundational technology to be the thing. What is parallel universe Dan building to make that stuff work?
Speaker 2
Yeah, I would have been doing more of the same. I mean, I would have, it was already valuable enough to do search over the things you've seen, said or heard in the past. The Large Links model just makes that search more useful and more actionable and more, it sort of takes that, task that you would have had to do. Like, let me search through all my emails with Sam. Now we figure, OK, when did we meet? Who introduced us? Instead of doing that manual task, now you do it in an automatic way. And so it just made me an analogy to Roz. Before we were on this evolution from horse and buggies to self -driving cars, our evolution before chat GPT was maybe we added a car with manual transmission way better than horse and buggy. But what chat GPT and the underlying API that we use enabled us to do is go to automatic transmission. So we're not quite at self driving cars, where the autonomous AI simulations of your mind doing things for you. Yeah, but we are able to actually save you time and give you a lot of value through this sort of evolution. So that's what we would have been doing. And by the way, now, our focus is basically just banking on the models getting better, like everything we do is under the premise of just that the models are going to get better. So like just just collect all the data in the best possible way to ride that wave. So our mindset has definitely shifted. We're not sort of doggedly just pursuing what we've done before. Now we realize, okay, there's this amazing wave of change or these models are just going to get better and better and extra believe. Why not just ride that wave and build a product that just gets better on its own, gets cheaper, gets better on its own as the model get better? The
Speaker 1
car analogy there is a little messy, but it's an interesting way to think about where we are with AI. We spent two decades or so with products like Google and Facebook, which built very smart and sophisticated systems for looking at a huge amount of stuff and ranking it a million different ways. Those systems are by and large very good, but the promise of AI is that it can take all that stuff and actually come to understand it. Not just find you the thing you're looking for by putting it at the top of the list, but by finding the perfect thing that you're not looking for or by using everything that you already know to help you do the next thing. We don't need AI to do Google searches. We really don't. Even Google is currently showing us how much we don't need AI as it tries to put more AI into Google search. But if all the AI boosters are right, we can use AI to not just find things, but build new things on top of them. In a personal context, that brings up the central with all of these AI systems. In order for an app like Rewind to know everything about you, it needs to know everything about you. Do you want your computer to store and save everything you do, everything you click on, all the words you type, all the TikToks you scroll through, all the pictures you look at, every single thing that you do wired at computer? Forget about the data security risks behind that for a second. Just like as a human, how does it feel to know that all of that stuff is being recorded and stored in perpetuity? And how useful does your computer need to become? What does it need to do with that data in order for it to be worth the trade? Dan has been thinking about this for a long time, and he calls it the personalized AI privacy paradox.
Speaker 2
And it goes something like this. In order to build a more useful personalized AI, you want more context recorded. But that raises more privacy concerns, which requires more need for data protection, which makes it harder to build a more personalized AI. So the desire of a personalized AI, you know, part of it is this, okay, we want to collect more things because that context is going to be useful. But in Exribil, you go down this path, and at some point, that's going to make the original goal harder to do. And you can see this paradox playing out in the world in two very different ways. With Microsoft, they were pretty cavalier on privacy when they launched Windows Recall. Lots of things to say on that, but let's say they launched a product called Windows Recall, which looked very familiar. But they took a very, very cavalier approach to privacy and that really hurt them. On the flip side, Apple Intelligence seems to respect privacy but it actually is limiting what Apple can actually do in terms of usefulness. So they're kind of on this both sides of this paradox. We've actually made many mistakes in this space. I think the most recent evolution of our thinking has really set us up well for the future. But it's not obvious. It's not straightforward. It's one of those things you have to kind of you're like tight roping between landmines, you know, you're trying to find the right path that respects privacy, and doesn't make it a choice between privacy and convenience. And the same time makes the product useful enough, because you're able to use the data and way to offer personalized AI.
Speaker 1
I agree with that. But then you have to put that in front of users, right? And I think even there's something about the idea of just like, here is an app that shows a timeline of every web page I've ever been to feels instinctively weird to some people, right? And I think it's been very funny watching a lot of this because like, yeah, of course, your browser knows all the web pages you've been to, like, that's your width, your web, like, yes, that's how it works. But I think a lot of what has been happening in these recent months is people are slowly starting to understand kind of how much awareness their technology has of what they're doing with it in a way that everybody probably should have had before, but didn't. And I feel like what you did, especially with Rewind in the early days is like really speed that process up. Right? You're like, this app knows everything and it's actually its job to know everything. And I wonder if to some extent it helps because it's like a thing you have to download. So by definition, you're going to get people who are more comfortable with it rather than like building it into the operating system. But you have to do this thing, I would think right away where you're like, okay, this is asking a lot. It's going to know a lot about you. I have to sort of immediately telegraph to you why it's worth it. And I feel like that's a that's a pretty big hurdle to clear right away.
Speaker 2
Yeah, I think we certainly paved the way there. But I wouldn't say we didn't make any mistakes. I think showing a timeline of everything you've seen is interesting. It's a cool trick. It helps create this magical moment of like, wow, I didn't realize I could. People love that. And that's partly why we named the first product in this space Rewind. But if I really had to be honest, people don't care about technology. They just care about their problems being solved. The technology is the means to the end. And this was too much of a cool, let's see what the technology can do part of the product, unless here's the problem we're trying to solve for you. So I do think the right user experience around personalized AI is to be very opinionated on the use cases. What are the problems in your life every day that we're trying to solve for you? How do we give you time back? How do we make it so that you're in time for dinner with your young kids at night? Those are the kinds of problems people care about. The technology and how sophisticated it is, they could care less. Like that data, the architecture, that is important to us, because it's how we build a product. But to a user, that's just implementation detail. So I do think Limitless does a much better job of this. This is why we've kind of evolved, and even partly reason than rename the product, because the core experience of rewinding time isn't the core thing people want. The job to be done is give me back more time, make me more productive, take things off my plate. What are things I do every day that frankly, I just don't need to do and a machine can do a better job of be more reliable and just give me back more time so I can do the things that I'm uniquely well suited to do. So those are the kinds of things and the kinds of experiences that I think will ultimately win in this world of personalized AI.
Speaker 1
What are some of those things that you identified early on as solutions those problems that you're describing?
Speaker 2
Yeah, I mean, big one is a blank piece of paper. Very often as knowledge workers, you start with a blank piece of paper. Maybe it's an article you're writing or an email you're sending, or even a simple text message you're sending to somebody. Writer's block is a big version of this problem, but it's this idea that starting from zero is much harder. And part of what a machine can do uniquely well is capture the context you might want when you're starting from zero. A good example is drafting emails. I gave an example of drafting emails, say a moment. Starting from scratch is a much... Why spend all the time and energy when a machine can surface to you perfectly the thing you might need and want. A way you can think about it is auto complete your life. Why do so many times you have to start with an empty line where a machine can provide an option? It doesn't do it autonomously. It just carides a draft, something you can edit and tweak and change and delete. That's a perfect win between, like I said before, horse and buggy and self -driving cars. There's going to be a day when AI can do things autonomously. We trust it. It can book you a trip for you and your wife to Italy in three months, and you know it'll put the right seat and everything. Right now, the use cases that I think the AI is well -suited for are these semi -autonomous use cases, things like drafting notes. That's just one. There's many, many others. Is that memory though? It is in the sense that the things that the context that's useful for those moments are memories. There are things from your past. There are details around your last conversation. If you think about this idea that we forget 90% of what happens after a week, you have a team, weekly team meeting. Many folks probably listen to this, have a weekly team meeting with their team at work. Maybe it's an hour long meeting. At best, the people remember six minutes of that last meeting. So simple things like following up with a context of what was said in that decision in the last meeting, the set of decisions, those kinds of use cases, meeting summaries, preparing you for meetings, live notes during meetings, all of these things are incredibly well, great sort of use cases that personalized AI can help you
Speaker 1
solve. It seems to me there's almost two different things going on there. Because on the one hand, there's the thing that's like, okay, I'm going to make it easier for you to remember at least all the important bits of your one hour long meeting last week, right? So that when you go to your next meeting, you can very quickly call up all the things you talked about last week. That is, I feel like, I think a lot of AI companies are pursuing, right? This like, take a bunch of notes, or we'll take the notes for you, and then we're gonna give you sort of quick recall of those notes. But I feel like you're also describing kind of a full step beyond that, which is just instead of helping you sort of actively remember something, we're going to use all things that you forgot to help you do new things.