[00:00:00] Luke: From privacy concerns to limitless potential, AI is rapidly impacting our evolving society. In this new season of the Brave Technologist podcast, we’re demystifying artificial intelligence, challenging the status quo, and empowering everyday people to embrace the digital revolution. I’m your host, Luke Malks, VP of Business Operations at Brave Software, makers of the privacy respecting Brave browser.

[00:00:24] and search engine, now powering AI with the Brave Search API. You’re listening to a new episode of The Brave Technologist, and this one features Brian Bondi, who co founded Brave Software with Brennan Eich in 2015, and now serves as the Chief Technology Officer and member of the board at Brave. We’re really excited to share this episode with you as we covered a lot of ground.

[00:00:43] Today we discuss some of the backstory behind Brave, the Brave Leo AI project, the roadmap, some of the differences between having AI outside of the browser versus having AI directly in the browser, how Brave’s already been using AI in different ways aside from the Leo [00:01:00] project. ways that brave sees that a i could be used in the browser directly in the future and now for this week’s episode of the brave technologist how you doing

[00:01:12] Brian: brian awesome how are you live.

[00:01:14] Luke: Doing well, doing well. Thanks for taking some time out to join the podcast today. No problem. It’d be super cool for people to kind of get a sense of your co founder at Brave. How’d you start getting involved? Give us a little bit of background on founding Brave

[00:01:26] Brian: Yeah, sure. So I was at Khan Academy at the time, and I randomly got a tweet from Brendan asking me to DM him, and I… Print to my wife and I’m like, Holy crap. The creator of JavaScript has messaged me and asked him to DM him. So, so I did. And we got to talking and it turned out that he wanted to start a browser company.

[00:01:44] I had built up a reputation at Mozilla for shipping a lot of different things and working on performance work and. In particular, knowing windows platform through and through and creating a browser from scratch that was specifically designed for the new immersive mode and windows. So he reached out to me and kind of wanted to talk [00:02:00] about starting things.

[00:02:00] And as we got to more and more talking and more and more planning over several months, he eventually asked me if I could fly out and meet him. And he had wanted to to start co founding the company. So we went and we got seed investment. The rest is history. Eight or nine years later, we have a couple hundred people working on creating a browser and a search engine, a cryptocurrency wallet and AI and a bunch of other stuff.

[00:02:23] So much

[00:02:24] Luke: stuff, you’re such a big player in, in a lot of it, right? Like, I mean, this whole element of like, okay, you can have a wallet or you can have AI or you can have these other things, but like doing that all from the browser, right? Like it’s really different and it takes like, you know, somebody like yourself to be hands on with, with getting that going.

[00:02:40] I saw like a bit about this Leo roadmap that we’re working on Brave to release. Maybe you can give people a little bit of background on what Leo is and then we can dive into the roadmap a little bit.

[00:02:51] Brian: Yeah, sure. Thanks. So the main thing to understand is that brave is a browser. So the same reason why we have a wallet and brave is because there’s dApps that you can visit.

[00:02:59] It just kind of makes [00:03:00] sense, right? So what we didn’t want to do was just create another chat GPT and add into the browser. So you have a browser and it also has a chat feature in it, which and they’re separated. That didn’t really make a lot of sense to us. So what we want to do is leverage that it.

[00:03:13] Users are visiting content. They want to understand that content as well as they can. So it makes sense to give the AI that you have open access to the page that you have open and to be able to converse about that page. And you can ask tangential questions about the content that’s on the page. You don’t need to, like, for example, if you’re reading an article about running, you don’t need to do a Google search and get out of the article and look for another similar topic to the article that you’re reading.

[00:03:39] You can just ask. Maybe when the author says this about VO2 Max, what do they mean? And you can really deep dive on the content, even if the author didn’t explicitly address a certain, certain question. So we really wanted to build that unified experience where the AI agent is not just a separate feature, but it’s there along with you on the ride, understanding the content and able [00:04:00] to Give you recommendations and help you explore the content in a new way.

[00:04:03] Luke: What’s different about how brave it’s approaching this compared to how other tools might approach it? Is there anything that really sets

[00:04:10] Brian: us apart? I think the biggest thing is just that we’re a browser, whereas like open AI is not a browser. So when you’re on a webpage and you have that, that article in front of you, being able to have context and sort of information is really relevant.

[00:04:23] We’re also a search engine, right? So. We have access to live data, so it’s not just outdated information when you’re communicating with the model. you can get up to the date information. We also have brave news that we can tie into this as well. So you can get really up to the minute kind of information when you’re talking with that.

[00:04:41] Brave is also not training its own model from scratch. So we have access to, I guess, connect the world’s leading LM models, whatever they are with users, allow these are to select between them as well. And also we’re not in the business of obviously collecting user data or anything like that. So really.

[00:04:58] Local models is something that we want to do [00:05:00] as well, not just a hosted solution, but right now what we have is a reverse proxy with llama two is the default and then there’s a premium version that gives you a bigger version of llama to the 70 billion parameter one and also a cloud anthropic access as well, just so you can get better answers when you’re asking questions over the llama to 13 B, which is the one that’s included for free.

[00:05:21] Luke: Leo’s been out in testing for, in Nightly for a little while, right? Yeah. Have there been any um, interesting use cases or developments that have just come from testers that you guys didn’t see in advance? Or, or are you guys learning anything interesting from having a lot of people trying it out that might be unexpected?

[00:05:38] Brian: Yeah, I mean, when we started building this, we didn’t know exactly what, what we wanted to build. Like we knew we wanted to have like AI in there and helping the user, but we didn’t know exactly how. And so like we started with just like a summarizer and I mean, it was very useful to be able to summarize long articles, but it just didn’t unlock the full power and potential of what we really want to get at.

[00:05:58] And I was reading a book named neural [00:06:00] networks from scratch. And what the article did in that book was they created a word document with the whole contents of the book. And if while you’re reading the book, if you had any questions, you can go into that word document, you could highlight the section that you had a question about, you could ask the question, and then the author would go in within 24 hours and respond to you and give you an answer of what you didn’t quite understand in the book or what you had a question about.

[00:06:19] So the idea was really, how can we take that and apply it to the entire internet? Why does it have to be about just this one book with this great feature? Like, why don’t we apply that feature to the entire internet? So that if a user has a question about any article, Allow them to basically almost communicate with the author.

[00:06:35] And in many ways it’s better than communicating with the author because the author has certain biases. So you can, you can do bias checking and things like that as well with us. That’s awesome.

[00:06:43] Luke: I remember seeing like one of those big things that jumped out at me too, was I think in one of the demos you were showing us, you pulled up a page with like a YouTube video and how, like, you know, Leo could basically like get onto the key frame, you know, noting and, and things that you’re starting to see people do in comments, right?

[00:06:57] Like for, for YouTube videos, like, but, but Leo [00:07:00] kind of do that. Like, it seems like there’s a whole bunch of use cases that we haven’t even discovered yet. Right. And then the users will probably help

[00:07:06] Brian: us. I think a big part of what we want Brave to do is give these models. The deepest understanding of the content that you’re viewing.

[00:07:15] So yes, you can do it with a blog post that you’re reading and it’s long and you want to summarize it and ask questions about it. But like you said, you can also go to a YouTube video and it knows how to, for example, retrieve a transcript and knows where the timestamps are on that transcript. And it knows which parts of the video talks about different things.

[00:07:31] But I see a lot of the work that we will be doing in the future is to keep giving it deeper and deeper understandings of the pages. So hooking into like accessibility information and things like that, just to really. Make the model aware of the structure of this page of the images on that page of the documents and links on that page and just taking it to a whole nother deeper level because the more that you can help the model understand the page, the more the model can help answer the questions that you may have for it.

[00:07:58] Awesome.

[00:07:58] Luke: Awesome. Where do you think [00:08:00] AI is going to have been successful so far? And where do you kind of see

[00:08:03] Brian: the gaps? Yeah, I mean, generating content has just been amazing to see over the last year or so, asking it to create or even providing an image and say, recreate this, this game on this web page and it spits out the code for it.

[00:08:16] And then you’re like, oh, no, wait a minute. I want you to create it in this language instead, and then it translates the whole thing into a different language. It’s really, really amazing to see a little bit scary as well. So I think that content generation is a great place. I think where it falls short is inaccuracy.

[00:08:30] Sometimes, sometimes it can hallucinate. Sometimes we hallucinate as humans too, right? It’s true. We don’t always get accurate information either. Right. So we do the best we can, but I do see that that problem will be less and less as time goes on because the model can check itself with other models. It can check itself against recent information on the web.

[00:08:50] It can check to see if it’s right. It can even reward itself. By doing like a brave search, for example, and seeing if content is valid or not, and kind of create that reinforcement learning loop [00:09:00] by itself. It’s really exciting to see where it goes.

[00:09:02] Luke: What stuff is most concerning you right now? Folks know about this like Doomerism stuff in the abstract or whatever, but I mean like from a practical

[00:09:09] Brian: level.

[00:09:10] Most concerning, I’d say, just the pace of the development. So I think that as long as we have time to adjust to the realities of the progression of the technology, it’ll be fine. But anything that’s too sudden would be really bad. I guess for people that don’t have time to train, to get new jobs in time to adjust their lifestyle.

[00:09:29] I really think overall that the potential that this technology brings is just massive and to improve the average life of people. And I mean, even for things like schooling. Everyone, when it first was worried, like, Oh, people are just going to use this to cheat and create essays and submit them. But if you just treat it as your own personal tutor, and that’s what Khan Academy is doing, for example, they’re just embracing the AI and saying, yeah, let it be your tutor because people succeed a lot better when they have tutors.

[00:09:54] And if you just allow it to do that, and just when you set up the pre prompt for it, you just say, [00:10:00] I don’t want you to give me an answer. I want you to help me understand the answer. And this is going to be transformative for people’s learning and their education.

[00:10:06] Luke: People being able to ask questions to the model.

[00:10:08] They might not feel comfortable asking in a broader classroom. For sure. Yeah. That’s one of the cool things about doing this from a browser too, is just from an accessibility standpoint too. Anybody with a browser can access this thing and start to use it, which is pretty exciting. I know you’re heavily on the web three side of things too, like from the browser and AI and web three are kind of separate buckets, but you see them kind of coming together at

[00:10:31] Brian: some point.

[00:10:32] Yeah, so one of the things with Leo is that like right now, we’re not hooking it up to third party services, but we do plan in the very near term to be hooking up to brave search. We also plan to hook it up to different web three services. I can see, for example, like, there’s always been this problem where, like, a wallet interface is kind of confusing.

[00:10:50] And is it really? Up to par with what my grandma would use like, for example, she just said, I want to transfer money from here to here and that AI [00:11:00] agent could understand any way that she asked it and, and just kind of create that transaction for her, then that becomes a lot more usable for her. Right.

[00:11:09] Awesome.

[00:11:09] Luke: Yeah. And I see too there’s a bit in the, the Leo remap about exploring around ways that you could even incorporate something like rewards right into Leo or in a monitor that different types of monetization models too.

[00:11:21] Brian: Right. Yeah. So like we have this private ad system that you can opt into and then you can get rewarded if you do opt into it.

[00:11:27] But what’s to say that you can’t. Tell the agent, here’s the ad catalog and then you could ask questions about it and just get honest feedback from, from the model about what’s in the catalog and what might interest you in the catalog. And is there better alternatives out there or is this the best alternative there is?

[00:11:43] And just being able to use natural language processing to help the user do that would be pretty amazing. I think.

[00:11:50] Luke: Awesome. Is most of what you guys are working on now kind of an English language model or how much does that localization impact something like a release plan for, for

[00:11:58] Brian: something like this?

[00:11:59] By default, we [00:12:00] ship with Lama 2. 13b. That’s primarily trained on English, but it does support many different languages. It’s just that most of the training data, maybe like 80 or 90 percent of it is English. I believe, like you can converse with it in any language you want, but if you want the best results, you’ll, you’ll converse in English, but the premium version, like anthropic, for example, is better at that, for example.

[00:12:19] And we’ll be hooking up other third party providers in the future. I’m sure as well. So. Those will be better at that as well.

[00:12:25] Luke: It’s one of those things, it seems like it’ll just keep

[00:12:26] Brian: improving with time. Yeah. And we’re about to add like even like Code Llama, for example, to Leo as well. So it doesn’t just end up like human spoken languages.

[00:12:34] Like you can have help with developer related programming languages. Oh, that’s super interesting. Yeah, and right now the primary interface for Leo is just on your sidebar. You just open it next to the page that you’re viewing and you can talk with it. We’re looking at other things like when you’re in a text box on any web page, you can reformat your context or reword it in different ways.

[00:12:53] And also for Code Llama, being able to open up your developer tools and have it be your assistant within developer tools, there’s just a massive amount of [00:13:00] different potential for that mind blowing stuff.

[00:13:02] Luke: It’s awesome. I mean, this is being the browser, right? Like, this is what’s so cool about

[00:13:06] Brian: it. Yeah. Just imagine being like on a page and being like, I want you to restyle this page because I want it in dark mode and the page doesn’t support dark mode or any, and you do that, or you just, you can make any other change in the fly, just with spoken language.

[00:13:18] Yeah,

[00:13:19] Luke: has there been any thought because you see there’s like plugins for AI that have come out and browsers kind of have these old extension libraries to around how these two things could possibly converge like over time, like you could have Leo, maybe work with extensions or something like I’m kind of getting way out there.

[00:13:34] But I don’t know how much this stuff has been a part of the thought processor or something. You guys are just kicking around. Yeah, I think

[00:13:39] Brian: the way to do it is to basically tell the model. Here’s basically some functions that you can call. Here’s how you would call them. And then Yeah. You allow it to generate a response while calling those functions and you specify what the functions are, what the parameters are, and then it’s kind of figure out when it should call it and when it’s appropriate to call it.

[00:13:57] And then it’s just kind of like a user setting about, I want to approve the [00:14:00] things before it actually happens on my computer would probably be the default, but you can maybe set it to like, I want to let it do whatever it wants to do and kind of auto gen mode type thing. Awesome.

[00:14:10] Luke: Yeah. Well, what’s been the most impactful personal or professional example of using AI that you’ve from your own personal life or work?

[00:14:17] Like, are you top AI

[00:14:18] Brian: hack? I think it’s really cool that you can, you can tell the model, the rules that you want it to work within. Just as an example, you can say, like, I want you to be a text based game and With dragons and whatever else, and it’s just completely role plays. And it’s like, what do you want to do next?

[00:14:34] And you can say like, I want to turn left. And it’s like, you turn left and you see this and it generates it on the fly. So you can just have this pre conversation with it before you ask her actual question and kind of model exactly what you’re looking for. And it kind of role plays in that way. And it’s kind of perfect for what you need.

[00:14:50] Another example of that is just like studying. You could just say, like, I want you to create cue cards to help me learn this, this content, and then it’ll generate the cue cards of [00:15:00] the content of the article that you based in or whatever. Right. Yeah. If you’re using Leo, then the article that’s open even more simply.

[00:15:06] Yeah,

[00:15:06] Luke: that’s awesome. Yeah, my own personal one’s been around making a children’s stories for nighttime. It’s kind of like Mad Libs, like, Oh, tell me a funny story. Make it short. It’s actually pretty

[00:15:16] Brian: well. Yeah. And once I further is making that interactive, right?

[00:15:19] Luke: Exactly. And now we’ll say, okay, make a sequel of this.

[00:15:23] Like what’s the next chapter kind of thing? It’s pretty rad. Where do you think all this will take

[00:15:27] Brian: us? I’m somewhat worried about things like neural link where of course that has a lot of good use cases for people with different disabilities and things like that, but at the same time, it’s kind of scary.

[00:15:39] I guess the divide that that can create between people that have that neural neural link set up into their brain and those people that don’t, it can create a huge competitive advantage for the haves versus the have nots. I think it takes us to a more efficient. Place is really hard to predict that I don’t know what we’ll see in a few years.

[00:15:57] Luke: It seems like a real like plowshare to [00:16:00] tractor moment in everything we’re doing and that we’re just kind of at the beginning of. Yeah. What’s kind of the number one myth about AI that you wish people would stop believing?

[00:16:08] Brian: Just that it doesn’t have consciousness. It doesn’t have feelings. So some people might answer an AI agent in a certain, in a rude way.

[00:16:18] But like the AI agent doesn’t feel bad about that. It’s not going to hunt you down in the future. Maybe, maybe it will if it looks at its logs, but not right now. Anyway,

[00:16:27] Luke: it’s not hell yet. Awesome. You know, one of the things that we’re trying to do here, I mean, you know, like a lot of our users are developers or maybe people into tech, but maybe they haven’t messed around with AI stuff yet.

[00:16:41] What are some resources you’d recommend or ones that you’ve used and found value from for people that are just kind of getting their feet wet

[00:16:47] Brian: into AI? Yeah, so I would start probably learning like ML type stuff competitions, like on Kaggle, that’s a great site where it has just different ML challenges and everyone competes and tries to have the best solution [00:17:00] different books from O’Reilly.

[00:17:01] There’s a lot of good books, like natural language processing with transformers with PyTorch, two books there hands on machine learning with scikit learn, Kerris and TensorFlow is a good book. I mentioned as well neural networks from scratch from Harrison Kinsley. Really great book to help you understand neural networks at a really deep level.

[00:17:19] And it does take some, some mathematical knowledge, but he kind of like it says from scratch, like it takes you from scratch and doing it. I mean, he also has a YouTube series dating back a decade or even longer, maybe with just. Learning ML concepts right from scratch as well. So you could probably listen to dozens and dozens of hours of that to help you come up to speed on different ML learnings.

[00:17:41] Luke: You mentioned a little bit earlier, like about local models and so far, so much of this stuff’s been based in the cloud. How far of a leap? Do you think it is at this stage? Like, do you think we’re just kind of starting to get involved with local model? I mean, a browser obviously is like one of the best local model to kind of work from, right?

[00:17:57] Because the whole browsing corpus and all sorts [00:18:00] of other info there, but is this something you see becoming a reality within a year or many years?

[00:18:05] Brian: Yeah, it’s already there, right? Like you can already run a lot of these models on, on a typical desktop computer. I would say maybe mobile is a little more challenging, so that’s probably still.

[00:18:16] A little bit ahead to be able to work with smaller and smaller models that perform better and better. Yeah, we’re already there. And even within a couple quarters, you’d probably see a feature in Brave that allows you to interact with local models of different types. And the real advantage there is, is data privacy and.

[00:18:32] Even though we have the way that it’s set up right now with our current Leo is there’s a reverse proxy. So even if you’re hooked up to a premium account and sending things to cloud, they still don’t know like which computer is coming from you do have to shield your requests. I guess if you’re talking to like anthropic cloud, for example, and you don’t want to be ever submitting like PII data to it, but there are a number of measures that where that’s stripped out as well.

[00:18:54] Like Amazon has services for stripping out PII from queries and things like that as well. We do our best to strip [00:19:00] out any kind of PII that you might accidentally submit as well. But yeah, once you have widespread local models, you can do a lot of cool things. Like you can feed in the history of the browser and you can feed in your bookmarks and all the things that you wouldn’t want to send up to a server.

[00:19:13] You can have your, your local basically profile tied in with this model. So it understands. The things that you’ve read, the places that you visited, the forums that you’re filling out, et cetera.

[00:19:24] Luke: Do you see kind of a future where your Leo profile can kind of sync across your devices too? Kind of like what we’ve done with Sync with history

[00:19:31] Brian: and things like that?

[00:19:32] Yeah, but I would say not in a way that like Google Syncs where all your data is going to Google servers and then they process it in my name. In a way that Brave Sync works where you’re the only device that holds the encryption key for that data and that encryption key you can transfer to your other computers.

[00:19:49] And then you can decrypt it there, but for sure, I don’t see a reason to not have your devices work together in a collaborative way for that type of thing. It’s

[00:19:58] Luke: awesome. It’s kind of like [00:20:00] really bringing convenience factor, still not giving everything away to like a, you know, cloud somewhere where who knows what can happen.

[00:20:07] Do you have like kind of a favorite movie or show or book featuring AI that you’d recommend listeners check out? I know we kind of got on technical side, but like, I don’t know, something fun. I’m

[00:20:15] Brian: always a fan of Matrix, but I know it’s like really old, Terminator, things like that. Terminator

[00:20:24] Luke: Truth and fiction are kind of brushing up together these days. No, that’s awesome. Is there anything that we didn’t cover that you kind of want to go into some detail around or anything that you want people to know about?

[00:20:35] Brian: Yeah, I would just say that like we focused a lot on Leo, which is the new, the new feature that’s coming out in Brave to converse with pages basically.

[00:20:42] But Brave’s using like AI in a lot of different places as well. Like Brave Search has, has had a summarizer in there for a while and it’s the beginning stages as well of being able to ask it questions and chat from there. Like you said, Brave Wallet will have, have features like that. Brave News, we’re using AI.

[00:20:56] Our ad matching has always used ML models in [00:21:00] it for several years now, maybe like. Seven or eight years now, where it’s matching a user’s interests locally with ads that are an ad catalog that’s delivered locally as well. So that no server ever needs to have it. So I really do see that this AI technology is going to be used across the whole web browser.

[00:21:16] It’s not going to be local to just one little sub. Side side panel, it’s going to be used for, for everything. Like if you want better names for your, for your bookmarks, for your, for your history, if you want to be able to search the text or ask questions about your history. And yeah, so I just see everything that you can do in a browser, it would just be better, better basically with us coming up.

[00:21:35] That’s a good point

[00:21:36] Luke: too. I think like even some of our ad blocking, right? Like we’ve, we’ve got in place has been using, you know, some AI based methods.

[00:21:42] Brian: Right. Yeah, absolutely. I definitely see filter lists being fine tuned validated. There’s a lot of hand curation for, for adblock lists now. It uses adblock filters, adblock plus filter syntax is the name of it.

[00:21:54] It uses these rules that are created, mostly created by community. But yeah, I see that being auto managed for [00:22:00] sure.

[00:22:00] Luke: Awesome. Where can people find you online if they want to talk to you more about anything or read about your running marathons? We didn’t even touch this. I think this might be one of the few interviews where we didn’t really talk about your marathon running.

[00:22:11] So anything around that.

[00:22:13] Brian: I have a website, brian bondi. com twitter. com slash Brian Bondi is probably the best spot as well. And it’s probably the social network where I’m most active anyways.

[00:22:21] Luke: Awesome, man. Well, Brian, thank you so much for taking time out of your day to answer some questions here and share more about what Brave’s up to with AI and Leo.

[00:22:30] And yeah, do everybody check out check out Brian’s site and follow him on Twitter or X or whatever we’re calling it these days and really appreciate it, man, have a good one, maybe we’ll have you come back to give us some updates as we roll more and more of this out. Sounds good. Thanks Luke.

[00:22:44] Brian: All right, man.

[00:22:44] Have a good one.

[00:22:46] Luke: Thanks for listening to the Brave Technologist podcast. To never miss an episode, make sure you hit follow in your podcast app. If you haven’t already made the switch to the Brave browser, you can download it for free today at brave. com and start using Brave Search, which [00:23:00] enables you to search the web privately.

[00:23:02] Brave also shields you from the ads, trackers, and other creepy stuff following you across the web.