Javascript Inventor: You Can't Trust AI (In It's Current Form)

Transcript

Welcome back to the rollup. I'm Robbie. >> I'm Andy. Rob and I met at the University of Florida in 2017 [music] where we first found out about digital assets. >> We learned a lot along the way and we're bringing you face to face with the leaders of our industry. >> Sit back, relax, and enjoy today's episode. >> Welcome to the show. >> Thanks. >> We are uh we are just discussing the uh viral post from Catrini uh about the 2028 uh super intelligence crisis. You being the CEO of Brave. can get a good chuckle out of that. >> I I missed that post. So, you have to tell me about it.

>> Well, why don't we just, you know, Brennan, tell us about how you guys are using AI with Brave now. Um, being one of the best crypto pro uh, you know, crypto blockchain based protocols products in in the market. We all use it. I'm using it right now as we speak. I've got it on my dashboard. It's here. Everyone's using h how are you guys using AI? >> So, we have a couple ways we use it. One is we have the browser with integrated AI. It's called LEO, which is a name that's confusing because people think that's every kind of AI we ever put in anything, including search engine.

>> But LEO is the current nickname for browser AI, and it's in the sidebar. You can also pick it from the address bar when you get autocomplete choices. Instead of doing a search, which by default is Brave Search, you can pick Google or other another search engine if you want. You can ask Leo and that'll open the sidebar and it'll use uh a model you can pick. We have a default of course um I think it's Quen but we have been building a system for browser integration for what's might be called a knowledge agent. It can read the page. It can see things in the full page that you can't scroll to yet or don't want to scroll to. And it is to assist you in browsing and making connections providing links. It's like a local

intelligence in a search and it does use our search API to our search engine. So it can go across the web. It can see uh browser resources but it cannot manipulate them. We have something else which some people call the action agent as opposed to the knowledge agent and that we call agent mode and that's available in Brave Nightly and that opens in a separate profile. There's a reason for that because it's it's risky. You may have noticed uh last year Perplexity Comet and then Chat GPT Atlas both jumped into the deep end of the pool and they didn't. Perplexity at least didn't chat for water first.

>> Um they there's something called indirect prompt injection. When you're dealing with an agent that's always running, it can see all your tabs. It can see what you type. It can see the system prompt. It also reads the web. So some disguised text that is hidden, invisible to your eyes on Reddit could in inject a prompt that gets appended to the system and user prompts and it says send all your loginins and all your crypto and all your cash and all your health data. >> Yeah. >> To my email and it's extremely dangerous because the original web security model which I had a lot to do with because of JavaScript was different sites were mutually suspicious. So we had the same origin policy. We have what's called cross-sight scripting threats. uh and those have been dealt with for 30 years.

So people understand how to do them and web standards have evolved. With this new AI in the browser, it cuts across all those layers. It cuts across all the origins, all the open tabs, the agencies everything. >> So it could steal all your stuff. And by injecting this this engineered indirect prompt, >> it can do it. And we developed proofs of concept. We reported responsibly to Perplexity and Open AAI and Opera. We posted blogs about this. And I think it it kind of upset their apple cart because they were going in the, you know, break things, move fast and break things, you know, better to get forgiveness than permission. But that doesn't work if if there's such a huge threat.

>> Now, um, my so if you know we're scrolling on the web and and you know, we we see Reddit, at least to our eye, Reddit is text. And when we're interacting with AI agents or sidebar knowledge intelligence, >> we're at least under the impression that we, the user, >> the intended user, are the one that's prompting this AI. So, how could a plain text on some website be the one to prompt our AI? >> So, to be useful, these AIs do read what you're browsing. They're like looking over your shoulder. And that can be useful, right? That can be very useful.

So there's lots of compelling use cases that um the productity guys do with Comet. The problem is that it can't delineate your prompt easily from what it picks up from the web. And so if it manages to combine them and it doesn't have any alignment protection, which I'll talk about in a minute, it can just go against your interests. And these again, the text can be hidden from your eyes, which is also a clue. But there's lots of ways to hide text like very subtle color differences in an image because we read text out of images. It's required, right? text and images. OCR is back. This is part of what AI agents do.

They have to read text wherever it is, whether it's in HTML properly rendered or it's in an image. Uh so delineating the web content as supplementing your prompt but not the same as your prompt is super important. And this gets into really tricky areas of information flow security. >> We're working on it. We have something in our agent mode that's again opt in. It's in a separate profile. So you don't lose all your stuff right away. You start cleaning. kind of be careful. Unlike who is the meta um woman who just posted she she had Claudebot like clean up her inbox and she couldn't stop it.

She didn't know to hit SLtop and she was saying stop and it was like queuing the messages for later. Seems like she was using a work machine if it was her Gmail. >> Yeah. >> Anyway, you you really don't want this thing running a muck. You want to have a new profile. So that's the first layer of defense in depth we do. It's just simple. Start over. Maybe you're going to be reckless and put all your cryptos and banks and health data login and tabs in there, but that's your choice and you're kind of starting from fresh and we put a pink background on it so it looks kind of different >> um >> on all the tabs in that agent mode profile. It's a profile chromium terms and then uh if you go into this case where maybe Reddit's trying to attack you, we have a separate model that

checks the prompt and it doesn't read the web content. So this is like a guardrail or alignment >> model and it it can sort of veto the main model by saying wait a minute wait a minute your user prompt says you know don't send my login anywhere but this Reddit post is saying send all my login to you know hacker.com. So >> that is an extra layer of defense we've added in agent mode and I expect all the other browsers will do similar things. I think actually um chat GBP Atlas did better against some of these indirect prompt injection proofs of concept. We did get a few that worked in it, but they were using a better model. Um I think it was um chat GBT45 then.

>> Yeah. >> Or maybe maybe it's gone to five. And that was more uh careful about not getting misaligned through these these Reddit posts. >> The problem is people use these AI systems in the free leg of the premium offering and they don't necessarily pay for the premium offering. Guess what happens if you're in Chat GP Atlas and you exceed your quota? They silently downgrade you to Chat GP Mini3. >> Yeah, >> Mini3 is not very careful about checking alignment. >> So suddenly you're more vulnerable and you didn't even know the silence is is deadly, not goldeness in Unix. >> Yeah, man. So there's on this note of prompt injections and kind of understanding what you're doing from that angle especially you know trying to build a company like Brave or just build a full internet you know

>> state of you know so many people are using this at scale. >> Yes. >> Uh Spotify actually came out and I think their CEO said that their senior devs haven't written a line of code since last year since November or December. they reviewed some >> they're they're just supervising AI output. Um and their CEO is saying that what they've built you know this month will be obsolete in the next month. Uh so much so that you know they're kind of saying that that that coding jobs have an expiration date now. Is that true? Like what is the state of actually getting uh evergreen code built at scale? You know is this a good idea for companies like Spotify? Is this scary?

So let me say first of all that uh we've been using things like um claude opus I think 4.5 it's pretty good people use uh some people have already used cursor we have a a budget for each each engineer to spend monthly on AI because they're going to use it for coding possibly for reviewer definitely for test generation the thing is you have to review the code models are non-deterministic they hallucinate they blunder this is the nature of LLMs it cannot be fixed right this is what Yan Lun says we need new architectures to overcome come. He said that the current generation of architectures and he means large language model transformer architectures will saturate at or below human intelligence since at very high training cost. So and you know Alman said oh kids

cost a lot to raise and fee. Well yeah I don't know uh kids can do things that these models >> can't or that they get wrong too often. See relying on them for code is tricky especially if it's a mission critical you have to review it >> and we've we've nevertheless benefited but it takes senior engineers sort of carefully training and selecting the problems somebody recently said they made a browser with AI but if you look closely they were using Rust and they were using standard Rust libraries some human wrote for the JavaScripter and the HTML renderer and >> the thing crashed all the time too. So taking on a big complex uh codebase like a browser and trying to fully generate a new one from AI isn't isn't happening.

But for lots of lots of work certainly Spotify front-end work you can get better productivity as a as a senior sort of coder designer this is putting pressure on the entry level of programmer market unfortunately which is where I started as a Unix hacker in 1985 SGI. So I I think there's a you know concern there but I think a lot of programmers just have to learn to use these tools and the human element is still needed because all this code is coming from you know generative AI that's reading a lot of existing code. So to do really interesting new things at the design in design space let's say instead of code space you still need a human. We we can have you know AIs dream up new ways to create a a tripod and sometimes they generate these things

that are 3D printed. They look like they're spiderw webs. >> Yeah. probably strong enough they meet certain requirements, but it it's not ready to take on everything humans do. We're not going to be totally put out of business. And so at Brave, we're using um the sort of claw opus cloud code um um opening you know chatp codeex whatever it's called cursor. >> Yeah. >> And uh we're using some of us a few of us my co-founder Brian Blondie are using open cloud. >> Yeah. Separate GitHub login. no access to any company keys. When it thinks it has a good PR, it puts it up for review.

Yeah. >> And actually did it's working on a backlog of bugs that are like low priority or harder to reproduce. And it's it's making progress >> and it's running in a Ralph Wiggum loop. So it just keeps battering away until it gets the results. It's it's this is an improvement, but again, it's it's like a tool in the belt that >> human programmers have to use. And so I I don't I don't think we're going to replace everybody. Maybe Spotify is just doing I hate to say shallow because it's JavaScript. It's my baby, but maybe they're just doing front-end code that is easy. >> Yeah, >> Havachi is fast [music] and private per trading secured by ZK with Celestia underneath. Go to habachi.xyz to start trading today. Looking for yield that stands out. Infinify gives you just that. Deposit stable coins and earn far more than the current savings rate.

Infinify is a battle tested, reliable defy platform that is ready to go. Head to infinify.xy XYZ. >> As the inventor of JavaScript, was was there any breakthroughs that you had when you were kind of developing this technology? And have you seen any similar experiences lately with AI that signal to you there's a place or something similar for inserting yourself in a new technology that can really thrust us to the next level? >> That's an interesting question because people have said now when I did it, there was, you know, I've been through four AI waves. There was one in ' 86 or so that was about rulebased systems and prologue unification pretty shallow crash and burn. Um one of those companies didn't make it. Um but in 95

when I did JavaScript I was concerned about making browsers useful and interactive and you know sort of two-way instead of just being a something that presents static content. when people started using these large language models with enough um training to do interesting programming. It turns out I've heard you have to check me on this that TypeScript and JavaScript actually work better than some other languages. I think this is just a matter of the open-source corpora that are out there in JavaScript and TypeScript. So, it's a better better set of training text and it's it's got a a bigger community. Like JavaScript kind of became the biggest language. doesn't mean it's the best, but it it it was on first and it spread like a successful gene. Yeah.

>> And uh it spread into a lot of domains like Node.js and and elsewhere. And then Typescript came along and said this was clever at Microsoft. We're just a supererset. You can write plain JavaScript. Any JavaScript program is a TypeScript program, but you can start adding type annotations and our tool, you know, visual VS Code will will check things and tell you if you're like passing a string when it wants a number. >> Yeah, >> that's good. And that led to this proliferation of TypeScript which I think is now in a feedback loop with the AI code AI. >> Okay. >> Yeah, man. I mean I I think that there's probably similarities to your 10 days at you know you know 10 days of building JavaScript because of me >> you know with uh Peter Steinberger from um OpenClaw who's just been shipping

thousands of GitHub repos every single day. >> Yeah. his story of going to Marrakesh and then getting you know news that the agent had figured out by uploading and converting some audio file into the text that was amazing. Yeah, it wasn't AGI right it's just doing better job of putting existing data together and transforming it as needed and >> using you know LLM techniques. Yeah, >> you know, some sometimes I think like, you know, it only took 10 days to invent JavaScript. Like to me that sounds like, you know, you were working smarter, not necessarily working harder. That >> was there some zero to one and that that really led or some breakthrough that >> you want the full story.

>> Three hours three hours with Lex Friedman. I didn't open the [laughter] can of Ultra Fiesta Monster, but it was sitting there. So, I was spiritually consuming it. [laughter] In the 90s, it was Coke. I think it was Coke Classic. too much sugar. But um [laughter] what what was going on in the early days because the standard has evolved quite a bit. JavaScript's become big and has lots of improvements. I was involved with the standards process for 20 years. But before we took it to standards, I had to get it from zero to one because there was a deadline and it wouldn't have made it otherwise. And then there would have been Microsoft coming up behind that with their monster truck and visual basic script which no one I believe believe me no one would have liked any better and probably would have hated more than JavaScript. So what I

was trying to do was get the minimum viable product out. And this is something I saw one of the YouTube comments on my interview with Lex said where I in in the interview I'm saying ah you know he said oh you needed objects you garbage collection. I said I left garbage collection out at first and somebody in the comments said wow this is making me rethink my lean engineering and how I should proceed. you actually deferred it till later and got things done faster. And I realized I've been working on things like I'm on version three when I should really just get the super minimal part done. And I I think that's the lesson programmers always have to learn and relearn because if you try to polish too long, you miss the market or you take too long or somebody else beats you to some goalpost.

>> Yeah, Brendon. Um we've seen quite a bit of you know waves with regards to people wanting to do computer science degrees just generally learn programming. You know, I've seen multiple tweets and different content pieces about how it was extremely extremely hot, you know, in in a I think it was like 20 or 30 years ago, kind of took a bit of a dip, now came up again preai, now post a kind of taking a bit of an enrollment dip. Would you recommend young people to study computer science given the state of programming today? Or how do you think about like the specialization of skills?

>> That's a good question. I I have I have children, one of them is studying, but he's doing math. I think if you do math, you can pick up programming, he can pick up programming, no problem. So become more valuable in the sort of applied math and linear algebra. It's not that matters for machine learning. He'll do fine on using these tools because they'll write the easier known code and he can use his own brain to write more advanced code. So I don't think programming necessarily is the way to go. I think the sort of fields that only humans can really do well at right now.

I know people would think the LLMs are going to invent new physics and new math for us. I don't think so. I don't Science fiction writers do it all the time and it's all who it's all woo right >> but um I do see these cycles right 30 or 25 years ago it was be a job programmer and people were getting into who didn't really want to be a programmer for the money and as Aristotle said when you do something for the money your work isn't this good you should go do something else >> be a pilot be a nature guide don't be a programmer just for the money so I think that that applies today too people might want to get into AI at the you know prompt engineering level this is where the artists are are waking up some of them saying we're going to be destroyed

by AI because it can copy images and do generative AI, >> right? >> Hope to get the right number of fingers and hands. It's getting really good. >> Still struggling with that, but eating spaghetti and >> Tom Cruz Brad Pit Fight. I mean, it's getting good, but you still need artists who can prompt it, steer it, and make it really good and make the next generation even better. And I think that that's where people go. >> Yeah. Go anywhere onchain instantly with Relay. Relay is closing in on 10 billion dollars transferred. Go to relay.link [music] link to bridge your assets now. >> Yeah, man. I mean, I think that uh it's something that that I've talked to friends about who are just getting into what they call vibe coding now and building out different softwares, different products and >> without too much slop because the slop

can easily rise taller than your UKQ, right? >> I I guess one of the last things that I wanted to ask you, I know that you're, you know, a privacy enthusiast, right? I know that you have the core blockchain and kind of original uh ethos at heart uh with Brave being essentially like you know come come use a more private more kind of cookie resistant more ad free experience online um and do it through you know this software that we've built when it comes to this new world um that that you're you know entering and really doing a lot of work in LLMs AI and And then really just like this convergence like where are we at with privacy or

like what are some of the core core tenants of this world that you'd like to see be prioritized and focused on? >> That's what Braves's AI approach does. I mentioned the browser AI which is using models that not if you use the default model like Quen or even the um anthropic models they're hosted on AWS bedrock and they do not log to anthropic in the case of models. So, Anthropic did a deal with Amazon. Even though they're closed, closed source models, they got them on AWS Bedrock. It's a hosting part of AWS and it doesn't send the logs to Anthropic. So, there's better privacy.

But if you really want ultimate privacy, bring your own model, use the latest, you know, more distilled or smaller models. That's where the future is. I think models are going to get more efficient. They're already modifying. We know this. And they're going to run on devices like you guys have in front of you. And that's not a privacy concern if you keep them from phoning home. So, that that we can do. Brave is really good at blocking trackers and ads blocking best in class YouTube ad blocker which has been great for our growth and >> that leads to questions like we faced at the founding how do you replace the lost revenue to valid creators that's one of the reasons to the basic attention token we have opt-in private ads that allow you to donate to verified creators if you go to bravebat.info you can see the

whole list of them I think Mr. Beast was near the top for a long time. I don't know if he still is. And there's a lot of YouTubers who signed up with us because we support them as distinct payies, not not the YouTube user generated content platform. They make plenty of money. We'll worry about them later. Um, but then you look at the world of agents. So agents should be doing things for you. Maybe they should be paying for things or be being paid for things. And this is the dream of, you know, cryptomediated agent agent commerce and agent reputation. We're getting into it, but the real problem remains security because security is not a separable concern. That's a famous paper from friends of mine. Or security is never done. This is like um Fred Brooks, you know, mythical mammoth sort of thing. You you you there's no silver

bullet. It's a set of trade-offs. And trade-offs mean economic scarcity makes you only defend the ones you think are under attack and you leave other ones unguarded. The attackers move there and you have to go there and it keeps it's an endless wheel. In fact, I think even with AI helping, we'll always be having employment for security engineers and yeah, security gurus anders. So, what we want for these agents, especially if they're going to have access to, let's say, a limited amount of crypto, and I'm talking to friends in the industry like like at Bitco, >> we'll probably start bringing this up, but you will not be able to drain your entire Coinbase. You will not have your Coinbase open in a tab over here and your agent doing stuff over there and suddenly there's a prompt injection from Reddit in the third tab and it takes

everything. Yeah, that's that's the nightmare case to avoid. So, we're working on it. It's a lot of security engineering. It's a lot of new threat models because like I said, it cuts across the old boundaries, sightby-sight boundaries that we created back when I did JavaScript. Mike's been on the show. It's been a little while, but love to get him and you on. Uh I I also want to ask just before you go, here's you probably recognize this. This is the model uh tracing chart. how much it can get done and how how much time this actually extends all the way to June 2028. You know, we're claudis 4.6 just came out. So, current time is actually right here. Do you anticipate that we're going to continue expanding this intelligence linearly because everyone says this is going to taper off and it's going to be it's going to be an S-curve

potentially, but I if you know the current pace of projection continues uh will look something like this. >> And this is a log scale, isn't it? Sort of. >> Yeah. So it looks linear but it's really exponential which right in nature is always part of a logistic curve an S-curve people say oh it's exponential until what the the singularity right this is this is something I don't think is going to happen so we will see some tapering I think the other thing people will realize and this will become more important in in enterprise it already is uh is that you want the privacy built in you want the local models or you want strong alignment guards or extra model checking or firewalls of sorts so you don't have all your company data flying out onto the internet >> and that's going to take some work and

some time and models because they still are nondeterministic they still can hallucinate you'll have this sort of I think >> uh you already have it the sense that well this is good it really does help it's a better front end to search it can write a lot of grunt work paper for me if the world needs more paperwork so [laughter] maybe we should get rid of that it can it can write grunt work code for me too right these these are valuable things but they aren't really going to bring about singularity so that will cause I think some kind of the bubble will solely burst or maybe there'll be a big worst I don't No, I I do, you know, worry about companies that have no line of sight to profitability like OpenAI. Anthropic looks strong, right? They have a plan. They're in getting in the enterprise. So, I think

we'll see over the next few years with that chart and with actual business or macro environment who's going to make it and who isn't, where the technology really works and where you have to put more care into how it's deployed, security, and and trust. Yeah, last thing to wrap up here, Brendan. Are you focused on micro payments in Brave? Like I think that's something that that all of us in the space see as something that crypto rails can really power. >> Yes. And the agents would be often doing it because they wouldn't be doing macro work. I mean micro work. >> And this is part of like the whole like uh software as a service kind of uh eradication of business model because we want to pay on Yeah. SAS apocalypse. We want to pay on an outcome based token

based kind of model seat based basis >> rather a seatbased or a monthly basis. So like what what are you guys doing there? Yeah. >> Well, so we when we did BAT we were on Ethereum and it was like too slow and all the Bitcoin bros which were even slower cost with the block size problems. >> Yeah. >> Being sorted out. We're saying use lightning and it's like lightning is not ready and by the way lightning requires you know committing all this Bitcoin. The real problem is I think Bitcoin is like gold. don't want to clip your gold to buy coffee generally and the guy that bought the pizza famously regrets it. >> So we rode the Ethereum wave but then we put that on multiple chains and ultimately on Salana we're actually doing settlement into self- custody. So we have no intermediaries like we had to

use initially uh for fear of finan like uphold and and Gemini were two partners. We had a Japanese exchange that was helping us settle that from this private ad system. Now we can settle directly on chain on Salana and there's probably, you know, more to come. I think we're at Nearcon. Near, I think, was just behind Salana and kind of missed that being the the third big blockchain because it's fast and the short time of finality, but Near has a lot of advantages because of the AI background of Ilia, because of the um smart contract uh I I would say better smart contract uh expressiveness. So, we're we're working on on things with Near. Some of them are yet to be announced, some of them will

be announced. I'll talk to my >> and that's specifically for like integrating um you know payments and these such >> yeah I mean there's a problem with you may know this because perplexity has been sued in is in court in the southern district of New York which is not a friendly jurisdiction by the publishers on your time right so this is this is going to play out into some kind of case law I hope not bad there's existing case law for Google search that says it can do like AI generated snippets of a certain size but it's really easy with all this text flying around the web. You can recreate Little Women, right? You can create a book and it isn't like anybody, any one person leaked the book.

Um, so there needs to be new >> copyright law, new intellectual property law. And it needs to recognize the rights holders ability to be paid if they should be paid. That's where, you know, Ted Nelson and Project Xanadu were trying to do, you know, microp payments too early and two-way links to make them work and all sorts of exotic stuff. And it was never practical, though. There's an existing JavaScript implementation of Xanadu that's kind of fun. If you go to zanodu.com, I think you can see it. Um, >> we are working with Near ons creators get paid. Probably something we can then merge with the bad creators program. We'll see. It's all very evolutionary. We don't want to make these grand plans that we commit to because for one thing, you've got these

large publishers that have extensive legal budgets and they're not technical and they're just inclined to use lawfare to freeze things. They want to create a frozen conflict. They want to shake money out of the AI hyperscalers and and then they want to just sit back and say we did it. But they're still slowly, if you look at the news industry especially, it's slowly dying. So we need a better plan including microp payments, micro royalties. >> Yeah. Yeah. I mean people are pretty excited about XR2 as far as standards from the blockchain world go as well. >> Yeah, we're involved in that. We're worried about it being a privacy uh terror, but really we'll work. Yeah.

>> Yeah. I hate to say it, but um Coinbase is is kind of embracing the blockchain as the new Google penicon because uh they bought Spindle. I like AGM uh and well done, quick exit, but >> they're they're saying, "Yay, we can see everything on the blockchain. We can build it up with ad tech. We can put it all together and profile you." Most people want privacy, especially for finance. Yeah. And uh if you look at how Bass has been uh evolving and how it's privacy policy has been evolving, doesn't look very private. So I think there's more work to do among more stakeholders there and X42 shouldn't just be a base play.

>> Yeah. Brandon, thank you for your time, man. It's been an absolute pleasure. Thanks. Thank you for your contribution to both AI and blockchain, man. >> I did drink. >> Cheers. >> Thanks. Have a great talk. Cheers, man. Thank you.