#30: Richard Socher - The Future of Search Engines

All right.

Welcome to another episode of Cyber Patterns. Today we have Richard Socheron, the founder and CEO of u.com, the search engine you control. Richard previously started and sold a company. He also loves paramotoring and photography. This is a really fascinating conversation with one of the brightest minds in tech. Hope you enjoy. Cool. So I think a lot of my listeners won't know what Yew is yet.

Are we jumping right in?

Yeah, let's jump right in. I mean, might as well. Awesome. So if you could describe what Yew is, how would you do that? I feel like I'm saying Yew so many times, but, and why did you start you?

Yeah, uh, U.com is a new search engine, uh, acts in kindness, uh, for us internally, uh, and it's a search engine that you control. What does that mean? You control your privacy, you control your sources, uh, you can actually contribute to it through an open platform, uh, which is especially relevant for, for developers. And, uh, it's a, a search engine Because you can control it, helps you save time, save your privacy, and give you control over your information diet.

Got you. And, um, so what's wrong with current search engines out there? What's wrong with the algorithmic search engine? Um, I mean, I hate SEO stuff, so is that part of it?

Yeah, that's right. So the beautiful thing is once you control your search engine, that gives us multiple ways to improve current search. So one, of course, there are too many ads on Google. That's one thing. But then even after you try to block the ads or you scroll past the ads, you'll see SEO microsites. And it's because you don't control the search engine. The search engine just wants to funnel you these days to the highest bidding advertiser. And sometimes that advertiser isn't directly in the search results. The advertiser comes in after you click on one of those SEO microsites and then you click on ads right there. And so like these SEO microsites are just trying to funnel a Google either SEO or direct ad into another ad from like Amazon or something where you then also search for stuff and see mostly like ads nowadays. So, so these SEO sites are essentially second-order effects of the advertising world and the privacy invading world. And so once you're in control of your search engine at u.com, you can also actually just downvote certain sources. 'Cause sometimes, you know, CNET and PCMag and so on, they do have interesting and useful content, right? And some people like it. Some people like Geeks for Geeks or W3Schools. It comes up very high in a lot of search engines. Some people viscerally hate W3School, and they, you know, on u.com, they can just vote down on it, and then it will update in their, in their, in the AI, in the ranker. They have control over it, and then they see it less. And so to us, putting user back in control of search actually solves a lot of the fundamental issues of search, of privacy invasion, losing a lot of time based on getting stuck in engagement loops, SEO, ads, and so on. Ads are complex, like we probably will have to get private ads at some point, similar to DuckDuckGo, just to be able to survive and then run a search engine. But we wanna make sure that they always stay private and commit to never doing targeted ads and stuff similar and basically share some of the privacy ideals with DuckDuckGo.

[Speaker:JASON_LEVIN] Yeah, so similar to Brave in that sense where they have private network of advertisers.

You know, yeah, like I think Brave is all about crypto and the BAT token and otherwise like it's a really great Chromium browser and yeah, we don't have any cryptocurrencies right now. We don't kind of have accounts for people's cryptocurrencies and stuff like that. You can use the private mode of view.com without any logins and things like that.

Got you, got you.

So, uh, the apps are completely private too.

Yeah. So when I, I, I think from like a crypto lens, just because I'm, that's where my full-time job is and all my, a lot of my writing, not all of it, but, um, when you were saying that people could plug in kind of different APIs into the search engine, it just reminds me of almost like a composable protocol. And, uh, wondering if you could explain more about that and and how people can do that and what use cases there are.

Yeah, great, great question. So there are so many use cases. It's actually really exciting. Can I share my screen?

Yeah, let me— I feel like it should be automatic to let people share screens or not in the advanced option. So I agree. If you could fix that too, that'd be great. While you're fixing the world.

So here you see, if you go to u.com/apps, you actually see all the different apps that you can like. And you know, some people love Reddit, other people like Twitter. I have to like both. And you know, there are all these different apps that people will like and that they can set the preferences for. And we actually have now started to have external apps. So here, it's the first version, still a lot more, you know, better, better never done. Search is kind of a pretty really interesting and exciting and tough problem. But you kind of see here, if you look for best headphones, for instance, there's a What HiFi app that has like useful summaries that we generate for people. And, you know, this is actually currently just their affiliate link, so they should be happy with just us hosting their app for now. And then you have Loria app. Loria is a new startup, and they actually use large language models and AI to summarize user reviews from thousands and thousands of reviews and then, you know, and star ratings and everything. And then they basically generate these AI-powered summaries and do price comparisons. And so to me, this is kind of a really useful app, you know, best for over-the-ear headphones, best for gaming, best for most wireless and so on. And to me, this is like a really great example of an open platform app. So in that sense, we actually are, very inspired by Web3 ideas where we basically want to collaborate with people, want to be much more open, and we want people to benefit when they contribute. So when you build an app like Luria, you can actually have also revenue generation, like the links when they do price comparison for you, probably their affiliate links, And that should be mutually beneficial. So here's another example for developers. If you look for Python Fibonacci direct computation, we have a Stack Overflow app, which I already liked. And so it comes up a lot and you have the code for it. And then there's a copy and paste button and then boom, you're done. And you compare that to your old search engine, you're like, well, it's a lot of blue links, a lot more blue links. A lot more blue links. I know, it's just like, it's not useful. Here you have like the useful things. You see the code right here in different ways, you know, some with a lot of comments, others with very few comments. You can read up on them directly here about more complex, you know, things if you want to read up on sort of details. You have videos, you have GitHub repositories and GitHub issues. So, and you have archived papers, and if the archive paper has like an open source implementation and you just see right away that open source implementation on GitHub and, and you know, like you compare and then you understand why, why we're growing and why people like us. It's just, you save time ultimately. And the more time you spend matters to use one that wants to actually help you save your time.

Yeah. The guys at Google must hate you, man.

You know, I, I mean, it's a, it's a trillion-dollar market and, and they're making so much money from invading everyone's privacy and everything that, uh, I, I don't know if we're that much of a threat yet, but I have heard that, um, they are working on copying some of our features, so I'm sure that will happen in the next year or so.

Yeah, they're so dependent on ads though that— yeah, I like the idea of a private network though, especially if you're, if you're focused a lot on developers and coders, you know, you could kind of tailor it. But yeah, so it's permissionless to build an app on it? On Yew?

It's not. Yeah, so we actually, we thought a lot about, could we put this all on some blockchain and so on? And it's actually unfortunately not, like the technology isn't ready yet for search in some fundamental way in the sense that in search, everything has to be in hundreds of milliseconds. Like it just can't, you can't wait for, you can build the best search engine in the world. So massive AI, it'll like write stuff for you. If it takes 10 seconds to get your results, no one, no one will use your search engine. So speed is absolutely crucial. And so right now we have to serve a lot of those things still, and it's still more centralized. Now we are very open and excited to see the technology progress. And then distribute this more and more. Ultimately, I can even see us eventually like open sourcing more and more of our own apps. That will probably happen very soon and maybe even eventually distribute everything in a much more sort of decentralized way so that people, you know, again, it's a search engine you control. You know, we're building this together with a larger and larger community. So there are lots of ideas I love about it. But right now you can allow people to just submit an app to us. They can build a little configuration, have a connection to some APIs, and then like launch an app. And if we had built that on a different stack than the one we currently have, it probably would've taken us a lot longer.

Got you. And yeah, I just love the idea even of like, say, say I was a developer and I wanted to build an app for myself, just like having my own personalized kind of search engine is what we should all have. Like, I, you know, you're definitely right about that. And it's, um, you know, uh, down the line it's, it's cool. I am curious, um, kind of just because like this isn't related totally to you, but more just you're a smart guy, um, curious what your thoughts are on like the state of crypto right now and, and if you're into it at all, or, um, you know what's kind of your head state?

Yeah, I dabble a little bit in it. I love certain aspects of it philosophically. You know, I think helping people be more autonomous, working together on things. I love especially the Web3 aspect of making your earliest users kind of participate in the upside of the platform that they are helping to create. I think, That is beautiful. I do think, especially your search, right? Again, like the latency, for instance, is just a problem. It needs to be so, so fast for search that, you know, unfortunately we can't have all the web on some blockchain and then just kind of query it from that chain and everyone who kind of hosts it benefits from it and things like that. There's some non-trivial things that we'd still need to figure out and we continue watching it. And then of course there's the whole financial aspect of it. And a lot of these things make a lot of sense, right? If you have a very concrete economic model, like for instance, people pay for internet. Now, if you provide internet through this decentralized system like Helium, then it makes sense to kind of map those existing economics into the Web3 kind of blockchain world. And then you can be very successful like the Helium network that's based on Solana. Now, I think if in search, for instance, things are a little bit different, right? It's usually user gets nothing and Google gets everything. And really the economics of search are happening when Google sells your data and your queries to the highest bidding advertiser and not between a user and Google, right? The user just gets like a useful function for free, like Maps and Gmail and everything, and they pay sort of behind the curtains with their personalized data, and then the advertisers are the actual product. And so it's much harder to map search into that world if you don't want to make it based on personalized targeted advertisements. And again, I think there's a big difference between private ads that feel much more like a billboard than this hyper-targeting following around the internet. I mean, most people don't know, but it's not just like that Google tracks you on google.com, right? It's not just on YouTube, but it's also on like very private websites that people usually go into incognito mode for. Even then, Google continues to track you on these private sites to understand you better and know how to market to you better. Now, you know, on the one hand, we could talk a lot about privacy. The truth is, while somewhere like, and this is plus minus depending on the survey, 80% of people say they care about privacy, but only 5% of people really act on that preference, right? Like most of us are still on Facebook, on Instagram, on YouTube, and so on. And so the reality is at You.com, when you're in control, we want you to have that personalization and convenience that is possible if and when you want it. So we have a one-click switch between personal mode and private mode. And in personal mode, you can still log in. You can set your preferences for the apps that you want to see or not. And you're not completely, you know, an unknown entity to the system. Because the truth is, like, it seems like most people prefer to have some of that convenience rather than extreme forms of privacy, right? At some point, like billions of people buy stuff on Amazon, like, well, Amazon will know where to ship your product. So they have your address and the convenience of them having your address outweighs the disadvantages of them having your address. And so similarly, you see that with Review.com and their features, you know, where we allow you to write essays, for instance, or write, blog posts with you, right? That's a very expensive AI-powered feature, large language models, takes a lot of money to host that and so on. So you can swipe your credit card if you want to use that kind of feature a lot. And again, you get convenience in exchange for certain capabilities.

Yeah, 100%. I mean, it reminds me of just— I forget where I was reading it, but it needs to be some middle ground between sovereignty and like none, you know, not— nobody really wants to be living off the grid eating like their own, killing animals and stuff. Like, you know, there needs to be something, there needs to be some government or whatever structure to take care of us and, and, or help out and, and whatnot. Um, I am curious though, uh, I checked out your website, just your personal, uh, website And saw that two things, AIX Ventures, would love to hear more about your thesis there. And then I also would love to talk travel and photography and just more about kind of your passions outside of work.

Yeah, yeah, those are both super fun to talk about. So I, ever since my first acquisition, kind of loved the idea of helping smart people increase their impact. In the world, and AI is such an exciting field, and there's so many different facets of AI that I just couldn't help but use the money from my first acquisition to actually invest in my friends and mentor some of my, you know, former interns and employees and coworkers and so on. And so yeah, I started investing a bunch in the past, and then, you know, there were a bunch of really smart people, and so they're doing very well. And so I thought maybe we can scale this up and bring in lots of other super smart people into this kind of system where we have people who are actually full-time practitioners, and we partner them with incredibly smart full-time investors. And so at AIX Ventures, we have a bunch of incredible investing partners, like Peter Thiel, one of the most famous roboticists in the world and professor in Berkeley and a startup founder too of multiple successful startups. Chris Manning, who's the most cited NLP researcher in the world and was actually one of my two doctoral advisors at Stanford, and Sunjii Saria, who's an incredible professor for healthcare and AI at Johns Hopkins and also a startup founder, and Anthony Goldblum, who started Kaggle, a really powerful and interesting machine learning community and competition website. And so all of those folks have full-time jobs, but they have all kinds of interesting people that they know also, interesting deal flow. And just incredible expertise to help those kinds of companies succeed. And then we partner that with some incredible full-time folks who are just like actually running the fund full-time. And so, yeah, I love that. And we've had a lot of success with incredible investments like Hugging Face or Weights Biases or Athelas, which is blood measurements and so on. So yeah, it's been a really exciting journey and just fun to see the area of AI kind of blow up. And then on photography, I guess I love paramotor photography.

Can you say exactly what that means?

I'd be happy to. Yeah. So paramotoring is essentially a propeller backpack that you can strap on your back and then you have a parachute-like wing, you know, slightly different geometry, and then you can fly anywhere. You can just fly it from like your backyard or a random like public field or anything like that. And it's just the best thing ever. And maybe I'll show you just a short video of what that looks like. So this is kind of a picture of a paramotor.

Oh my God. Yeah, I just Googled that too. I didn't know this thing existed.

This is crazy. You comment. Yeah. No more Googling.

But yeah, so it's— No, actually it was on you. But are you this? Yeah.

Yes, you can Google on you.com. You know, we may have to get comfortable with that.

Yeah.

Yeah. You can just go into the most amazing places, uh, in the world. Uh, yeah, I'm in Iceland and yeah, it's, uh, yeah, I'm in Utah. Uh, it's just the best thing. And it's so, you know, in comparison to any other form of flight, it's so cheap. Uh, but you can go into like the most incredible spots and you can make photos too of, places that probably in that angle no human had ever been, which is getting pretty hard, right? Like I also love motorcycles, but like if you wanna do something novel and interesting on a motorcycle, I mean you have to damn near kill yourself with like a triple backflip at this point. Whereas in a paramotor, you know, even as a hobbyist, you can go into lots of cool spots and you experience kind of very unique things. Here I'm in the middle of Arizona in some remote canyon, hadn't seen a human in a long time. There are no hiking paths there and like a bunch of wild horses all of a sudden, like, and it's just, yeah, it's the best thing ever. I love paramotoring and then combining that with photography.

That's incredible. So do you usually go solo and how'd you get into it? Also, like, how does somebody like myself get into it?

How much money do you need?

Or like, how did— this is crazy. I didn't know this existed.

So I used to do mountain biking a lot and my mountain biking buddy got into hang gliding where you have a rigid wing, fly a drone, like you'd be lying on your belly And so I got a little bit— I'm actually afraid of heights, but I believe in fear exposure theory, so I'm just like exposing myself to it, and it worked out a lot better. I'm still sometimes less comfortable up high, which is silly because it's actually safer because you can throw a reserve parachute and stuff. But, uh, long story short, I got into hang gliding, then paragliding, and then paramotoring. But you can actually just start, uh, directly into paramotoring. I have lots of friends who just go from no flight experience to flying a paramotor in a handful of days. And so it costs about, all in, where you get good training, you get good sort of instruction and all of that, you get certified, you take the theory exams to make sure you and everyone else is safe, and then you buy like some really good high-quality gear that you know, not too heavy, not too light, but very reliable 'cause it's a two-stroke engine, it's like a lawnmower. And a good wing that is also new and doesn't, you know, have any issues and so on. It's a very safe wing for your skill level. And so all in, it's maybe $15,000, one five. And then once you have that gear, you know, you basically fly for free. It's just gas money at that point. And it's not that much gas. So it's like, you know, 1 or 2 gallons get you to fly for an hour plus. And so it is the cheapest form of sort of real aviation compared to like a plane or something like that. But, you know, $15K is also still, it's still money. You can get down to maybe $5K, but at that point you're kind of getting closer and closer to a death wish because, you know, you don't want to have gear that's too old and too unreliable and the wing is porous and you want to have some good instruction too. You know, in the US it's kind of crazy, but it is completely unregulated, like, or very— there's very little regulation overall. Like, you don't need a legal certification to paramotor. You could just watch some YouTube videos and go try, trial and error yourself legally. Now I think it's a really terrible idea and I would encourage everyone to take proper instruction. I'm now a tandem instructor-rated pilot. Just 'cause I read.

That's amazing. My girlfriend is, uh, she's in Switzerland with her family and just went paragliding. Nice. And so I gotta, I gotta tell her about this. Um, man, I really wanna try.

Unfortunately overregulated. And, uh, it's Swiss allow paramotors, um, only if they're electric because of the noise, which is kind of funny cuz they do allow helicopters, I guess in Switzerland you're rich enough to make your noise.

Yeah. How, how fast are you going?

When you, uh, usually fly like 20 to most 40 miles an hour, you know, there's, it depends on a lot of factors. Like your wing is the biggest factor. And then of course there's wind speed, airspeed, and ground speed. So there are different kinds of speed, you know, as you're moving through the wind or over the ground, you have a tailwind and you know, you add your tailwind to, to your, uh, that sort of adds to your ground speed and so on. But yeah, somewhere between 20 to 40 miles an hour.

Gotcha. All right. I got a couple random questions if that's cool. Favorite books or current book that you're reading or just ones that have kind of impacted you as a person, helped you grow and kind of—

Yeah, maybe 3 very different ones. In terms of sort of really hardcore math, I really like the Convex Optimization by Boyd and Vandenberghe. It's just so well written for such a heavy subject. And very engaging, good examples, and just a lot of clarity and just good instruction. In terms of sci-fi slash just general fiction, I really enjoyed the trilogy of The Three-Body Problem. It's just so— especially the later books and the end of those. I feel like I have a pretty good sequence model of prediction. And most movies, I'm bored because I feel like they're so predictable. And especially, it's especially bad when the previews just tell you the whole story already. You piece together everything else. But the end of the three-body problem was so far out in the future that I had never even contemplated what those steps would be. And it gets you really stoked to learn more about physics again and just realizing that, man, we're really quite stuck in some fundamental things in physics. I think AI obviously is the most exciting thing. And it has a little bit of AI, but it touches more on the physics. And of course, as a really big believer in AI, I do think AI can also help us achieve scientific breakthroughs, which is super exciting. And then in terms of nonfiction that isn't hardcore math, like convex optimization, I really enjoyed— a book called From Dawn to Decadence: 500 Years of Western Cultural Life. It's basically a history book, but instead of the standard history books of like this king fought against this empire and blah blah, like it's all the intellectual ideas of like, you know, at what point— and it's the last 500 years of basically Europe and the Western world. And, uh, and, you know, sort of when was the first salon started, or like when did people realize that artists is not just like another craft, but its own kind of category of jobs? Or like, when did it change that Western, the Western world considered women to be more equal? And like all of these really interesting ideas and important ideas that shaped all of us. And obviously it has to talk a lot about religion too, and Calvinists, and how did it come about that somehow doing well monetarily is like something that has to do with religion. And your chances to go to heaven. And all these kinds of interesting ideas that influence our culture and their origins in different places. It was a really, really engaging book. It's kind of heavy. It's really big, but just has so many interesting nuggets in it.

No, that's amazing. I read The Three-Body Problem. And just as somebody who hadn't looked at physics in 6 years, 5, 6 years, there's a lot of things that I had to, you know, recalibrate myself with, which was cool. And, you know, I'm an English guy, I'm a writer, but it's still cool to like learn a little bit, learn enough that my brain can handle. And then, yeah, I read a book called Americana, which was like 500 years of American history, which was pretty cool. Or just like, not American, but America hasn't been around 500 years, but just the whole kind of area and like all about capitalism and that was a cool one learning about like the history of advertising. Yeah, that was interesting as well just to think about where it's going. Yeah, and then I guess another kind of random question, what do you think is gonna be kind of where AI is going in the next 10, 20, 30 years and And is the world going to look different? Are we going to figure out some new cures to diseases and all that good stuff?

Yeah, hopefully cures, not new diseases. You never know. It's general technology. So I actually just posted a longer tweet about this, or thread. I might just pull it up really quick. But the, I think, idea right now, and what we will see in the next few years, is more pre-training. I published some papers on word vectors before, and I think those were super impactful. Then I thought, well, how can we increase more pre-training and transfer learning between these kinds of tasks? And so worked on contextual vectors of sentences and so on. And similar ideas have eventually been expanded into computer vision. Of self-supervised learning and so on in context. And then the next level was to basically train these multitask systems where you pre-train not just the encoder, just the contextual encoder, but also the task and the output. And that was kind of started by DECA-NLP, a single model to train 10 different tasks in natural language processing like translation and summarization and sentiment analysis and so on. And then inspired kind of these really large language models to just basically incorporate the tasks, but everything's just a sequence, which is even more clever than DECA-NLP, which had a separate encoding for the question and the context. But then also all of this sort of alignment, and by that I mean basically co-attention that we just call it now. And so I think we're going to see more progress on that in the next couple of years by mapping different modalities into that same space. And this is something we've explored also 10 years ago where we map image vectors and sentence vectors into the same space, and then you can find one from the other and describe images through sentences. Then after that, I think we'll try to merge even more complex things together, train a massive combined models for language and video and sound. So bring in more temporal sequences. I think interesting stuff will happen if we then add programming code into that mix, and then all of a sudden you can, you know, ask it to do something about all of these modalities that isn't just based on a textual description, but maybe can actually execute more complex instructions over that kind of data. And that, I think, will be a very, very powerful combination of sort of reasoning that is discrete and fuzzy. Right now we have massive models with billions of parameters and floating point operations. But if you ask any of these models a complex arithmetic question, it still fails at it. It's kind of hilarious. But like, you ask it like, oh, what's 37.8 divided by 15.73? And it will do billions of floating point like multiplications and matrix multiplication and so on, but it will not be able to give you back the right answer. And so that kind of combination of logical and set-theoretic and algorithmic reasoning and fuzzy reasoning where everything's kind of continuous and a little bit more data-driven and correlation-based, the merging of these two fields will be super impactful. And then we could eventually pre-train models with good simulations and actual real-world grounding. And then since you asked for 10 to 20, it gets harder and harder to predict the future. And a lot of people who exude a lot of confidence, I think, have usually been shown to be wrong. It's just that 30 years later, no one is like, remember this guy who said 20 years ago that we're going to have flying cars by now? Well, he was wrong. Ha ha. It's like, ah, whatever. So no one gets really penalized for making bad predictions in the far distant future. So people keep enjoying it. People keep reading about it and being inspired or scared about it. But I think some of these are fairly safe to make. But again, the error bars get larger for your predictions the further out you go into the future. But I think one thing is clear to me to make that happen, you know, you may have to work on it yourself. But after all this pre-training, the next level we have to zoom out from is to actually let the AI choose what it is actually optimizing for. And that is kind of the technical terms, objective functions or goals, and let the AI kind of choose what that should be. Right now, all these AI systems are just making a very specific objective function better. And that objective function is usually, like, for large language models that will generate text and so on, it's just predict the next word. And that, you know, they have no desire to say anything in particular. They're just trying to minimize some loss function or maximize the probability of getting the right word correctly that comes after. And so I think we have to kind of really creatively start thinking about what could objective functions be for AI to train it in a way that it can get better and better over time. And, you know, people say, well, it's survival, but I don't think that actually makes sense for real AI systems because unlike biological life, you know, you don't have to constantly fight for reproduction and fight for resources with competing factions and the environment and so on. You could just be connected to a massive solar panel and float around in space forever. And so you don't need to constantly be fighting for it and then get sort of evolutionarily find your niche and exploit that niche and so on. So I think there are other interesting ways, but we're going to— I don't know how long you want my answer to be, so I'll stop here because it gets further and further into sci-fi and kind of interesting. But Jason Levin.

No, that's super exciting. I followed some of it. No, my brain was, uh, oh, um, I gotta read the thread. Uh, definitely a visual guy, but yeah, I think those were, those were all the main questions. Um, and I, uh, I appreciate your time, man, and appreciate that you're, you're building a project that is, you know, a digital public good. Um, actually, yeah, I'm writing an essay about that right now and gotta include you. So Yeah, yeah, that, uh, thank you so much. Really appreciate it.

It's great to get the word out. And you know, the whole internet is going to take, like, everything in the economy is moving online and onto the internet. And so it's very important that that entryway is not a gatekeeper that just wants to sell you to the highest bidding advertiser as your main model, but it is a more collaborative, uh, place and platform for, for people to work. Come together. So yeah, really excited about it also. And thanks so much for your time.

Yeah, yeah, 100%. Um, wow. One last thought, uh, I just had a conversation with somebody talking about how like virtual worlds are almost like campfires or their own world, and you is almost like the entrance to the different worlds. Um, and that entrance should be a friendly one, not, not a scary one where you're getting hazed by ads. So, but all right, thank you so much. Appreciate your time.

Thank you. Have a good one. Bye.