Engineering A ML-Powered Developer-First Search Engine with Richard Socher - #582

Key Takeaways

all right everyone welcome to another episode of the twiml ai podcast i am of course your host sam cherrington and today i'm joined by richard socher richard is co-founder and ceo at u.com uh before we get into our conversation today please be sure to take a moment to head over to apple podcast or your listening platform of choice and if you enjoy the show please leave us a five star rating and review richard it's been almost exactly two years since we spoke you were at salesforce then welcome back to the podcast it's great to be back man a lot has changed in those two years a lot has changed i think uh the last time we spoke to you were like holed up in some exotic bunker uh you know early in the pandemic um but we made that conversation work and i'm super excited to get caught up with you you've been a busy guy for the past couple of years it's been yeah busy b days um but exciting can complain awesome well for those who uh don't know you i'd love to have us uh start with you sharing a little bit about your background uh before we dive into what you've been up to recently sounds great yeah hello everyone i'm richard uh where do i start i'm originally from germany did my phd at stanford um in deep learning for natural language processing and computer vision um and i thought it would be great to have more people use deep learning for natural language processing which was uh still quite uh um received with quite a lot of skepticism back then uh and so i started teaching at stanford deep learning for natural language processing uh in 2014 and uh well um needless to say nowadays it's kind of obvious that like that is the right technology after four years for a variety of reasons but one one being that everyone was then teaching deep learning um as the main approach for natural language processing i stopped teaching my main job was uh doing a startup metamind an enterprise ai platform we got acquired by salesforce where i became the chief scientist and then build out the research group there and eventually also a lot of the aia product engineering um teams that i let and then couldn't quite take this idea off i actually during the last couple of days of my phd i had implemented a first little prototype of a new search engine that was going to summarize the web for people and be just much more useful and quicker in google at the time i thought oh man all my smart friends are going to google no one's ever complaining about it where as far as i can hear it's just maybe too audacious of an idea but i couldn't quite shake the idea off over the last eight nine years and after after many amazing wonderful years with the great teams at salesforce i decided i think i need to i think i need to do this i think the world needs a better search engine for a lot of macro reasons as well as sort of user user reasons and so uh i decided in uh the summer of 2020 to start this with brian mccann um one of my amazing collaborators uh and co-workers uh both at stanford and at salesforce and and yeah couldn't be happier with brian and the team we've built at you.com when you talk about uh the meta and user reasons for a new search engine what exactly does that mean what's the the motivation for for you.com yeah so and a high level reason it's kind of crazy that uh the entire economy is moving online and you have the single gatekeeper in the beginning of most people's online journey that mostly wants to sell you to the highest bidding advertiser and you and your queries at the same time we are in an information age and there's information overload 20 years ago it's hard to get access to information but nowadays it's actually it's almost too easy to get access to a lot of not that useful information and you need ai to help you deal with this flux of information to help you summarize all the things that are going on and get quickly to what you want to actually achieve and do and you give that intent usually to a search engine but now sixty percent of all google queries are zero zero clicks meaning they don't leave the google ecosystem anymore they stay within youtube within maps and they try to suck you into these engagement loops rather than trying to be as useful as they can summarize and then get you on your way either somewhere else onto the internet or um or just that you execute on the intent that you have so if i search for um how to uh to sort array by value or something in python i just want the code snippet and that's what you know you.com one of the many many features that we just give you there's a code snippet and a copy and paste button because we know that that's probably what you want instead of a list of 10 links and you go there and you don't have good string matching and whatnot so those are just that's just one of many examples um that's helpful for the user connects to the macro then there's sort of you know outside of ai and machine learning um reasons uh and just that when every company has to pay this tax to exist on that first page which is you know by by paying for ads it creates some really odd incentives that we've that i've observed now i had multiple people kind of tell me a story where they got organically up in the google ranker and then uh you know they make start make millions of dollars because the content uh was just good um and so google ads team comes over the sales team and say hey do you wanna you know buy some ads to get even more traffic you're like no we're good we're getting so much traffic they basically disappeared and went to page nine and yeah and then like sorry yeah we'll buy the ads and then magically they come back to page one with their ads too that's kind of uh one of the one of the many reasons to do it and then in some crazy way also now people are actually complaining more and more about uh just the ranking and the relevance um do you have all these seoed microsites that kind of look at try to reverse engineer the algorithm that google decides for everyone what to see and read and consume and buy and they're trying to reverse engineer it and so you get all these like really odd microsites that have a bunch of sort of language model samples on there that are known to resonate well um you look for like good machining tools for like uh building a roof or something and at the bottom there's some weird wikipedia article about california on the page that comes up highest on google it's because they know oh that stuff ranks well in the algorithm so there's all this reverse engineering and i think part of the problem there is that an is partially an ai problem but it's also a system systemic problem and how you approach ai is that google decides and wants to be able to decide for everyone what they read consume and buy because they have so much power by then showing you mostly ads which is also just becoming ninety percent of any page of the results that you know when you actually want to buy something and have a monetization it's like it's it's it's it's degraded surprisingly over the years yeah and so i think it's important um to uh help people still get stuff done um but not just try to do that through ads and by giving them also some control so we have of course a large neural network too to rank uh what we actually think about more is the apps that you're looking for like big content islands like reddit things like that stuff where you know social signals that people actually care about rather than seo microsites and but we actually give people control and say oh i like the source or i don't like the source and that way if you try to manipulate this too much by playing sort of you know seo games people will just downvote you and you'll disappear uh and so i think giving people control over the ai that influences so much of their information diet is yet another reason i could go on forever there's so many reasons why but yeah i'm it's also one of the most exciting ai applications that's so important in an information age it tackles summarization which is one of the big hard unsolved problems in ai and and so on it's just the most exciting thing i couldn't not do it awesome awesome when you you know starting to build a search engine in 2020 do you start with something that's fundamentally similar to a pagerank type of an approach or are you you approaching it very differently yeah it's a good question we are actually approaching it quite differently um we we have not sort of replicated the list of blue links we're getting that from an api but what we actually are doing is using large neural networks to understand what the intent actually is and then try to give you the most useful application and we want to be a much more open platform for you know building these applications out so that different people can actually implement um and contribute to that first page of the internet and make it very very seamless and so we have essentially relied much more on the content and the semantics um and then of course we can nowadays already extract a lot of popularity signals that you used to need pagerank for because a lot of like recipe sites and code snippet sites like stack overflow they actually have votes on how popular something is so you can extract that that's part of the signal together with the natural language processing signals so on top of the kind of core search engine you also recently announced a couple of extensions well at least the code is new is right new also yeah yeah it's a lot of new stuff coming out you know we're thinking about how can we be more useful and of course if you build a search engine nowadays the really tricky balance is how many of your resources do you spend on just catching up to google versus how many resources do you spend on doing something that's different that google doesn't yet do right and i'm sure in the beginning when people saw google flights they're like oh why does a search engine help you book flights and we get similar things now when we write an essay for you like you search how to write an essay about world war ii or the american revolution and you know this and that uh it'll just write you an essay and you can you know modify it and then generate new ones and that helps you with blank page problems of like okay how do i start uh you can want to write a blog post about you know your airbnb project or whatever it is um and so we think like that is something unique that ai can also bring we have these exciting large language model applications now twice uh in the search engine uh one is uh in code completion so if you look for you know how to do something in python or other languages uh the code complete uh app comes up and it's just a full-on like github co-pilot like model that essentially just auto-generates the code that you're looking for and then of course you have you know stack overflow apps with copy and paste buttons you have simple tutorials and you know to sort of get you started with things and you have hugging face and pytorch kinds of um official documentation tool and it's all just right within their first page with code snippets and copy and paste buttons to be very easy and then on the writing natural language pro like natural language side we have this you write app under you dot com slash right where you basically can have any i just write an essay or blog post for you got it got it so kind of stepping back you've got ai throughout this platform you you've described a couple of applications that uh are you know i don't know if you think of them as part of the search engine or kind of a json or on top of the search engine engine but then you're using ai as part of surfacing relevant results to to folks you know thinking about that that core maybe dig a little bit deeper into some of the ways that you're using uh machine learning to deliver relevant results to folks did the large language models come into play there or is it more on the summarization side yeah it's a great question we there are a bunch of summary features too um and ai is sort of both in terms of our product um sort of very deeply um and we can get into that like intent classification slot filling uh ranking of different applications and so on uh it's also um a user persona that we know a lot about and care a lot about just people who program uh ai applications and you know hence we have all these specific tutorials and uh documentation sets for for hugging face for instance um for pytorch and a bunch of other things we have github issues we crawled all of github so you can find all the different issues about your code that you're working on directly all in the search engine when you when you're basically trying to understand a query you want to understand like what programming language someone is using or we like if you say i want directions from san francisco to la it's kind of a standard example this is a good one for slot filling where you want to extract sort of certain things from the query directly and input them into an app and then basically let people immediately uh kind of have that filled out the slots of you know the front direction the two direction so you get the time that it takes you to drive somewhere and the directions uh for that so that that is an example of slot filling that we have to do for 150 or so apps that that we have um if you look for dow jobs or web 3 jobs for instance then we kind of extract that's the kind of job type you're looking for and we have a bunch of apps that uh basically show you job listings uh for these kinds of uh categories so slot filling is a is a pretty big part of the search engine then of course you have the intent kind of knowing oh it's just the weather intent and that leads then to influencing and kind of end-to-end training a large um a large ranking model that basically ranks all these different apps uh in terms of their priority for helping you get something done and so when you're are the the intents or the apps are those kind of top down you know created by you or a project manager hey we're going to need our travel app we're going to need our directions out we're going to need you know this list of you know hundreds of things or are they kind of bubbled up from the queries that people are making themselves yeah great question uh you bring up two good points one is namely privacy uh we we care a lot about privacy i think it's really important uh important right uh and so we have a private mode in which we don't track anything so we don't know really what people are doing we don't have the queries but we also have a personalized mode and so some people actually want to give us feedback um and and tell uh tell us what they want so we got we have a very active community of thousands of people that give a lot of like feature requests and you know it's kind of tricky sometimes because search touches really everything right everything we do often online like starts with search um and then uh we also yeah can look at sort of um churn queries like things that you know people tried to do that they couldn't do and then they leave forever uh and so uh sports was a good example of that that just kept bubbling up uh people wanting to see sports results um and so we now have um actually releasing this week a bunch of you know results for for sports that are live coming in from uh different apis uh so you can kind of see what's going on so yeah it's a it's a mix of uh direct feedback and then some indirect feedback can you talk a little bit about the way that the ways that you're using summarization you've referenced that a couple of times yeah yeah i think i think summarization in the limit is actually one of the hardest and most interesting and impactful nlp applications of our time right now because we're in the an attention economy and in the information age we need better summarization but really if you think about it uh if you're deeply ingrained in a space and you want to get a summary for a new paper it's very different to when you've never read anything in that in that space right i sort of um because of progen this protein generation model i've been working on uh at salesforce um you know i've been reading a lot more bio papers and like it there's just so much lingo that you don't know that and if you were to try to summarize that make it even shorter you you would not understand anything and so the summary might need to simplify things but also add some explanations for some things and then you think about things where you're an expert like oh the elmo paper was basically like the contextual vector code paper but instead of translation they use language modeling and then bert was essentially like elmo but instead of you know an lstm model they use a transformer model and so that's like a one sentence summary if you know exactly what all of these things mean already you're like boom i get it um and you know their models are larger it's probably also a good add-on to that summary um but if you don't know what a language model is what a word vector is um what you know what a neural network is on lstm and bird and and all of these things are then these these are non useful summaries for you so long story short summary super interesting super hard ai problem um but uh for us you know we're to say okay well we don't know that much about what what a user knows um how can we start with something that's useful for pretty much everyone uh where we started is basically with coding like if you look for this here's like the most relevant code snippet um that's in some ways a you know sort of multimodal if you think of programming as a different modality to natural language multimodal kind of summary and then another one is just pros and cons so if you look for best headphones for instance you want to just extract what are the main pros and cons of this particular headset so we kind of extract those from professional reviews uh review sites you can very quickly skim a bunch of results and know what are the pros and cons of the different headphones another one is recipes we heard a lot of people complain about having to read the whole life story of someone if they just want to get like a lasagna or chocolate chip cookie recipe so we extract it like here are the ingredients here are the 10 steps to actually make the cookie and then you're done so those are examples uh that you already see in the product on you.com that you can just kind of see a useful summary that pretty much for everyone is going to be universally useful and are summaries the way that you're using them are they tending towards more generative summaries versus extractive summaries or do you use both in different places you have to you have to kind of use both our values are trust facts and kindness and if you think about it as much as i love these language models you can't quite trust their facts yet right they they make stuff up right you can just say write me an article about how hillary clinton won the election and it will write you a perfectly reasonable sounding article how that that happened you know and so the veracity and just like factual correctness of uh of large language models is still you know uh iffy sometimes so you can't just let them generate uh you might get some some pretty not so great tasting uh cookies if you just do that so you have to you have to start with with some clear extractions but then you can kind of simplify things um and get rid of redundancies for multi-document summarization and things like that uh your example reminded me of um one of the issues that came up with google sometime this past year um where i think the you know just one of these examples if you type in a query what should you do if you know someone's having a seizure and like it extracted the things the list of things not to do that's right and um you know i'm curious if you can talk about you know in in the context of using similar technology uh and trying to to present information you know in a similar kind of summarized way like how do you you know when you're starting from the ground up build guard rails you know so that you can achieve this level of trust that you're aspiring to yeah it's a good question i actually think you you have to keep your users and yourself and the use cases in mind and the more important use case is the more important it is to really have uh some human oversight you know i love ai i think i can change anything from cow diets to reduce methane to you know agriculture to medicine and creating new protein structures and all of these different things but when it comes to like life-threatening information you still need to have some human oversight so for instance when you when you ask how to commit suicide we just have a handcrafted message that your life matters and you know don't do it here's a suicide hotline you don't want some ai to kind of do that and of course there are some interesting uh sort of chatbot applications too to have longer conversations with someone who is thinking about this and i think like companies like replica uh where full disclosure i'm an investor to you know have done a phenomenal job kind of being essentially a journal that talks back to you and seems to care about you um and and kind of helps you work through issues so there are some exciting and interesting applications uh in that space as well but yeah we are basically thinking about the more impactful the application is the more careful we have to be in just letting ai run run off and do its thing you know when you were when you were working on metamind i'm imagining that uh you know you had to be pretty far out in the you know let's call it the research domain to you know get these raw tools to do the kind of things that um that you wanted them to do you know how how uh much has that changed or not now like are you able to kind of are you using mostly off-the-shelf stuff or are you you you know thinking are you required to do a lot of you know research oriented things uh you know novel architectures or novel training methods or or that kind of thing yeah that's a great question i think it's actually gotten so much easier to build ai companies and just functionally like useful algorithms it's it's been really incredible to see like back in metamind we still build those cnns in c plus like and so it sounds like the your general take is that there you have a lot more ability to kind of pull things off the shelf and um and you get the you have raw horsepower there that you're not kind of needing uh to push the innovation frontier in that way yeah so uh yeah um that i think you can build a lot more you can have a lot of impact with the applications in fact we've done so much research that you can just have i think currently more impact in just applying that research to all these different domains and problems that you see in the world and workflows that are not yet automated as much as they can um then you could i think in pure research right now it's kind of been in terms of pure research and sort of major ideas it's been mostly executing on make the models larger um efficient to optimize and you know paralyze on current hardware um and and get more training data and collect that training data in reasonable ways and think about the bias and so on um and the models haven't really changed and honestly even if like there's a different kind of model we just need a very large general function approximator that's efficient and has enough knobs to tune um and then you can kind of optimize that whole that whole setup my hunch is if it wasn't for the particular hardware that we're using we would probably be we could use any other large model too i think lstms aren't inherently worse they're just like less easy to optimize and paralyze and then train henceforth on our current hardware so so that's one one aspect of the answer the second one is we definitely have to still innovate because there is no you know hugging phase model to build a neural network ranking mechanism that takes into consideration your intent your slots and then 150 of the apps that you know we built for the first time ever no search engine ever had before and then ranking them properly so you do still have to innovate and build brand new models now for now we can't really publish those um i actually hope that at some point we are safe enough as a in terms of the existence of the company that we can be much more open in the future uh but yeah so the answers of course we still have to massively innovate because of these hard and interesting new problems that we're tackling where we also don't have that much data right so large language models enable us to to know that if you look for you know a chinese restaurant near me or an east asian restaurant in my area or close to where i am and all of these things they all mean the same thing that was you know something that in the past you couldn't know and you would have to get a ton of training data and then you know be able to actually map them all to a similar kind of uh you know natural language sorry api language to then triggers we essentially have to translate human language into the language of computers and apis so so there has to be a lot of innovation on our side for that and then maybe a last note since we brought up metamind is that back then metamine tried to do i think so many different things like help you kind of label your data and drag and drop it into the browser um that's you know scale and crowdflower now and scale is like an eight billion dollar company or something uh that that piece of itself that was just like one of the many features of metamind and then you know deploying it and scaling that deployment and helping you do error analysis and then just making it a simple python interface to actually run your ai classifier or model in production all of these things now have companies that are valued in the hundreds of millions or billions of dollars each of these one like separate features that we had implemented metamine from scratch pre you know having anything like pi torch or tensorflow and so it's just fascinating how good the tooling has gotten for ai um and that was sort of my tangent on like we're investing in uh both vertically integrated but also sort of tooling companies at aix ventures and and because of that it's gotten a lot easier for people to to have impact in those applications you reference the particulars of the hardware that you're using is that say that you're using kind of non-traditional exotic things or just uh that you know there's this affinity between gpus and and transformers that allows you to scale yeah there's uh we definitely are using very standard hardware uh because we need to scale we need to be able to fill up the data center in um uh in europe or in asia and in different different places so we don't really want to rely on any non-standard hardware right now and even gpus are often there's a shortage so we have to sometimes map some of the gpu models and see if we can make them fast enough on a cpu just so that we can have more data centers uh and have you know less lag time um when there are not enough gpus available in certain geos so yeah it's uh it's an interesting interesting challenge for sure and how how uh are you finding the level of maturity from an engineering perspective to allow you to achieve that that kind of scale i guess it's just about the people like we've gotten really lucky and having hired like an incredible um ai and engineering team and also devops you know just like like spawning out all these machines with a click of a button you have a whole new data center and reduced latency for people in a different geo it's it's yeah mostly about the people um and i'm imagining you've invested significantly in kind of building out a enabling platform that allows you to kind of develop and train models quickly and get them into production quickly was that a big focus yeah yeah that's definitely um it's definitely something we also are relying on you know things like weights and biases full disclosure i'm also an investor in them to help with experiments uh and then running those uh and you know there are a lot of a lot of good tools that you can use now but ultimately to actually spawn out the whole system and like have a new end-to-end you know search engine that runs with a bunch of different machines and so on they're all communicating um that is still something that we had to build ourselves from scratch and then we also want to make it easy to create a new application so now we just have it so that you have a new json like data dump and then boom with like a config file and you have a new application within you.com and so i'm excited to in the future essentially let anyone kind of build that and have search capabilities over all that data so i think automating search over new kinds of data sets that has been uh you know an interesting and tough challenge that we tackled we spoke a little bit earlier about the code module code application in a lot of ways the kind of the use case that you described of hey you know i'm searching stack overflow really i just want this code snippet uh i think i mentioned that in the conversation with greg brockman about codex you know how that's you know ultimately what we want you know what we need um you know in terms of the process of building that module can you maybe compare contrast with what you've seen uh others do around you know codex copilot um that kind of thing yeah i think um you know i think there's basically i i think about this in kind of two levels either you're trying to solve a problem that people have solved before and at that point you just want the direct code snippet um as is or you're trying to combine it uh and and have you know sort of a new combination of problems that no one has yet quite solved like this before and at that point you need like codex um or like our code complete and you dot com to just give you the answer and generate something novel uh and and it's kind of incredible how these models aren't just kind of able to deal with things inside the convexcell but really in the hypercube of like you know different combinations uh and combinatorial i mean combinatorial combinations of things they have seen in the training data to just generate and then combine them so uh yeah that's kind of how i think about these these two levels of generation it's a ground up model that you built as opposed to an api that you're using no yeah we're also using api for the code generation oh for the code generation you are okay uh for the models that you're building what are some of the training data sources that you rely most heavily on i guess there's you know sort of large-scale um internet available data um that is there and then we have to crawl a ton too um so that is probably the biggest one and it's been something that i think a lot of uh machine learning leaders can relate to which is everyone wants to come in and build cool models and you know but then they realize man that's really hard so they download hugging face models and they just kind of work out the box and and then the biggest thing is that you know everyone wants to not very few ml engineers want to spend a lot of time on data um which you know let andrew to say oh let's just have data competitions of like whoever can get the most interesting data set for for this problem and that's something for us that that often meant we had to spend a lot of time crawling and eventually we just hired some people who are actually excited about crawling data and getting us that data that we need to then be able to actually train summarizing models summarization models and so on um later on uh and and so so yeah it's been a continuous challenge to crawl and it's one of the many places that the monopoly of google comes into because there are some sites that say only google is allowed to crawl us no one else is and they're like well you know how are we ever going to beat google if we can't do that um and so there yeah all kinds of interesting challenges both on the technical side but also the sort of systemic side okay you mentioned earlier um when i asked about pagerank uh i thought your response was saying that you weren't crawling for kind of the the index um but rather you were consuming that via an api um so is it that's the index but you are calling for some of these other applications you're building is that the idea yeah sorry that was uh yeah there's some ambiguity there so we actually think that the list of like a blue link of lists a blue list of links isn't going to be as important anymore as you know the actual larger content islands like reddit like medium like twitter and so on uh and then in order to be able to summarize things you also can't really do that on the fly these large models are not fast enough people want things in hundreds of milliseconds and took us a long time to for ninety percent of queries now be faster than you know duckduckgo and other competitors in the search engine space and almost as fast as google at least when you're close to our data centers we don't have as many all over the world of course but uh long story short um we are actually crawling a ton of data in order to build these apps and make them fast enough there are also some several times where we thought we could rely on an api from someone else but then just the scale and burstiness of search um and when you get tens of thousands of queries um in a few hours like you just no api was able to deal with that we have to build uh and have that content ourselves index it be able to do interesting vector search operations and things like that with the data all in a few hundred milliseconds to then be able to surface the right kind of content uh very quickly so yeah we're kind of slowly crawling um the web through the most important content islands like stack overflow like github uh like you know pie torch or hugging phase documentation or all of medium which is also pretty pretty large so there are a bunch of interesting uh things that we are you know sort of we have to crawl ourselves just to be able to have the speed and the ai capabilities that you have to run offline the the goal is to produce a a better general search engine um but you've also specialized in some ways that makes it a super interesting search engine today for more technical folks um yeah how do you think about like when you hit the knee of the curve that it's like better for for everyone yeah um what we've learned so far is that we're better for developers already like a lot of people um i posted a couple of uh features on a twitter thread for our you know you code kind of special search and it blew up like crazy hundreds of thousands like 300 400 000 impressions thousands of likes and so it resonates a lot with that crowd now what we've learned is that um i sort of jokingly say it turns out developers are people too and they want to know what the weather outside is and what the sports results are and how to travel and like all of these things and so um you know where to buy food and maybe order food and so we have to basically if we want to be the best search engine for developers and be your default and be there every day and in your navbar through you know chrome extensions and things like that um we have to be uh able to do everything else in search too which is tough for a small startup um but we've now gotten to a point where once we launch sports results there's maybe only the travel category where we're not as good for everyone else and then most other things we actually are uh you know we have answers for from movies and and things like that and you know there are still some apis like the movie api that is a little bit slow so it takes like two or three seconds to load rather than you know less than one second um and we've gotten complains about that um also um but but yeah um there there's some proprietary data that we could probably just crawl i guess the law just kind of changed a little bit because linkedin lost a big lawsuit um that you know they tried to prevent folks from from crawling data but long story short a lot of crawling's happening and the speed of speed is always super important um and we we are we are having to build um a lot of that in-house of course congrats on on the the launch of view.com and and u.com code and right uh before we part ways i did want to circle back on a couple of the things that we spoke about um that you started at salesforce it sounds like you you're still working on those the protein generation one we spent a fair amount of time talking about that the last time we spoke and we'll include the link to that in the show notes we didn't i don't think spend much time talking about the ai economist i think the timing didn't quite work out to dig into that um so we'd love to have you share a bit about that project and i think you have some recent news there any recent updates that's right this week in machine learning um we we actually got the science advances paper out about the ai economist and um maybe just at a very high level what is the ai economist it was word by uh stefan stefan chang and alex trot and and a few others at salesforce and and myself and basically the idea is using reinforcement learning for some of the most important applications that we can think about for humanity period instead of having rl play games that are kind of interesting but ultimately themselves not very useful um why don't we see if we can build a very realistic simulation and we're far from that in terms of realism right now this will have to scale up over time too but i think it's a brand new area of ai that can have a huge amount of impact and so the high level ideas you have a two level reinforcement learning problem where you have an rl agent that set that sets taxes and subsidies um for a bunch of other rl agents that which themselves are just trying to optimize their own utility as in they try to maximize resources they can collect money they can make houses they can build blocking of other people from resources by you know building houses around them things like that um and are basically to some degree more selfish um and you know but may also eventually identify patterns to to collaborate um towards their own selfish goals uh and maximizing their own utility functions and so the interesting thing is now you can give that top level um rla agent that sort of the ai economist the ability to subsidize or tax uh different income groups differently in order to maximize a specific objective that you've given that ai so the idea here is that you can now say oh i want to help the middle class or i want to maximize productivity of this economy or i want to maximize equality in this economy or a combination of all of these things that you wait and you say okay i care about uh the one we chose in the end was productivity multiplied with equality which is one minus the genie index uh it's essentially thinking about how how equal you want to be you don't like you know in the limit you don't want to be like everyone is extremely equal but extremely poor right that's that's not helpful too so you have this like overall productivity um in there as well and so that that is kind of the high level and so what that means is that you know if you if you take that idea and you really scale it out and you make uh the simulation more realistic you increase the size of the number of agents to hundreds of thousands uh and you actually put in sort of historic data into this that you know to start the model that in the future if a politician says oh i'm doing like these following five things to help the middle class or to help this particular group of people whoever it is like worldwide right then you can run that suggestion across and compare it and contrast it with millions and millions of years of simulated taxation where you basically try to identify what the fairest or best or most sustainable or most equal or most productive way is to tax that entire system that touches on highly philosophical things like you know communism capitalism socialism market market economy uh and so on um and combinations of these systems um on the one side but it's very concrete like it'll you know could change and be another input um into economists models to be more realistic it's kind of crazy but you know there's there's models that are being used right now like the science formula very famous uh berkeley professor in economics um who has this provably correct uh way to tax different income brackets but it's proven to be correct in a one-step economy it turns out people iterate right like turns out time moves on and and so this ai economist model can actually recover the approvally correct solution for a one-step economy but then as the models learn as the agents adapt as the time continues like that model just is so much more powerful and realistic than any of the linear models and one step models that we're currently using that it's just hard for me to not see how that won't change all of economics which and the grand scheme of things has been an area that hasn't been impacted by i nearly as much as i think it could or should and if you think about how much bloodshed there has been in human history to identify what the right model is of taxation and representation and things like that like it's just so powerful to be able to try to offload that into a simulation get millions of years of taxation going and then uh you know learn from that and see if we can use some of these things and of course you know like we don't want like in a dictator either like we need it as like another data point as a model that helps us make more um you know uh more accurate uh like decisions um but ultimately people still want to decide what the objective is so that's still like a very important one and you have to sanity check it of course before you implement these things i'm not like absolutists like this this has to be like a new religion or anything but like it i think it's just it's such a powerful tool and and i have high hopes that just like what we've seen in linguistics and natural language processing or we've seen in computer vision or we've seen in robotics uh we've seen in self-driving all of these different application areas of ai that economics could be another such application area sounds like a model that would be really good at the sims yeah it's it's not actually crazy to think that that is like a pretty good simulation now of course the problem is that it doesn't capture like sort of as realistic utility functions that people have like in the sims like like people might not get as tired and then like just don't want to work anymore because they want to sleep and things like that so you wanna you know adjust and for most people like you know there's sort of logarithmic happiness curves too like making like an order of magnitude you need to make an order of magnitude more money often in order to be like a little bit happier and then it sort of levels out logarithmically there are all kinds of interesting things that we have found in psychology but what's fascinating too is that you can actually say well i think people are this and that like i think people are going to want to work more or want to work less you can actually make that very explicit in the beginning of the simulation and then see how those assumptions about how people you know define their own utilities will actually influence the optimal political model um or not political and to some degree you know if you kind of group all of these different uh um policies uh into one cluster but you know just generally sort of uh taxation and financially does the work on behavioral economics you know things like predictably irrational all of that you know does it does it say that kind of akin to what you're saying that everyone has their own utility functions and they're not as uniform as traditional economics might like or you know is there is it more that you know there's just an emotional irrational component and if that is the case like how do you even model something like that yeah it's a great question so you can you can have a prior distribution and then you sample uh different like you know you sample from that prior distribution that you have for instance for utility uh functions um and then you know based on that sample uh and based on how you define your prior distribution uh you can get different sets of agents uh that that come out of it um and so so that's that's one one aspect and then uh and then there's some things where the irrationality has not been captured yet uh as as realistically in that simulation just the idea that you know sometimes people do something that they know is actually suboptimal for them but they think because of fairness they want to still do it um and so those uh are not yet modeled in in the simulation that we ran but at the same time those rl agents that have neural networks and can try to adjust their behavior to others and so on still much more realistic than anything that uh economists use nowadays which is like often linear models and one step kind of like provably correct formulas and have you have you published the models themselves or the simulation environment yes uh it's actually extremely important you know imagine someone is like i know what is right for everyone and i had a i said it let's all trust it that's of course a terrible idea so you have to open source these models you have to open source all the assumptions you made that went into the simulation and the simulation itself that could be some pretty insidious bugs right if you said oh this is how everyone's going to tax get taxed and then there's a bug and you're like oops so you know a lot of people need to do this and and that's you know one thing i loved about salesforce research too uh and still love that you know for these kinds of important uh human kinds of uh applications um we did open source uh all that model and there are some really exciting ongoing projects now um that you know you can use this also to avoid things like tragedy of the commons where it's like old example of if all the sheep farmers put all their sheep into one field and the field just gets completely destroyed and no one has a field anymore for any sheep so you have to kind of partner up and make sure you don't use your resources too much it's something that i think we're going to hit worldwide in terms of sustainability and deforestation and things like that um and water um so we all have to kind of avoid tragedy of the commons uh in like sort of worldwide yeah given the meta factors that we've talked about um you know with regard to the way revenue models impact uh you know search engine behavior like how does u.com become a viable company uh if it's not going to be ad based and and fall into the same traps that you know we saw with google yeah great question so uh the main goal is to have these applications that we're building actually provide enough value that people would want to pay for them uh you write as one example um you know costs a lot of money to run a large language model um as it writes a blog post for you or an essay like you can pay for that and that's one thing i also think that private ads can be used especially in our private mode where we don't log anything we don't know what's going on at all and we can't really monetize it in any way other than through private ads and what i mean by private ads is just ads that are dependent on the query and that's a luxury that you can have as a as a search engine because people give you an intent of what they want to do um if you're a social network they don't really tell you like i want to buy an air purifier right now um they just talk to their friends and be like oh keep sneezing or coughing and maybe i have dust mites in my home and and then you kind of have to like spy on them if you want to sell them ads to like know what they might want to buy in the future but as a search engine um even if you don't know anything about the user uh if you just look at a query and then you based on that query give an ad i think that is is better and it's kind of what we've seen duckduckgo doing too you know they care about privacy as well um and they have these private ads that basically are not user dependent they're only query dependent and the advertisers don't really know which user is seeing the ad um and you know just it's just basically really only based on the query and so i think that that can be kind of a backup but i really hope that we can build something that's useful enough that people are going to want to pay money for certain things well richard it has been wonderful catching up uh again congrats on the the recent launches and all the amazing work that's gone into uh building what you've built over the couple years the past couple years and uh looking forward to catching up again soon thanks so much great questions and yeah been a pleasure chatting with you you