Orchestrate all the Things - You.com raises $50M to lead AI for Knowledge Workers. Featuring you.com Co-Founder Richard Socher

Episode Date: September 4, 2024

You.com showcases the state of AI today The story of you.com is multi-faceted and telling in many ways. You.com was founded in 2020 by Richard Socher, one of the leading NLP (Natural Language Processing) researchers in the world, to offer a better search experience to users and compete with Google. With a startup exit and a Chief Data Scientist stint at Salesforce, Socher got the experience, network and backing he needed to pursue his long-time ambition of taking on Google. That's something few people have tried, with moderate success. Socher diagnosed early enough that the way to success is by carving a niche for you.com. You.com focuses on serving knowledge workers in "complex informational / action searches": elaborate queries, and queries that are really about accomplishing a task, respectively. In 2022, in the pre-ChatGPT era, Socher set out a course for you.com based on AI, apps, privacy, and personalization. In 2024, you.com is staying the course, but a few things have changed. In the GenAI era the competition is growing, and borrowing pages from you.com’s book. Language model providers such as OpenAI and Anthropic now offer services similar to you.com. Upstarts such as perplexity.ai have sprung up, and Google itself is embracing the AI approach to search. You.com is making progress too. Since launching in November 2021, you.com has served 1 billion queries and has millions of active users, including from Fortune 500. The company's ARR has grown by 500% since January 2024. Today, you.com announced a $50 million Series B funding round, as well as a new team plan called Multiplayer AI. We caught up with Socher, talked about the news, and took you.com for a spin. Article published on Orchestrate all the Things: https://linkeddataorchestration.com/2024/09/04/you-com-raises-50m-to-lead-ai-for-knowledge-workers/

Welcome to Orchestrate All the Things. I'm George Anatiotis and we'll be connecting the dots together. Stories about technology, data, AI and media and how they flow into each other, shaping our lives. The story of u.com is multifaceted and telling in many ways. u.com was founded in 2020 by Richard Schroescher, one of the leading natural language processing researchers in the world, to offer a better search experience to users and compete with Google. With a startup exit and a chief data scientist stint at Salesforce, Suchar got the experience, network and backing he needed to pursue his long-time ambition of taking on Google. That's something few people have tried with moderate success. Suchar diagnosed early enough that the way to success is by carving a niche for u.com.

u.com focuses on serving knowledge workers in complex informational and action searches, elaborate queries and queries that are really about accomplishing a task, respectively. In 2022, in the pre-ChatGPT era, Searchers set out a course for u.com based on AI, apps, privacy and personalization. In 2024, u.com is staying the course, but a few things have changed. In the GenAI era, the competition is borrowing pages from u.com's book. Not only have appstarts such as Perplexity AI Sprang Up, but Google itself is embracing the AI approach to search. u.com is making progress too. Since launching in November 2021, u.com has served 1 billion queries and has millions of active users,

including from Fortune 500. The company's annual recurring revenue has grown by 500% since January 2024, and today U.com announced the 50 million Series B funding round, as well as a new team plan called Multiplayer AI. We caught up with Sotcher, discussed about the news, and took U.com for a spin. I hope you will enjoy this. If you like my work and orchestrate all the things, you can subscribe to my podcast, available on all major platforms, my self-published newsletter, also syndicated on Substack,

Hackernan, Medium and Dzone, or follow and orchestrate all the things on your social media of choice. Hi everyone, I'm Richard Sosier. I'm the CEO here at you.com. And we build an accurate productivity engine for knowledge workers. Okay, that was really, really brief, which is good. So actually, the the occasion uh for uh for having this conversation today is that you're about to announce some news well first you're getting a new round of funding and second you're

announcing somewhat of a switch in your business model so i thought i would start with the funding news. And you have quite a list of backers there. The funding is led by Georgian and you also have SPVA, which is the former SoftBank Ventures Asia, Salesforce Ventures, NVIDIA, DuckDuckGo and Day One Ventures. And I thought I would just ask you to say just a few words for each of these backers, which I'm sure you have worked with probably personally. So what would you say was the thing that got each one of them interested in Buckingham and just to say in advance, I find NVIDIA and DuckDuckGo out of this list somewhat unexpected

and particularly interesting it's for a different reason yeah we're very excited for all our investors georgian has been an incredible partner already in our journey over the last few weeks and they have a very deep technical team that's helping with go to market as well i think nvidia has seen kind of how we have innovated and out innovated many in the field you know we have brought large language models into a search engine context a year before ChatGP came out with essay writing, LLN widgets, with programming widgets that would write code for you if that's what you look for. So we've been at that forefront for a while.

We're the first to connect the search engine to a chat window so that the chat and LLN language models can have citations. And I think NVIDIA saw that innovation and then saw the combination of now that we have innovation and revenue and are focused on both, they're very excited to join us and we're very excited to have them. There's a lot more collaborations that i hope we get to announce in the future with them okay um i also found uh the fact that you got uh support from dak dak go interesting because first i don't think i've seen them uh support any other venture previously to yours and also that because of, well, in theory at least, the kind of competition, right?

So you must be doing something right if your competition is bugging you. We are not quite in the search engine market anymore. There are a lot of questions that users ask that are quite fast and simple, navigational, where a chat answer engine doesn't quite shine. For instance, if you just ask, how old is Obama? When was Greece founded? What's the weather tomorrow? These kinds of questions you can get a very quick answer for,

and it's hard to use LMs to make that 10x better. Maybe make it one or two x better for simple questions. What we found is that questions that are really complicated, that knowledge workers are asking in fintech, in marketing, in sales, for preparing a meeting, for writing a marketing campaign, for doing deep research in biotech or at a hedge fund. That is where we can actually provide a 10x better experience. Why? Because we have deep search expertise and deep LM expertise. So we're very excited to bring that technology to other companies in the business context.

And when we do that, there are two lines of business. One is enterprise site licenses, where an entire company can make all the employees more productive. We really lean in. For larger companies, we also offer a consumption-based pricing model so they don't pay for hype. They pay for actual usage with their employees.

And we also offer APIs. And that latter one is particularly interested for DuckDuckGo. We share their values around privacy. We're very excited to work with them. And we're very excited uh to work with them and and we're very excited with Dave and Esther okay uh great i that's one answer that is a great segue to ask you something that was was in my list anyway so you basically very quickly described the kind of pivot that you had from being like a so to say direct competitor to the googles of the world so a straight up search engine to switching to a more sophisticated question answering system that targets a specific niche so knowledge workers and you have been doing that for the last couple of years, if I'm not mistaken anyway.

But now you're making this switch, let's say, to what you called the multiplayer AI thing. So I'm wondering, besides the obvious, which you already sort of sketched out, like giving enterprises like one access point to multiple employees instead of having like each person sort of bring their own AI, let's say. Does that also mean that you're going to support features that enterprises are typically looking for like security and access control and all of that? That is exactly right. Yeah, we are getting single sign-on. We are just in, we've finished complying

with all the SOC 2 compliance things and are just waiting now a few more weeks to be in this official waiting period of SOC 2. And we're offering teams and enterprises ways to collaborate around AI to start projects, have massive context windows so that a custom agent that you built will know all the company context that it needs to know to be actually a useful knowledgeable agent for your workflows okay so that also includes collaboration not just you know here's many accounts go out and play but people can

actually work with each other and presumably what will what will they exactly be able to do will they be able to like collaboratively offer workflows or I don't know prompts or what? It would be easiest to just share my screen but I haven't used this video software before but at a high level what you can think about is that you can share your chats, other people can continue those chats, you can upload files to a project but then everyone will have access to and can answer and ask questions about and you can comment and continue the workflow that someone else may have started and you all have a shared knowledge base

as you're doing knowledge work okay yeah so yes that definitely sounds like a collaboration so like a shared workspace and many different artifacts i presume that people can i can work on on that workspace that's right and if you want, I can share with you some screenshots. Yes, that would be great indeed. So I was wondering, obviously, in order to get all those investors on board, you must have shown them your business plan and your growth already, which, by the way, seems to be coming along pretty nicely. You have some Fortune 500 clients and your annual recurring revenue, you said,

has grown by 500% since the beginning of the year. So it seems to be going well, in other words. What I'm wondering, however, is, and that's something that others have have come across before you obviously it's one thing to sell b2c let's say and it's another thing to sell b2b so enterprise sales and I'm wondering if your sales is ready for that basically because the requirements are different yeah we've actually hired some amazing folks in the team for enterprise sales, and it's just been really wonderful seeing them in action. And I had been at Salesforce myself in the past, so I wasn't doing sales there as the chief scientist,

but I've observed and learned a lot about very effective sales teams. And so, yeah, we're excited to have expanded the team. We'll make some announcements around that probably in a few weeks as well. And it's a new muscle, but it seems like there's a lot of market pull. What we actually found is that a lot of companies had tried to build their own prototypes, the really big ones. Some just don't have the resources and we can help them build the first version. Some larger customers had actually built prototypes similar to u.com where they tested out the competition. And then we're just dismayed by how inaccurate that is. You can quickly in Gen AI pack up a prototype, but it's very hard to build something that really works at scale every day, very accurately, has strong search, strong LM orchestration, and puts it all together in a package that you can rely on in your day-to-day work. And so a lot of those companies that have tried in the past with other competitors or their own prototypes are then coming to us and seeing the difference.

And so we do have what is called a hybrid of product-led growth, where people can try U.com out and be wowed by its accuracy and then also set up okay so i know because i've actually played a little bit with uh with the product myself that you have different different modes in your in your user interface you have something called smart mode called smart another one called genius and you have another mode called research. Would you like to just quickly explain what each one does and how it's different? Yeah, that's a great question and maybe at some point it will get merged, but you can think of these as very different levels of capabilities. Smart mode will give you a quick, smart, accurate answer. It won't go into too much detail. It won't give you a huge research report.

But it will be fast. It will be multimodal. It will have images and things like that. Genius is for much more complicated questions. And I wish I could share my screen with you. Do you think that will work in your video? Yes, sure. All right, so let me. Let me try this one. All right, amazing. So can you see? Yeah. So here is an example of what a genius mode can do. And this is an easier to understand example. It's not deep in someone's work context, but

it shows you the power of what's possible in genius mode. What should be the initial investment in a compounding interest index fund for my one-year-old child to ensure that it fully covers the cost of her Stanford creation by the time she can complete the side-score? Imagine the complexity of that answer and what is needed to really get this right. And what genius mode will realize is that it will have to do multiple different queries and put all of these facts together. It will have to do multiple different queries and put all of these facts together. It will have to look for Stanford tuition, the expected increase in that tuition,

expected annual return of an index fund, number of years until a child goes to college, which is wonderful. Also, no pressure for the kid. And so you can see here that this is exactly what it does. It goes out on the web and searches for the cost of Stanford tuition. It has nice clean citations for everything. Then it goes out and looks for the expected annual increase of that tuition based on historical trends. Then it will go out and look for annual S&P 500 returns, which you've never mentioned, right? You just said some compounding interest fund,

realizes that that's right around 10%, and then realizes that we'll probably have another 17 years or so to do this. And then instead of making up a bunch of numbers, which a lot of the competition does, or not being helpful, it actually takes all of those facts and does the right math by programming all of those facts into code and then actually executing that code

and then giving you an answer. That is just light years ahead of what you will see anywhere else. So that's the answer. And then lastly, well, I guess I'll share with you also research mode. Research mode will write very careful research reports for you. Sometimes you want a quick answer. If you know you want a very carefully researched, longer report on a subject

that will do literally 10 plus queries, rephrase your input and your question into multiple different searches and then find you the right answer, then research mode is for you and you're realizing that you call the modes nowadays people call them agents right these are agents that have the ability to search when they need to and perform that search action they can code when they need to and then decide also to run that code they can ask you follow-up questions too so these agents are quite powerful and there's

all kinds of things you can do including for instance have an interactive session with such an agent that will help you for instance write seo optimized content all right well another All right. Well, another thing that struck me about how u.com works and also the user interface was that it seems like you're also acting as your own in-house language models that you've trained, but you also offer, in addition to that, you also offer access to other language models such as, you know, the GPT family or CLODE or a bunch of others, really. And at first I was, to be honest, I was having trouble understanding like why would you want to do that? But then again, reading through some of your some of some of the things you've said in those interviews, it struck me that it seems like what you're effectively doing is something like a mixture of experts approach.

So you said you're routing each user query to one or potentially more of those models and then synthesizing the results that you get out of those according well to obviously use some some kind of heuristics to determine which model is the best to address its query to and i was wondering if indeed my impression is correct if this is what you do and by extension this also, since you're doing it anyway, I guess you figured that it may make sense to offer access to those models to your users as well. And then the actual challenges of doing that, because well, it's an approach that many, many users,

many of us also use, you know, in real life, but actually engineering that and at the scale that you need to do that to offer a service like this, I'm pretty sure it's no easy feat. That's exactly right. We believe that u.com part of its job is to help people and businesses to be future proof. Gen AI moves very, very quickly.

And we've noticed that a lot of our customers feel like they're locked in when they sign a big annual contract with one of the foundational model providers. And then a month later or a few weeks later another better lm comes out and now they're stuck with that original foundational model so we decided every crisis and opportunity we want to enable our customers to basically be always at the state of the art by giving them direct access to these models and then on the plus, we also in our own modes and these powerful research genius agents, we are able to essentially orchestrate these LMs and use a different LM for different kinds of questions to give the most accurate answers.

For instance, now Claude from Anthropic has actually outperformed, uh, opening eye when it comes to programming. And so we can route these different kinds of questions that require programming to a different element. Okay and well another challenge that comes with dealing with that many models and actually you know even with the same model that as you said all models keep evolving over time and some of them pretty fast as well and so what people are finding out, among other things, is that the different prompts that they use to communicate with those models and they have people have actually invested lots of time and effort into crafting and fine tuning those prompts.

So what they're finding out is that oftentimes when new model versions are released, these prompts start working well, or sometimes they don't start working at all. And I'm sure that's a problem you must have come across as well. And so I'm wondering what your solution to this problem is and how do you manage your prompt engineering? Yeah, so I think there are indeed, there's a lot of complexity and the way we're thinking about this is that you might not have only one prompt, part of our orchestration layer is actually to do what is called dynamic prompting. So and of course on the fundamental level we have to do this for every query. For every query that comes in, we need to understand whether you want to actually ask a factual question or whether you want to have the model actually hallucinate something. So, for example, if I ask, I'm about to go into a big meeting with Nike, and I want to ask

about their executive team and the business and the updates and how much revenue they're making and things like that, then I don't want it to hallucinate and I want to have very accurate citations for every fact that I take into this meeting. Now, if I then afterwards ask for writing a poem to my wife, I don't really need a citation for every line of that poem. And each will require different things. And you have already this dynamic prompting where you allow the model to decide to look for facts on the internet and then feed those into the prompt.

All of those are things that we need to accommodate for and we basically abstract away. And it's a little bit like self-driving cars in the sense that it's very easy to build this demo that will drive on the highway and it's a beautiful wide open highway, good weather in California. And you're like, look, we have a self-driving car, but then it takes many more years to actually get that self-driving car to be in the city, driving every day without any issues. And it's similar to doing knowledge work and providing very accurate answers. It's really a never-ending type of work, but we believe, we've been added longer than anyone else in the world, we've now gotten to the highest point of accuracy.

And well, speaking of accuracy, which is a point that you also mentioned in the updates that you're about to release, and I understand that it's central in what you do anyhow and so another question I had was precisely around accuracy and so how what kind of techniques do you apply to be able to uh to increase accuracy in the replies that you have some people are using uh various uh uh various various variances on rug other people are usingAC, other people are using context windows, other people are using combination of those and you obviously also mentioned already that you work with workflows and what I just said, if none of the agents in your workflows is 100% accurate, then obviously the end result is not going to be accurate as well.

So you need to increase the accuracy on each and every step until you manage to have a pipeline that will give you the right result. So what kind of techniques do you apply to to to increase accuracy? I could talk about that for hour and put my old professorial back on from Stanford days and answer that question in all its detail. But at a high level, there are several dozens of modules. There's intent classification when the user asks a question. There's query rewriting.

Imagine the user asks a question. The answer includes Salesforce. And then you ask, oh, who's their CEO? If you just send who's their CEO to your search back end it's not going to do anything useful so you need to read phrase based on the context of the whole conversation the kind of query you want to send to your search back end then of course making search accurate is a non-trivial engineering task and as you can see now from some of our investors we have folks that have been very impressed by our

search capabilities and the technology we've built there then you want this whole thing to be up to date you need to do dynamic prompting you need to do LLM orchestration so there's a lot of different aspects that will go into making this technology. Okay, so I understand we need to wrap up. So I'll just give you one final question. And that's again on accuracy. So as it happened, I was writing an article this week and I wanted to do a very quick piece of research for a quote. It's a quote attributed to Umberto Eco which goes like, there's no news in August. So having the plan for this conversation, the quote goes, there is no news in August

and it's attributed to Umberto Eco so I thought it would make for a nice quick little test if I tried to do like a prompt around this and test it with u.com and perplexity so I did that and I tested with both and I have to say that both replies were different, both got some aspects right, but also both got some aspects very, very wrong. So that made me wonder, okay, so how do you actually objectively evaluate such a thing as accuracy with competing LLMs or in that case, services based off of LLMs? I know that's not a very easy question to answer but well if if you're making

accuracies as a central point you have to at least have given it some some thought yeah so we actually have done a bunch of uh let me uh let me put this uh this question into printv.com here too. So we have done a lot of extensive evaluations. You can actually go on blog.v.com and see how we evaluate it and do user studies. We ask users, what is the answer you prefer? Do the research and verify which answer is more accurate. And we're much more accurate than either ChatGVT or some of our smaller perplexing competitors. Thanks for sticking around.

For more stories like this, check the link in bio and follow link data orchestration.