You.com is your AI-powered productivity engine 🎛

CV Deep Dive

Today, we’re talking with Richard Socher, founder and CEO of You.com.

When Richard Socher started You.com in 2020, it was because he believed that the highest impact application of Natural Language Processing (NLP) was search. As the fourth most-cited researcher in NLP and the former Chief Scientist of Salesforce, Richard saw an opportunity at the time to build a better way to search that was much more about summarization and giving answers rather than a list of ten blue links.

You.com is Richard’s second startup. His previous startup MetaMind was acquired by Salesforce in 2016. While at Salesforce, Socher was key to the company’s Einstein AI platform and worked with Bryan McCann, You.com’s co-founder and CTO, on NLP and AI research. In 2021, You.com became the first search engine to embed LLM technology directly into the search experience — an innovation that laid the foundation for their work since. In December 2022, You.com was the first to connect a LLM to live web access for up-to-date responses with citations.

Today, You.com has millions of active users, particularly among people who care deeply about accuracy in the answers they receive. You.com is widely used for knowledge work across fields like finance, biotech, and legal, as well as for students and academic research. You.com specializes in serving as an AI-powered productivity engine for users with complex informational queries, setting it apart from traditional search and other AI chatbots.

In this conversation, Richard walks us through the founding of You.com, its focus on productivity, and his deep background in AI research.

Let’s dive in ⚡️

Read time: 8 mins

Our Chat with Richard 💬

Richard - welcome to Cerebral Valley! First off, just give us a bit about your background and what led you to found You.com?

Hey there! I'm Richard Socher, co-founder and CEO of You.com. Early in my career, I did my PhD at Stanford and brought neural networks into the field of natural language processing in 2010. I was very excited about a single model for all NLP and vision tasks, and my PhD focused on deep learning for vision and NLP. I also co-invented word vectors, contextual vectors, and prompt engineering — all with this elusive goal of training a single model for the hardest NLP tasks. I had a lot of fun in academia as a professor at Stanford for a couple of years, but I also wanted to see this technology in the real world.

So, I started a company, MetaMind, which was acquired by Salesforce. There, I became Chief Scientist and built out the research team, and we were able to ship these technologies into real products, which was fun. In 2020, having seen that we could train a single model for all of NLP now and ask it any kind of question, I realized that the biggest application of NLP—search—hadn't really changed in many years.

So, I started You.com together with Bryan McCann, the first author of the decaNLP paper that introduced prompt engineering. In 2021, we were the first to connect LLMs into a search engine context.

How would you describe You.com today to an uninitiated consumer or knowledge worker curious about AI platforms?

Great question. It actually has evolved over time, and the way I would describe it now is as a productivity engine that helps you with search and work. It does the searching and working for you, can program for you, handle many queries, and summarize content. If you're a knowledge worker or student and you want amazing answers for more complex questions, that's where we really shine.

Everyone, including Google, can now give you decent answers to short questions like "When was Obama born?" It's hard to do that 10x better. But if you have a complex question like "Explain the whole background and consequences of the Peloponnesian War" and you want a full essay with images, multiple citations, and no hallucinations, that's when You.com really shines.

We see companies like biotech firms, university students, researchers, and knowledge workers in legal and finance sectors—people who really care about accuracy—using You.com for productive work. After comparing us to others, they usually stick with us and we have extremely high retention. That's why we're transitioning more and more away from being just a chat assistant to becoming a productivity engine that helps you with chat and search, but also with productive work.

Anytime you have a complex informational query, we can support you. Whether you're trying to understand cancer treatments, solve a math or physics problem that requires coding, or just find a quick answer and make sure it's correct, we can help. From simple questions like "What's the weather?" or "What's the stock price?" or “What’s a good restaurant?” to highly complex questions that require programming, math, or logical reasoning to healthcare questions or finance questions, and so on, we can answer those for you.

You.com was the first company to actually combine the power of LLMs with web search. Could you share a little bit about those early days, and the insights that led you to almost preempt the rise of LLM-powered search?

We actually had LLMs even before ChatGPT came out. We had a whole host of different LLM apps or widgets within a search engine results page, including LLMs that would write code or essays for you when you had those kinds of queries. Initially, when we innovated too far away from a Google-like experience, the majority of the feedback was, "I'm just so used to Google. I don't want it to be too different." But that all changed with ChatGPT and its massive, visceral experience of conversational interaction.

This unlock helped people realize they could search in a different way, moving away from a list of blue links. We quickly adapted, and two or three weeks after ChatGPT came out, we connected the first LLM to the web and included citations. As amazing as LLMs are, they do hallucinate and can't be trained with constantly updating news articles. They often don't specify where their facts come from.

We have the most accurate citations. We've seen many competitors copy our citation feature, but when you double click into it and run some benchmarks and analyses, half of their citations are just random links behind unrelated sentences. Citation logic is non-trivial. To stay accurate and up-to-date, we've built our own web index and crawler to provide the best results to feed into the LLM. Surprisingly, we are even more accurate than ChatGPT-4 premium.

We also allow everyone on a YouPro subscription to use all the other models. We recently added the new Google Flash model and GPT-4o on the day it came out. We're trying to help everyone stay up-to-date in this evolving space by ensuring they always have access to the latest and best models and, in each case, they will be connected to the web.

We also just added a Custom Assistants feature last week. You can now take one of these top AI models, give it specific instructions, and add an internet connection. For example, imagine creating a kid-friendly AI chatbot, or an interactive trivia chatbot to quiz you on world geography, or an assistant to help you come up with catchy headlines for your articles. Our users have already made thousands of these Custom Assistants.

Walk us through the evolution of You.com, from those early days as LLM-powered search through to chat and now productivity. How would you describe that arc?

The arc for us has been focusing on technology and accuracy to provide amazing answers. We're now seeing that we need to focus more on marketing. In the beginning, there were still quite a lot of hallucinations, and it didn't feel right to claim we could replace your Google experience entirely. Now, we have weather, stock market, and map widgets integrated into the chat. We also offer links similar to Google's if you want them.

Most importantly, our chat-first experience is now very accurate, with almost no hallucinations. Based on various independent benchmark studies, we know we are more accurate than any other chat engine. I'm really proud of the experience we offer now. The biggest transition for us is getting ready to scale this up and help people be more productive with their knowledge work through amazing answers.

What would you say is unique about You.com’s technical approach relative to some of the other answer-driven AI applications available today?

At a high level, it's an organizational decision to spend more resources on technology and accuracy rather than marketing. This means taking a scientific approach to our product. We're still working with cutting-edge AI. In several cases, we've shipped ideas on how to make a model better or orchestrate things more effectively, and then a few weeks later, we see a paper from Stanford publishing that same idea. In the industry, especially in a small startup, you don't spend as much time writing up a detailed paper with all the careful comparisons and analyses. You realize it works better, is cost-efficient, and you ship it.

This is still very much research. There are over a dozen different AI sub-modules involved in providing a quick answer. For example, you need a classifier to determine whether you should search the web or not. If someone asks for a poem for their anniversary, you don't need a citation for each line of that poem. But if it's a factual question in the news, you need citations, ideally multiple ones to present different viewpoints if there are contrasting facts.

We've rethought search to build an index that is designed for LLMs to consume. Instead of short snippets like you get from Google or Bing, we provide long snippets that are too much for a human to consume but perfect for an LLM. This way, the LLM can find the right information and provide an amazing answer.

We also offer an API for other LLMs, allowing them to give more up-to-date answers by connecting to the web. This API empowers other companies and chatbots to provide accurate information by consuming the extensive snippets we generate.

We're now at a point where image and video models have reached high fidelity and are getting more impressive week by week. How does multimodal factor into the plans you have for You.com as this productivity-oriented tool?

I love this question. We were actually the first to enhance chat models with the convenience of search engine widgets. For example, if you ask for the weather, we can give you a whole paragraph about it, but instead, we provide a little widget showing you the weather by day and hour. If you ask for a stock price, the best answer is not a lengthy paragraph detailing every time step of the day, month, or week. It's best to visualize it with a graph and a stock ticker, which is what You.com does.

We were the first to integrate more and more of these widgets into chat, enhancing the overall chat experience. We also added image generation. If the best answer to "What's an image of a dancing bear on a balloon?" is just to give you that image, we provide it instead of telling you how to create it on the web. Initially, some people thought this was unfocused for a search engine or chat engine, but now everyone, including Microsoft and other small players, has copied this idea. The concept of providing amazing, actionable answers has become more widespread.

Our Research mode writes very carefully crafted analyses and essays for you. Our Creative mode generates images. Over time, we plan to add even more modalities to help people in marketing, sales, outreach, and other areas to be more productive and possibly create multimodal experiences.

What do you foresee as the hardest technical challenge today about getting You.com to be an all-encompassing productivity suite powered by accurate answers?

Today, it's about bringing in more and more useful datasets and continually improving the search index. We aim to make all these different subsystems extremely fast and super accurate. Eventually, we want to have a single premium mode. Right now, we have Genius and Research modes, and if you know which one to use, they are better than anything out there. But sometimes, people aren't sure which one to use. So, we're working on merging them and recommending switches to different modes when we understand that a question is best answered by a specific mode.

Concretely, Genius mode will program for you and handle multi-step reasoning. For example, you might need three things done, rethink your query, get an image output, and then get a programmatic response. Research mode will go to many dozens of websites and summarize the information with many citations—up to 10, 20, or 30 for a single answer. These modes tackle different use cases really well.

The technical challenge is always knowing which use case a user has in mind when they make a query. To be honest, there's a set of queries where it's just impossible to know. Sometimes, I put the same query into different modes to get slightly different answers and viewpoints or a more in-depth answer after the first query. In some cases, it is knowable which mode to use.

Lastly, what do the next 6-12 months look like for You.com - any particular areas of focus? And how are you thinking about building the team to compete in such an intense arena?

The big goal is to keep our hockey stick revenue growth going throughout the year and grow our user numbers. But most importantly, we want to help people become more productive and develop not just an answer engine, but a productivity engine that works, searches, and programs for you. We're really excited to collaborate more and more with companies, developers, knowledge workers, and universities to help their user base be more productive, learn faster, research faster, and get more work done.

Team-wise, every person brings their DNA into a small company, and my DNA is heavily research-focused. For a startup, that means making decisions based on facts. You build trust with people by bringing the right facts, experiments, and datasets to the experience. Additionally, it's important to be kind in how we communicate and interact with each other and with our users. Trust, facts, and kindness are the main values that drive a lot of decision-making and the culture at You.com.

Conclusion

To stay up to date on the latest with You.com, follow them on X and learn more about them at You.com.

Read our past few Deep Dives below:

• Zilliz Cloud is powering vector search for the enterprise 🔋

• Kolena aims to redefine quality standards in ML 🎛

• Pinecone Serverless is turbo-charging your time-to-prod ⚡️

• Cleanlab is automating data curation at scale 🌐

If you would like us to ‘Deep Dive’ a founder, team or product launch, please reply to this email (newsletter@cerebralvalley.ai) or DM us on Twitter or LinkedIn.

Join Slack | All Events | Jobs