How Prompt Engineering Inventor Built $1.5B in 3 Years | You.com, Richard Socher
These notes were generated by AI and may contain inaccuracies.
Had the crazy idea in 2010 to use neural networks for natural language processing, which was very controversial. Many in the field said neural networks never work and won't ever work. Most papers got rejected because people hated neural nets for NLP. MIT and especially Berkeley hated neural networks for NLP. Invented prompt engineering and thought people should get access to it. Felt like someone's got to do it. In 2020 when starting You.com, many said search is dead, nothing you can do.
I don't really generally care about what's popular. I just care about what's meaningful.
Once you love an idea and it makes sense from first principles, you need belief to make it prevail through rejection and keep going.
Hi everyone, I'm Richard Socher. Founder and CEO of You.com and founder and partner at AI Accenture. You.com is AI search infrastructure. To make an LM not hallucinate, need good search infrastructure to inform the LM. People search on Google; AI, LLMs, and agents search on You.com. Companies like OpenAI, Amazon, Alibaba, Telegraph, Windsurf, Harvey use You.com's API infrastructure to make their LLMs up-to-date, accurate, with citations to verify facts.
Amazing to become a unicorn, but feels like what's next?
Originally from Germany. Thought a lot about the meaning of life in high school. Loved natural languages: studied English, French, Chinese, and math. Math and languages intersect in computers using math to understand language. Decided to study linguistic computer science in 2003. Not famous or popular in Germany; niche subject few studied. Dad thought it didn't sound useful.
I don't really generally care about what's popular. I just care about what's meaningful.
Felt getting it to work on research side would have amazing impact. Language is the most interesting manifestation of human intelligence. Civilizations that created written language advanced; those that didn't fell behind. Civilizations not using AI now are falling behind. Meaningful to understand language to understand humans and create intelligence. Anything we create, we understand better afterward.
Your brain and mine are jam-packed with neurons tightly connected and talking. In computers, build artificial neural network or supervised learning algorithm simulating neurons connected and talking.
Fortunate at Stanford to hear Andrew Ng talk about deep learning and neural nets for computer vision. Made sense from first principles neural networks are right. Research was about feature engineering; e.g., in sentiment analysis, positive words, negation—linguists came up with features. Wouldn't scale to translation. Wanted to unify the field, leading to prompt engineering.
After PhD, had main ingredients: large neural networks, lots of data. Showed in research papers. Time to scale up for real applications. Papers in last years improved things, but main ideas of end-to-end trainable neural networks on large datasets pushed field forward. Academia lacks resources to scale. Should have scaled more; raised 10-20 million but should have raised 200 million or billion. Clear to scale in industry; technology ready to move from academia to real world.
Maximize impact: main ideas researched, time to apply in real world, so startup made sense. Graduated 2014, started MetaMind.
MetaMind was AI platform making it easy to train neural networks. Started selling to Salesforce. To sell, focus on one niche, but built powerful platform. In hands of large sales force like Salesforce, much more impact. Impact within Salesforce much larger.
Thought would be happy at Salesforce doing research: not just prompt engineering but AR language model for proteins and biology.
Invented prompt engineering: trained one neural network giving different kinds of answers. People should get access. Published paper cited by OpenAI, Alec Radford, Ilya Sutskever; they extended it. But they're research labs. Needed to bring to real people. Google not doing anything as monopoly making money on ads, no need for interesting search modifications.
Felt someone's got to do it. Started You.com as new company to become better way of finding information online. First to put LM into search engine. Did things like Google Gemini "people also ask" and AI answers but in 2021.
Research background: PhD is to create new ideas and models. Think of useful new ideas. Obvious better to get answer from LM than list of blue links requiring clicking 10 tabs. From first principles, better to get answer than links that may or may not have it.
Biggest thing: pivot into enterprise, really good focus. Many realize need AI, but experts know to make AI accurate and LM not hallucinate, need good search infrastructure.
Built infrastructure layer since 2022. Companies want underlying infrastructure for own solutions over custom datasets, willing to pay.
Pivots: gave people answers, now selling answers differently. Follow revenue: people wanting free vs. paying for good answers over custom data. Follow people that pay, real revenue, not hype.
If you build something companies are willing to pay for, you've built something of value.
Some say slow down; should accelerate more. Hard to navigate AI: real impact (customers built over 100,000 agents automating real tasks, paying because valuable) vs. hype (superintelligence timelines off). Field moves quickly; think 2-4 week cycles.
Important: right applications with virtuous data cycles—do manually, collect data, improve decision process, automate when good enough.
Didn't want to invest in self-driving companies saying no steering wheel needed. Can't sell until perfect. AI not perfect right away; humans make driving mistakes, AI will too. Good to have steering wheel.
Companies getting to full self-driving: clever like Tesla—buy car, use product, create training data, AI improves to automate.
Small continuous improvements motivate. Motto: better, better, never done. Always improve self, company, processes.
Overall: excitement. Love AI from foundational research on superintelligence to making it work now for companies and people.
Keep EO in your library
Save the videos and channels worth coming back to, and find them again in one place.
