How Richard Socher built You.com into a $1.5B AI search company

Executive overview

Most AI models hallucinate because they lack accurate, real-time information. The fix is search infrastructure that grounds LLM outputs in verified, up-to-date sources.

Richard Socher pioneered neural networks for NLP and co-invented prompt engineering at Salesforce, then founded You.com in 2020 to bring AI-native search to the world. You.com became the first to embed an LLM in a search engine, then pivoted to B2B infrastructure — the layer that keeps AI accurate.

The fastest path to non-hallucinating AI is better search, not better models alone.

From academic contrarian to startup founder

• Neural networks for NLP were widely rejected in 2010; most early papers were desk-rejected
• Stanford and Andrew Ng's deep learning work confirmed the approach from first principles
• Key insight: hand-crafted linguistic features would never scale — end-to-end neural training would
• Sold MetaMind to Salesforce in 2014 to access the scale academia couldn't provide
• At Salesforce: built the largest protein language model and co-invented prompt engineering
• Published the prompt engineering paper; it was cited and extended by OpenAI researchers
• Left to start You.com because research labs weren't putting the technology into real people's hands

Building You.com and the pivot to enterprise

• Founded in 2020 when conventional wisdom said "search is dead"
• First to integrate an LLM into a search engine (2021) — predating Google's AI Overviews
• Core conviction: getting a direct answer is better than a list of links from first principles
• Pivoted from consumer search to B2B API infrastructure when enterprise demand emerged
• Customers include OpenAI, Amazon, Alibaba, Windsurf, Harvey — using You.com to keep their LLMs accurate and cited
• Rule: follow real revenue, not free users; if companies pay, you've built something of value
• Over 100,000 agents built on the platform, automating real enterprise tasks

Operating principles and AI outlook

• "Better, better, never done" — continuous small improvements compound over time
• Build virtuous data cycles: do a task manually, collect data, automate incrementally
• Avoid all-or-nothing bets (e.g. steering-wheel-less self-driving) — AI needs an on-ramp
• Tesla's model is the right template: product usage generates training data, which automates more
• Think in 2–4 week cycles; the field moves too fast for longer horizons
• Separate real impact (paying customers, working agents) from hype (superintelligence timelines)
• Accelerationist stance: slow down is the wrong call — more, faster