EO StudioHow Prompt Engineering Inventor Built $1.5B in 3 Years | You.com, Richard Socher
CHAPTERS
- 0:00 – 0:30
Betting on neural nets for language (and enduring early rejection)
Socher recalls his 2010 conviction that neural networks would work for NLP—an unpopular stance that led to paper rejections and skepticism from top schools. He frames this contrarian streak as a throughline that later enabled You.com despite people declaring “search is dead.”
- •Early NLP community resistance to neural networks
- •Paper rejections and institutional skepticism (MIT/Berkeley)
- •“Prompt engineering” origins and desire to democratize access
- •2020 launch context: widespread belief that search couldn’t be disrupted
- 0:30 – 1:01
Meaning over popularity: founder mindset and first-principles conviction
He explains that he optimizes for what’s meaningful rather than what’s fashionable. The ability to hold a first-principles belief through repeated rejection becomes a core entrepreneurial advantage.
- •Choosing meaningful problems vs. trendy ones
- •First-principles reasoning as a source of conviction
- •Perseverance through rejection as a requirement
- •Personal introduction: CEO of You.com, partner at AI X Ventures
- 1:01 – 1:31
You.com’s thesis: AI search infrastructure that keeps LMs accurate
Socher defines You.com as AI search infrastructure that feeds language models up-to-date, verifiable information. He emphasizes citations and grounding as the path to reducing hallucinations, and notes enterprise adoption via APIs.
- •Grounding LMs with search to reduce hallucinations
- •API infrastructure for freshness, accuracy, and citations
- •Agents and users search on You.com vs. traditional search
- •Enterprise customers cited: OpenAI, Amazon, Alibaba, etc.
- 1:31 – 2:31
From Germany to computational linguistics: languages + math as a life direction
He traces his background—German upbringing, love of multiple languages, and simultaneous pull toward mathematics. This intersection leads him to study computational linguistics despite it being niche and viewed as impractical.
- •Early fascination with meaning and “life questions”
- •Love of natural languages (English, French, Chinese)
- •Tension and eventual intersection of math and language via computing
- •Choosing a niche field in 2003 despite doubts from others
- 2:31 – 3:21
Why language is the operating system of intelligence
Socher argues language is a primary manifestation of human intelligence and a driver of civilizational progress. He draws a parallel between written language historically and AI adoption today as a competitive differentiator.
- •Language as a core expression of intelligence
- •Written language as a historical accelerant for civilizations
- •Analogy: societies not using AI risk falling behind
- •Engineering intelligence as a route to understanding it
- 3:21 – 4:10
Neural networks as a scalable alternative to brittle feature engineering
He introduces the intuition behind artificial neural networks and why they fit language problems. Hearing Andrew Ng on deep learning reinforces his view that hand-crafted linguistic features wouldn’t scale to harder tasks like translation.
- •Biological neurons inspire artificial neural networks
- •Deep learning inspiration from computer vision
- •Limits of feature engineering in NLP (e.g., sentiment features)
- •Goal to unify methods across language tasks
- 4:10 – 5:11
Scaling the breakthrough: from research ingredients to real-world impact
After his PhD, Socher felt the key ingredients were clear—large neural nets plus lots of data—and the next frontier was scaling them in production. He reflects that academia lacked resources and that even early industry efforts were underfunded relative to the opportunity.
- •Post-PhD realization: apply proven ingredients at scale
- •End-to-end trainable networks + large datasets as the key leap
- •Academia’s resource constraints for true scaling
- •Retrospective: should have raised far more capital to scale faster
- 5:11 – 6:12
MetaMind and the Salesforce chapter: platforms, focus, and distribution
He describes founding MetaMind as an AI platform to make training neural networks easy, then recognizing that large-scale distribution required a major sales force. At Salesforce, the work broadened into major applied research, including biology-focused language models.
- •MetaMind as an AI training platform company
- •Challenge: selling requires tight niche focus
- •Acquisition logic: larger org can amplify impact via distribution
- •Research breadth at Salesforce, including protein language models
- 6:12 – 6:43
Inventing and popularizing prompt engineering: one model, many tasks
Socher outlines how they trained a single network to produce many kinds of answers—an early framing of prompt engineering. He notes the idea’s influence through citations and extensions by researchers associated with OpenAI.
- •Prompt engineering framed as enabling multi-purpose answering
- •Publishing and broader research uptake
- •Bridging research ideas into tools for real people
- •Motivation to move beyond ‘research lab only’ availability
- 6:43 – 7:43
Founding You.com to challenge Google: reinventing search with LMs
He argues Google’s ad-driven monopoly reduced incentives to rethink search fundamentally. You.com was created to deliver direct answers with LMs—years before big incumbents shipped similar experiences—based on the principle that answers beat blue links for many queries.
- •Perceived stagnation in Google’s search paradigm
- •You.com as a new-company bet to change information access
- •First-principles UX: answers vs. ten tabs of links
- •Early LM-in-search deployment (2021) ahead of incumbents
- 7:43 – 9:14
The enterprise pivot: follow revenue and build what companies pay for
Socher explains that a critical shift was focusing on enterprise customers who needed high-quality answers over proprietary data. The core value stayed the same (answers), but monetization changed—guided by paying demand rather than hype.
- •Pivot from free consumer usage to enterprise needs
- •Custom datasets + accuracy requirements drive willingness to pay
- •“Follow the revenue” as product strategy
- •Value validation: payment signals real utility
- 9:14 – 11:15
Moving fast in AI: grounding, agents, hype vs. reality, and virtuous data loops
He advocates accelerating development while staying clear-eyed about hype and timelines. He emphasizes two-to-four-week iteration cycles, building applications that generate virtuous data loops, and using partial automation pathways (like a steering wheel) to reach reliability.
- •Acceleration mindset vs. calls to slow down
- •Customers building large numbers of agents for real work
- •2–4 week iteration cycles in fast-moving AI markets
- •Virtuous data cycles: manual → data → improvement → automation
- •Self-driving analogy: ship usable systems before perfection
- 11:15 – 12:03
A builder’s ethos: continuous improvement and excitement for AI’s full stack
Socher closes with an optimism grounded in practice—improving systems continuously while exploring both frontier research and practical deployment. His motto captures the operating cadence: relentless iteration toward better outcomes for users and companies.
- •Motto: “Better, better, never done.”
- •Continuous improvement across self, company, and processes
- •Excitement spanning superintelligence research to real deployment
- •Focus on getting AI into more hands to create tangible value
