AI summary by TwinMind
Snowflake CEO Sridhar Ramaswamy on Using Data to Create Simple, Reliable AI for Businesses
Sequoia Capital
59m 30s
Background and Introduction
- Sridhar Ramaswamy: CEO of Snowflake, computer scientist, former Google executive (search ads business)
- Grew Google's ads business from $1.5B to $120B revenue over decade
- Founded Neva (AI-driven search engine) before Snowflake acquisition
- Snowflake: AI data cloud platform, 10,000+ customers, $2.6B revenue last year
- Core thesis: cloud computing platform centered on data superior to generic cloud for enterprise customers
Enterprise AI Landscape Overview
- High awareness level among customers - people understand AI's transformational potential
- Thousands of customers implementing AI solutions across spectrum
- Use cases span from business user data access (Bayer example) to data transformation engines
- Document AI: extracts structured information from contracts, receipts
- Two main categories: data transformation (bottom) and data democratization (top)
Snowflake's AI Strategy and Right to Win
- Cortex AI: model garden integrated into Snowflake platform
- Any analyst with SQL access has AI access - massive democratization
- Core strength: making reliable AI vs raw language model output
- Talk-to-data applications: GPT-4 out-of-box gets 45% reliability, Snowflake achieves 90%+ reliability
- Focus on grounded applications vs ungrounded language model outputs
- Two-command solutions instead of software engineering projects
Data Strategy and Iceberg Integration
- Expanding beyond proprietary storage format to interoperable data
- 100-1000x more data sits in cloud storage than specialized platforms
- Iceberg: interoperable storage format for cloud data
- Polaris: cloud catalog for data accessibility
- Vision: 10-year future with data in cloud storage, interoperable formats, open catalogs
- Customers running video models on container services, extracting transcripts
Competitive Positioning vs Hyperscalers
- Snowflake at platform layer vs infrastructure layer (hyperscalers)
- 99.9% of customers want simple solutions, not complex engineering projects
- Ecosystems move upstream - comparison to car maintenance evolution
- Total cost of ownership advantage for banks, healthcare, media customers
- Focus on making hard problems simple vs technology for technology's sake
Product Velocity and Engineering Excellence
- Safety net first: regression tests, auto-experiment scaling frameworks
- Inner loop productivity: single change velocity determines total throughput
- System extensibility: 2-year process predating CEO to make unified product extensible
- Leadership focus: clarity, accountability, 2-week and 3-month promises
- Quality and speed both requirements, not trade-offs
- "Virtuosity trumps strategy" - speed of execution beats fixed strategy
Technical Deep Dive: Cortex Analyst
- Systems design approach vs pure AI model approach
- Problem decomposition: "whether to answer" vs "how to answer"
- Product structure constraints: restrictive vs free-flow domain
- Schema complexity: 200,000 tables, multiple "revenue" definitions
- Precision-recall trade-off: useful but precise
- Training for refusal vs attempting every question
AI Reliability and Software Engineering
- Grounded chatbots as norm vs raw language model outputs
- Test frameworks and observability for ongoing AI applications
- Regression testing and acceptance criteria for AI apps
- Models react "violently" to prompt changes (adding periods)
- Need predictable performance metrics for business deployment
Foundation Model Landscape and Future
- GPT-5 expected but unclear what step change it represents
- GPT-4: faster, multimodal, but no major reasoning capability leap
- Multi-step reasoning as potential breakthrough area
- Concern about billion-dollar training costs limiting competition
- Preference for competitive model landscape vs oligopoly
Search and Consumer AI
- Perplexity timing advantage vs Neva (2-3 years early)
- Search business preserved by contracts, not consumer choice
- Consumer choice "fiction" - people use default search engines
- Distribution advantage critical for search disruption
- Apple ChatGPT deal: phone as controlled environment with API mandate potential
Enterprise vs Consumer AI Disruption
- Incumbents learning from previous disruption cycles
- Cost of disruption higher than defensive investment required
- Enterprise companies (including Snowflake) leaning into AI proactively
- Startups safer creating new categories vs "faster because AI-enabled"
- New categories: image/video/voice creation, marketing specialization
AI Applications and Wishlist
- ChatGPT favorite application: utility for coding, language learning, data manipulation
- Desired: actual talk-to-phone mediating between apps
- Simple tasks like copying addresses from email to Uber still "dumb and hard"
- Mobile ecosystem potential for cross-app AI functionality
Future Vision for AI
- AI as enabler of software access for global population
- New layer between humans and software
- Democratization of both software creation and consumption
- Comparison to Google Search impact via Android
- Accessibility over fancy new technology as primary goal
