Interview
George Sivulka, Co-Founder & CEO @Hebbia: The Future of Foundation Models | E1250
- George Tan expects to achieve a valuation of $100 million in annual recurring revenue (ARR) for his venture through persistent effort and working 16-18 hour days, while maintaining a personal routine of waking at 8:20 AM for an 8:30 AM meeting and an hour of prayer.
- He plans to accept admission to Stanford University to become the youngest PhD student in the school's history but will remain on leave for one year to focus on his company, Hebbia, potentially living in a master bedroom closet in East Palo Alto to minimize costs.
- Tan predicts $100 trillion in economic value will be generated over the next 60 years by AI inference, with agentic applications contributing over 50% of GDP within a timeframe faster than a few decades, driven by a shift where businesses must adopt AI instantly or die.
- The outlook anticipates a transition from training to inference scaling as a dominant industry force, allowing companies to improve results by running smaller models more frequently rather than training larger ones, potentially reducing the cost of intelligence to zero within four years (a seven orders of magnitude drop).
- Industry adoption is expected to accelerate faster than the 18-24 month period seen with Excel in finance, with finance leading the way once specific value is proven, while the majority of the current market remains in an experimental budget phase.
- Tan foresees the commoditization of the model layer and a shift in value creation toward hardware, infrastructure, and the agent layer, with Hebbia's "Hebbia Matrix" intended to orchestrate hundreds of smaller LLM calls to handle complex, open-ended questions rather than simple document searches.
- Despite AI integration, team sizes are not expected to significantly decrease, with mass layoffs viewed as marketing tactics; instead, AI is predicted to increase assets under management (AUM) and drive employment by augmenting human capabilities.
- The market for AI inference tools may see 90% of usage as "fake," but real value is emerging in repeatable use cases, with four canonical pricing mechanisms expected: consumption-based, per-seat, "rent a salary," and flat pricing.
- Competitive dynamics are expected to shift as NVIDIA's dominance may be slightly destabilized by the move to inference, allowing AMD chips or custom architectures to be utilized for cheaper infrastructure deployment.
- Tan predicts that XAI could overtake OpenAI and Anthropic in value over the next 12 to 24 months, driven by Elon Musk's efficient positioning, and that the US government will become a major user of AI to improve efficiency.
- Energy is identified as a critical bottleneck for AI growth, necessitating resources like nuclear reactors to manage data center demands, while Silicon Valley is expected to remain the center of AI despite Hebbia's New York base.
- Salesforce is not expected to be displaced in the next generation due to sticky network effects and high switching costs for human habits, whereas legacy businesses will only survive if they successfully integrate AI.
- The future interface for AI agents is predicted to be a human-centric orchestration layer rather than chat, allowing humans to manage multiple expert agents, with Hebbia aiming to define this new standard similarly to how Excel revolutionized bookkeeping.
- Tan believes the US government possesses fundamentally different propulsion technology regarding UFOs and that his deep religious faith serves as a mental antivirus to sustain him through the pressures of founding.
- The outlook includes a prediction that the transition to AI will be instant compared to historical revolutions, with businesses lacking AI integration facing immediate obsolescence, and that the best results will come from mixing models like OpenAI, Anthropic, and Gemini.
- Investors are expected to benefit from AI tools that allow juniors to access quantifiable data and deals, with a "graduation pathway" preserved as AI augments rather than replaces structured thinking.
- Tan anticipates that the cost of intelligence will follow the trajectory of electricity or storage, becoming effectively free as the industry moves past vaporware and focuses on measurable value in specific use cases like diligence.
- The strategy involves building "better apps" for specific verticals by orchestrating smaller models to handle dense legalese, inconsistencies, and questions about data not explicitly stated in marketing materials.
- Hebbia's competitive moat is defined by its ability to scale at inference to answer complex problems that single-pass models cannot, allowing for higher accuracy and margin growth as spend increases.
- Tan expects to potentially invest in AMD in public markets, believing it will benefit more from the inference shift than others, while viewing the current S&P 500 and all AI companies as undervalued.
- Personal and creative elements include the belief that creativity stems from subconscious processing rather than brainstorming, and that physical exertion in the gym is a method to channel startup pressure and anger.