George Sivulka, Co-Founder & CEO @Hebbia: The Future of Foundation Models | E1250

Founder Archetypes and Personal History

• George identifies three common backgrounds among successful founders: those with difficult childhoods, those who are gay, and those who were adopted.
  • He cites Elon Musk (difficult childhood), Jeff Bezos and Steve Jobs (adopted), and Peter Thiel and Sam Altman (publicly gay) as examples.
  • He attributes this trend to a "chip on the shoulder" driving a desire to prove oneself.
• George describes his own upbringing in Staten Island and New Jersey, born to parents who intended to be professional athletes but had a son who was non-athletic and academically focused.
  • He felt like a disappointment compared to his talented athletic siblings and distant father.
  • He was a "math kid" who hacked school tablets to play StarCraft and was popular among other non-athletic, nerdy peers.
• His early ambition was to become an astronaut, leading to a persistent but initially rejected pursuit of a NASA internship.
  • At age 15, he sat outside the NASA Goddard Institute for Space Studies in the snow until a guard allowed him entry after he cold-called numbers from Google.
  • He eventually secured an unpaid internship, published international research the following year, and gained admission to Stanford.

Founding Hebbia and Early Struggles

• Hebbia was founded in June 2020 after George realized GPT-3 had preempted his PhD research on meta-learning.
  • He pivoted to solving a tangible product problem: the high pain points in financial services regarding unstructured data processing.
  • The company initially focused on productionizing RAG (Retrieval-Augmented Generation), becoming the first to build a semantic search engine behind an LLM.
• George lived in a rented closet in East Palo Alto while building the company on a Stanford stipend of $42,000.
  • He worked 16–18 hours a day with no weekends, sleeping on a mattress on the floor or a folding table.
  • He admitted this period was "masochistic" and detrimental to his health but served as a personal crucible.
• Fundraising involved significant personal risk and embarrassment, including pitching investors via Zoom with clothes hanging behind him while living in a closet.
  • He secured a pre-seed round involving Peter Thiel and Floodgate.
  • Thiel invested after a 5-hour breakfast meeting in 2020 where they discussed philosophy and business; Thiel did not invest due to the pandemic but wrote a check later.
  • Mike Volpe (Index Ventures) led the seed round after hearing about the company from his Stanford daughter.

Strategic Pivot: RAG vs. Deep Analysis

• George asserts that RAG is fundamentally flawed for many enterprise use cases because it relies on searching for explicitly stated answers in data.
  • He estimates 90% of enterprise queries are not answerable by simple document search (e.g., "Is this company a good investment?").
  • These queries require deep analysis to distill information that is not explicitly stated, such as assessing management strength or identifying inconsistencies.
• Hebbia shifted strategy to build "agents" that can perform high-level computation and decision-making rather than just retrieval.
  • This involves decomposing complex questions and running models recursively to derive insights rather than finding text.
  • The company distinguishes itself from RPA (Robotic Process Automation), which handles low-skilled tasks, by focusing on high-level, ambiguous decision-making.
• He argues that 90% of current enterprise AI adoption is "fugazi" (fake), with many demos failing in real-world production environments.
  • He positions Hebbia as a platform for driving measurable value in specific, high-stakes use cases like credit agreements and financial diligence.

Market Outlook and Future Technology

• George predicts $100 trillion in economic value will be created by AI agents and agentic applications over the next 60 years.
  • He believes this will represent additional GDP growth rather than value replacement, with over 50% of GDP contributed by agentic apps.
  • He cites the rapid adoption of Excel in finance (18–24 months) as a parallel for how quickly finance will adopt AI, contrary to general enterprise timelines.
• He introduces the concept of "scaling at inference" as a new paradigm distinct from "scaling at training."
  • Instead of training larger monolithic models, Hebbia orchestrates hundreds or thousands of smaller model calls to answer a single complex question.
  • He predicts this approach will yield better accuracy than training-only scaling as the world runs out of high-quality training data.
• He believes the model layer will become commoditized, while value will accrue to the infrastructure (hardware) and application (agent) layers.
  • He views NVIDIA's dominance as potentially destabilized by a shift from training to inference, favoring AMD and custom ASICs for inference workloads.
  • He predicts XAI is the most undervalued AI company and believes Elon Musk's operational efficiency and geopolitical positioning could allow it to overtake OpenAI and Anthropic in value.
• He rejects the notion that AI will simply replace human jobs, arguing it will increase firm AUM and create new roles focused on managing AI agents.
  • He believes AI tools will augment junior analysts by allowing them to access more historical data than any human veteran, making them better investors.
  • He views the shift in business apps from standalone tools to platforms for building agents as the next evolution.

Operational Philosophy and Personal Traits

• Hebbia uses a model-agnostic approach, mixing outputs from OpenAI, Anthropic, and Gemini to optimize for accuracy across different document types (e.g., legalese vs. colloquial text).
  • The company charges per-seat rather than consumption-based to incentivize usage and adoption among business users rather than discouraging it with API costs.
• George identifies religion as a key, often hidden driver of his success, stating he prays for an hour every morning.
  • He views prayer as an "antivirus for the human mind" and a source of fuel for enduring difficult periods.
  • He also cites large-scale oil painting as a primary method for processing subconscious ideas and connecting concepts.
• He believes the "chat" interface is a temporary step, and the future of AI will involve complex, multi-screen agent orchestration requiring human-centric design.
  • He views Hebbia as the "Bell Labs" of defining these new AI interfaces.
• In a "spicy round" of personal questions, he confirms he would not sell Hebbia for $2 billion today and believes most traditional enterprise B2B applications (like Salesforce) will not be replaced due to high switching costs and network effects.

Forward-Looking Statements

• He expects the shift to inference to be the defining macro trend in AI deployment, favoring specialized inference chips over general training hardware.
• He anticipates that the transition from experimental budgets to production value will be driven by business users rather than CTOs, as business users understand the context better.
• He believes the AI revolution will be faster than previous technological shifts (like the agricultural or computer revolutions) because the "encapsulation" of the technology into useful products (like Excel or Hebbia) will accelerate adoption.
• He maintains that while he believes in the potential of government efficiency via AI, he is cautious about the self-reinforcing mechanisms within large organizations like the US government.
• He predicts that the "holy grail" of AI will not be the model itself, but the orchestration layer that manages thousands of agents effectively.