The Twenty Minute VCGeorge Sivulka, Co-Founder & CEO @Hebbia: The Future of Foundation Models | E1250
CHAPTERS
- 0:00 – 3:28
Childhood roots: misfit athlete, math kid, and the chip on his shoulder
George describes growing up between Staten Island and New Jersey with parents who expected athletic greatness, while he gravitated toward math and academics. That mismatch—plus family dynamics—created a persistent drive to prove himself.
- •Parents’ athletic ambitions vs. George’s academic personality
- •Feeling like the “ugly duckling” among athletic siblings
- •Early life misalignment as fuel for ambition
- •Growing up without strong awareness of elite university pathways
- 3:28 – 5:23
Founder archetypes and the psychology of proving yourself
George shares his provocative framework for three common founder backgrounds and argues they often produce an intense need to validate oneself. Harry relates personally, leading into a candid exchange about insecurity as motivation.
- •Three founder “background buckets” George observes
- •Early experiences as a catalyst for relentless drive
- •Harry’s reflection on feeling like a disappointment
- •George’s admission of similar feelings
- 5:23 – 8:07
Cold-calling NASA at 15: persistence as a life strategy
George tells the story of repeatedly getting rejected for a NASA internship, then showing up in person on a snow day and cold-calling until someone let him in. He converts a botched interview into a research role through overnight preparation and sheer persistence.
- •Physically showing up to NASA’s NYC office without an appointment
- •Mother’s sales-driven coaching: ‘call every number’
- •Turning failure into opportunity by extreme preparation
- •Publishing research and leveraging it into Stanford admission
- 8:07 – 9:55
Relentlessness vs. burnout: when persistence helps and when it harms
Asked when to give up versus push through, George argues persistence is non-negotiable and success is often a function of time and endurance. He also acknowledges the unhealthy edge of his work habits and the physical cost of extreme effort.
- •“Never give up” as a founder operating principle
- •Belief you can brute-force outcomes with enough time
- •Stanford acceptance as fleeting validation; immediate next goal
- •Admitting the health downside of going too hard
- 9:55 – 11:45
The moment Hebbia was born: GPT-3, meta-learning, and choosing product over research
As a Stanford PhD focused on meta-learning, George sees GPT-3 as the research breakthrough he wanted to own. He pivots from building ‘the most important technology’ to building ‘the most important product,’ targeting painful enterprise workflows.
- •GPT-3 reframes meta-learning and changes George’s trajectory
- •Distinction between breakthrough tech and usable products
- •Skepticism that ChatGPT is a complete “product” for work
- •Identifying financial services as peak pain for unstructured data
- 11:45 – 15:59
Bootstrapping in a closet: the monk mode phase of early Hebbia
George describes leaving his PhD, living extremely cheaply, and working 16–18 hour days while training models and avoiding cloud spend. The chapter captures the gritty early reality: physical sacrifice, emotional lows, and fundraising while visibly struggling.
- •Living on PhD stipend; renting a closet-like space
- •Training models obsessively; waking at night to check runs
- •Health trade-offs and “chewing glass” founder moments
- •Early investor conversations while in visibly harsh conditions
- 15:59 – 19:36
Pitching Peter Thiel at breakfast: first capital, worldview lessons, and credibility
George recounts driving pre-dawn, over-caffeinated, to pitch Peter Thiel at his home—only for the meeting to stretch into hours of philosophy and critique. Thiel’s check becomes a credibility inflection point that changes fundraising dynamics.
- •Breakfast-only scheduling and the 3am drive
- •Conversation spanning business flaws, math, and philosophy
- •Thiel’s “not investing, but I’ll put in a check” outcome
- •The signaling power of top-tier early backing
- 19:36 – 21:26
Index follow-on and early product breakthroughs: productionizing RAG and semantic search
With early capital secured, Hebbia attracts Index’s Mike Volpe and accelerates product execution. George frames Hebbia as among the first to productionize Retrieval-Augmented Generation (RAG) in enterprise settings.
- •Volpe discovers Hebbia through Stanford network effects
- •Pre-seed to seed momentum and added capital
- •Early “product studio” approach to enterprise deployment
- •Defining RAG and the ‘search behind an LLM’ architecture
- 21:26 – 26:32
The RAG reversal: why ‘search’ fails and ‘reasoning over documents’ wins
George delivers a contrarian take: RAG is widely used but often fails real enterprise needs because key questions are not explicitly stated in the data. Hebbia’s focus shifts toward doing work on top of documents—distilling truths, surfacing inconsistencies, and producing decision-grade outputs with citations.
- •RAG works for retrieval, but many questions require analysis about the data
- •Examples: investment quality, risk factors, customer concentration
- •Enterprises over-index on demos; real usage exposes failure modes
- •Hebbia’s positioning: stop experimenting, start driving measurable value
- 26:32 – 31:42
Agents, platforms, and the ‘apps collapse into agents’ debate
Harry cites Satya Nadella’s view that apps will collapse into agents; George disagrees and argues the future is a mix of apps, agents, and platforms. He frames Hebbia as a platform designed around what agents/AGI would want to use—tools that orchestrate work and scale computation at inference.
- •Disagreeing with ‘apps collapse into agents’ as too simplistic
- •Platform thesis: tools agents would choose to use
- •Matrix orchestration: many LLM calls vs. one long context window
- •Analogy to compute history: bookkeeping apps, Excel platform, re-unbundling
- 31:42 – 36:17
Jobs, adoption, and the Klarna skepticism: AI as augmentation, not mass replacement
The conversation turns to whether AI removes the training pathway for juniors and whether headcount will shrink. George argues AI will make humans better, widen access to experience, and that high-profile “we fired half the staff” stories are often marketing rather than reality.
- •AI can expand juniors’ capabilities by analyzing far more precedent data
- •Prediction: shift to managing “AI juniors” for lower-cognition tasks
- •Skepticism toward loud job-displacement narratives (e.g., Klarna)
- •Competition is expected in a massive value-creation wave
- 36:17 – 38:33
Scaling at inference: Hebbia’s core technical bet (and why model choice matters less)
George explains “scaling at inference” as running many decomposed model calls to compute over large document sets, improving accuracy today without waiting for bigger trained models. He also discusses model-agnostic routing across OpenAI, Anthropic, Gemini, and trade-offs by document type.
- •Inference scaling vs training scaling: more compute at answer time
- •Decomposition into hundreds/thousands of sub-tasks across documents
- •Cost of intelligence trending down; accuracy becomes the differentiator
- •Pragmatic model mixing by task/document characteristics
- 38:33 – 42:58
Specialized vs general models, scaling laws, and why vertical fine-tunes keep losing
Using BloombergGPT as an example, George argues vertical models are often eclipsed by next-gen general models due to scaling effects. He then distinguishes training-era scaling from inference-era scaling and suggests inference scaling may be the next durable advantage.
- •BloombergGPT vs GPT-4: domain data still lost to frontier scaling
- •Debate on whether training scaling laws are slowing due to data/constraints
- •Inference scaling as ‘many smaller engines’ vs one massive engine
- •Implications: fine-tuning alone won’t keep pace with frontier progress
- 42:58 – 57:55
Commoditization, chips, cloud parallels, and the geopolitics of AI winners
George predicts the model layer will commoditize, with value accruing to hardware, infrastructure, and applications/agents. He compares model switching (easy) to cloud switching (hard), offers a bullish take on xAI, and argues geopolitics and energy constraints will shape outcomes.
- •Model APIs are easy to swap; clouds are sticky due to switching costs
- •NVIDIA’s moat as a people/network moat via CUDA; inference may diversify chips
- •Cloud providers using models as loss leaders to reinforce platform moats
- •Geopolitics, energy, and operational efficiency as decisive competitive factors
- 57:55 – 1:08:12
Quick-fire: deals, trust, UFOs, religion, creativity, and the enduring ‘chip’
In the closing segment, George answers rapid questions: he wouldn’t sell for $2B, doesn’t trust Sam Altman, believes UFOs are real, and describes deep religious practice as a source of resilience. He also discusses painting as a channel for subconscious connections and ends on whether his parents are proud—while admitting the chip remains.
- •Wouldn’t sell Hebbia for $2B; avoids naming worst VC meeting
- •“Do you trust Sam Altman?” → “No”
- •Daily hour-long prayer; faith as endurance fuel
- •Creativity via painting; parents’ pride and lingering drive
