Jan 22, 2025 · 1h 8m · news
George Sivulka, Co-Founder & CEO @Hebbia: The Future of Foundation Models | E1250 · 20VC with Harry Stebbings
0:00 / 0:00
▶ Watch on YouTube →
gold bands on the timeline = statements, start to end. Hover to read, click to jump. CC turns on captions
In this episode of 20VC, Hebbia Co-Founder and CEO George Sivulka explores the intense psychological drive and grit of elite founders, detailing his journey from living in a closet to raising seed capital from Peter Thiel, while outlining Hebbia's technical pivot toward inference scaling over flawed RAG systems.
How this conversation actually went
Every chapter scored 0–10 on four independent dynamics. Hover any point for the reasoning behind the score. Harry holds 19.2% of the talking time here. How this is scored →
Harry as informed peer 3.7 Guest teaching 4.1 Guest disagreement 3.8 Harry pushing back 3.3
speaking balance: gold is Harry, purple is the guest (3 minute bins)
Sharpest disagreement ▶ 26:50 Rejecting Satya Nadella's core thesis on business apps
George forcefully rejects Satya Nadella's claim that business apps will collapse into agents, saying 'I think he's completely wrong' and dismissing the premise entirely.
Hardest push from Harry ▶ 39:51 Challenging scaling laws consensus with industry heavyweightsHarry confronts George by directly citing opposing viewpoints from Reid Hoffman, Marc Benioff, and Daniel Dines on whether training scaling laws have hit a limit.
Biggest teaching moment ▶ 30:09 Reframing enterprise AI adoption speed using Excel's historical precedentWhen Harry argues that enterprises adopt tech slowly due to security and process friction, George counters with concrete historical evidence showing Excel reached 90 percent finance penetration in 18 months.
Harry holds his own ▶ 24:02 Pressing George on market consensus vs. his 'fugazi' claimsHarry hits back against George's contrarian stances by contrasting them with quotes from UiPath CEO Daniel Dines and questioning why the broader market continues to embrace RAG.
the scores for every segment, with the reasoning behind each
| Chapter | Topic | Harry as informed peer | Guest teaching | Guest disagreement | Harry pushing back | Why |
|---|---|---|---|---|---|---|
| George's Childhood, Parental Expectations, and the Math Kid | 1 | 1 | 1 | 1 | Harry opens with warm rapport and asks about George's childhood in New York and New Jersey. George jokingly admits to being a popular math kid who hacked school tablets to play StarCraft. | |
| Redefining the Three Archetypes of Entrepreneurial Success | 2 | 3 | 2 | 1 | George outlines his theory that top founders come from three archetypes (messed-up childhood, gay, or adopted), which Harry validates with his own childhood story. George then shares how he cold-called NASA during a snowstorm to land an internship. | |
| The Philosophy of Brutal Persistence | 2 | 4 | 2 | 2 | George shares his view that persistence can brute-force product-market fit for any business. He explains how OpenAI's GPT-3 paper inspired him to pivot from Stanford research to founding Hebbia. | |
| The East Palo Alto Closet and Extreme Monkish Focus | 1 | 2 | 1 | 1 | Harry asks about George's early housing struggles, and George describes living in a East Palo Alto closet on a PhD stipend, working up to 18 hours a day. | |
| The Crucible of Extreme Work | 4 | 2 | 3 | 4 | Harry challenges the productivity of working 18-hour days, drawing on his own self-described extreme tendencies. George acknowledges it hurt his health but defends it as a necessary crucible. | |
| Pitching Peter Thiel | 1 | 3 | 1 | 1 | George details driving five hours on 18 cups of coffee to pitch Peter Thiel at his house for breakfast, turning a short meeting into a five-hour discussion. | |
| Why Peter Thiel is Unique | 3 | 5 | 1 | 1 | George breaks down Thiel's intellectual traits into ontological and phenomenological intelligence. Harry asks structural questions about early funding rounds and definitions of RAG. | |
| Enterprise AI is 'Fugazi' | 2 | 6 | 5 | 2 | George delivers a contrarian take, stating that despite pioneering enterprise RAG, he believes RAG does not work because questions are about data rather than explicitly in data. | |
| RPA vs. AI Agents | 6 | 4 | 5 | 4 | Harry cites UiPath CEO Daniel Dines to frame RPA versus high-level cognitive AI agents. George pushes back by declaring 90 percent of current enterprise AI adoption to be 'fugazi'. | |
| A Mindfuck: Scaling at Inference | 5 | 7 | 7 | 5 | Harry presents Satya Nadella's view that business apps will collapse into agents. George forcefully rejects the premise, stating Nadella is 'completely wrong' and explaining Hebbia's matrix inference scaling model. | |
| The Speed of Enterprise AI Adoption | 6 | 6 | 6 | 6 | Harry brings up arguments from UiPath's CEO about enterprise adoption lag. George responds with historical data showing Excel conquered financial services in just 18 to 24 months. | |
| Klarna's Layoff Narrative is 'BS' | 4 | 5 | 7 | 4 | Harry raises public narratives like Klarna replacing staff with AI. George dismisses Klarna's claims as pure marketing 'BS' meant to mask internal corporate anxiety. | |
| Inference Scaling: The Tesla Metaphor | 7 | 6 | 4 | 6 | Harry presses on whether training scaling laws are plateauing, citing Reid Hoffman, Marc Benioff, and Daniel Dines. George counters by detailing inference scaling and using a Tesla multi-engine analogy. | |
| Hebbia's Compute Scaling and Accuracy | 5 | 5 | 4 | 5 | Harry questions if running thousands of LLM sub-calls per user query makes Hebbia capital inefficient. George argues the cost of inference is plummeting by orders of magnitude. | |
| AI Lab Valuations and the $100 Trillion Technological Unlock | 4 | 5 | 5 | 4 | Harry asks George to evaluate valuations across top AI labs. George gives a spicy prediction that xAI will surpass OpenAI and Anthropic, arguing all tech stocks are undervalued due to a $100T unlock. | |
| The Speed of the AI Transition & Hebbia's Role | 4 | 5 | 4 | 4 | Harry asks if chat interfaces are appropriate for enterprise workflows. George dismisses basic chatbots and frames Hebbia as the 'Bell Labs' for defining true AI user interfaces. | |
| Geopolitics, DOGE, and Trump's Business Impact | 5 | 4 | 4 | 5 | Harry probes George on DOGE, Trump's election impact, and Nvidia's chip dominance. Harry notes off-camera tech CEO double standards regarding political endorsements. | |
| Real Enterprise Value and CTO vs. Business User Workflows | 4 | 5 | 5 | 4 | Harry asks about agent pricing models. George asserts that CTOs know the least about actual business operations, making business-user-centric per-seat pricing superior. | |
| The Spicy Round: VC Meetings, Hebbia's Worth, and Trust | 5 | 3 | 6 | 5 | In a rapid-fire round, George refuses a $2B acquisition offer, flatly answers 'No' to whether he trusts Sam Altman, and reveals his daily one-hour morning prayer routine. | |
| Startup Stress, Gym Routines, and the New York Exception | 4 | 3 | 6 | 4 | Harry presents a 'shag, marry, kill' scenario with legacy software giants. George answers that he would 'kill them all', labeling traditional enterprise B2B apps unsexy. | |
| Subconscious Creativity, Mayor of NYC, and Dad's Pride | 2 | 2 | 1 | 1 | George discusses his passion for large-scale oil painting as a channel for subconscious creativity, expresses ambition to be Mayor of NYC, and reflects on making his father proud. |
Statements from this episode (44)
Sivulka: Great founders have messed up childhoods, are gay, or adopted
“You can bucket great founders into three backgrounds. The most common is that you have kind of a messed up childhood. The second most common would be, ah, you're gay, and the third most common would be you were adopted. Look at like a list of all-time greats. …”
Sivulka hacked his public school tablets to play StarCraft
“I was the type of kid that would hack the school tablets to put StarCraft on everyone's computer, and then, you know, we'd all, like, not be paying attention in public school playing StarCraft”
Sivulka: Founders can brute-force any business to $100M ARR
“Like, you can bring a lemonade stand to a hundred million dollars in ARR. Like, there's nothing that's actually stopped. You can brute force your way as a founder. You screw product market. You can literally brute force anything in the world.”
Sivulka: ChatGPT is a poor product that acts like a calculator
“I don't even think ChatGPT is a really good product. It's like a calculator. It's got the technology in there encapsulated in a very simple form. But it's not a product that, like in Excel, that lets you just build whatever you'd like with it.”
Sivulka rented a master closet in East Palo Alto starting Hebbia
“I asked my friends who were renting out a house in East Palo Alto to let me rent a room, the cheapest room they could possibly find, and they were all fully booked, and it was like, I think, over a thousand dollars of rent, and they said, I think it was actual…”
Hebbia raised pre-seed from Peter Thiel and Floodgate before Index seed
“We raised a pre-seed from Peter Thiel, and Floodgate, and then our seed from Mike Volpe at Index”
Peter Thiel gave George Sivulka his first offer from a venture investor
“We ended up talking for, I think, like, four or five hours about Not only the company and all of the flaws that I had in my business model, but then also math and deep esoteric philosophy and like just the world. And he said, you know, I'm not investing at the…”
Sivulka: Hebbia was first to productionize RAG in 2020
“We end up building a product studio, which is the first to productionize RAG, Retrieval Augmented Generation, also in 2020.”
Sivulka: Retrieval Augmented Generation (RAG) does not work at all
“That's a bit of a plot twist over here. We, I actually don't think RAG works at all.”
Hebbia hit $1M revenue before raising a $30M Series A from Index
“We ended up going from zero to a million dollars of revenue sometime in 2021 or 2022, and raise our series A also from Index, from Micon Index, which is thirty million bucks.”
Sivulka: 90% of enterprise AI queries require computation beyond RAG search
“And so all of these questions, it was actually almost 90% of the questions that people were asking these systems weren't answerable by search through the documents. But rather they had to be work done on top of the documents and then answered.”
Sivulka: 90% of enterprise AI is vapor and usage stats are fake
“I think, like, 90% of enterprise AI right now is almost like this vapor where, hey, we swear it works, like, look at this amazing demo where we ask, What does the CEO say about the investment? And the minute that they actually go to, you know, try to use it in…”
Sivulka: RPA is simple computation, not modern AI
“I think that I actually am not a big believer in RPA. I think RPA is a, is almost not an AI application in the new sense of AI. It's like AI in the old 10 years ago sense of AI where RPA, RPA is effectively, like, very simple computation.”
Sivulka: Satya Nadella is wrong about business apps collapsing into agents
“I actually don't agree with that at all. I think he's completely wrong. I think it depends on how you define a business app. But I actually think that if the new business apps are platforms, you'll, you'll actually start to see those platforms really take hold…”
Sivulka: AI compute will create $100T in market value within 60 years
“If there was a hundred trillion dollars of value that was created in the stock market from the introduction of the computer, or the fundamental unit of compute, I actually think a hundred trillion dollars of value will be created in the next 60 years from the …”
Sivulka: Agentic AI will generate over 50% of global GDP within decades
“I ultimately, genuinely believe that more than 50% of the GDP will be contributed by what you can call agentic applications in the next few decades, yeah. I actually think it'll happen faster than the next few decades.”
Sivulka: Excel reached 90% finance market penetration in 18-24 months
“If you look at finance and how fast, ah, Excel went to 90% market penetration in finance. 1985 to 1986. Literally 18 months to 24 months, Excel took over all of finance. Everyone switched from using a calculator, the HP 12 C, to using Excel.”
Sivulka: Finance adopts technology faster than any industry when value is proven
“And so finance is the slowest moving, most lethargic, you know, ah, Leviath. It's the worst possible customer base to go after. Unless you're providing outsized alpha or real value. In which case, the minute that there's something real, Finance moves faster th…”
Sivulka: AI will increase financial firms' AUM and drive employment
“I think that it will change the way that people do work, but I genuinely believe that, ah, it'll actually increase the AUM of the firms that use it. I think it will actually drive more employment.”
Sivulka: Klarna's claims of AI-driven layoffs are investor BS
“You know, there's all these stories, and, you know, you have, like, Klarna that's positioning for investors that they're firing half their staff, and no one really wants, and I think. I think it's BS, yeah.”
Sivulka: Loud corporate PR usually signals internal company panic
“When you're saying something and screaming it from the rooftops, that almost always means that inter, and Internally, you're freaking out about something, and so you know, I look at that really loud behavior, and I think that the behavior itself really negates…”
Sivulka: GPT-4 beat BloombergGPT at every financial task
“So then GPT-IV was released, I think, like a few weeks later. I don't know exactly any of the right timeline, but it just destroyed Bloomberg GPT at every single finance task.”
Sivulka: Fine-tuned models will never catch up to inference scaling
“And so you saw the idea of, ah, of post training or kind of like this refined, verticalized model creation, ah, just always would lose to scaling laws. And maybe we're at the end of scaling laws at training, but I actually think, you know, Hebbia and now OpenA…”
Sivulka: AI inference costs dropped 7 orders of magnitude in 4 years
“I think that since Hebbia started, the cost of inference over a fixed number of parameters has decreased by, like, seven orders of magnitude in four years.”
Sivulka: AI model layer will commoditize as value accrues to applications and hardware
“I, you know, I think that ultimately the model layer, and I think this is not a hot take anymore. I've been saying it for a few years, but I think it will become commoditized. I think that a lot of value will accrue at the hardware layer especially And we coul…”
Sivulka: Switching cloud providers costs startups $10M-$20M, unlike AI models
“I can switch models readily. Like, I think there's even entire businesses now There will be an entire industry of being able to switch models from OpenAI to Anthropic when OpenAI goes down, but just to switch clouds is like, you know, for any, like, substantia…”
Sivulka: xAI could overtake OpenAI and Anthropic in valuation within two years
“I think XAI is the most undervalued company, and a really spicy take. I actually think XAI might overtake OpenAI and Anthropic in value over the next, you know, 12 to 24 months.”
Sivulka: All AI companies and S&P 500 stocks are massively undervalued
“I genuinely believe all AI companies and the S&P 500 are all undervalued, which is a very hot take. But again, if we're about to create a hundred trillion dollars of value, I think this is a real tangible Technological shift. It's a massive unlock on the order…”
Sivulka: Current AI chatbots deliver mostly superficial value
“Right now we have these chatbots or these, you know, surface level search engines that, that give you facetious, you know, surface level value. You know, it'll help your kid cheat on their homework, but to drive to whether or not something's a good investment …”
Sivulka: Chat is not the long-term primary interface for AI
“Ultimately, do not think so. I think that chat was always a useful feature. It's an, it's a useful interface, but again, it's like a single cell in Excel. It's like asking if the TI-eighty-four was the right interface for computers, or the terminal was the rig…”
Sivulka: Capable AI agents turn user interface into a management problem
“The better agents are, the more work that they do, the more important it will be that they are easily understood by humans. The idea would be, okay, let's say we have a bunch of employees, 10,000 employees, or 10,000 AI agents drop at a company. They're all ex…”
Sivulka: DOGE will be Elon Musk's greatest challenge
“I think it will be his greatest challenge. I think there's a lot of self-reinforcing Self-protecting mechanisms in the largest organization in the world, which is kind of the U.S. Government by spend, by head, you know, it's just this massive, unruly organizat…”
Stebbings: 99% of tech CEOs publicly back Harris but privately support Trump
“There's 99% of CEOs come on the show, and they either shut up, or they say they vote for Kamala, and then it ends, and they're like, by the way, I'm so Trump, I am so Trump”
Sivulka: The strongest moats are people and networks, not technology
“I think that the best moats aren't technological moats. They're not data moats. They're actually people moats. People and networks have the most friction to change.”
Sivulka: AI shift from training to inference will destabilize Nvidia's dominance
“The shift away from training to inference as a fundamental, like, almost macro shift in how people deploy AI. I actually think that will destabilize slightly the dominance of NVIDIA chips.”
Sivulka: Big Tech custom chips and AMD will win AI inference over startups
“I think it'll probably be large tech providers and AMD.”
Sivulka: 90% of AI market remains in experimental budget phase
“I think that 90% of the market is still in experimental budget phase, but we're starting to see early promises of Actual value.”
Sivulka: CTOs know the least about business workflows for AI
“The CTO or the IT folks are actually the people that know the least about the business. The people that, that actually understand how to use AI in a business context are those that are closest to the business.”
Sivulka: Consumption pricing disincentivizes AI adoption by charging per interaction
“And I think that, ah, when you charge for consumption or API pricing, you're disincentivizing the change. You're saying, ok, well I'm gonna penalize you in a monetary way for every time you use an AI application.”
Sivulka: US government possesses non-traditional UFO propulsion technology
“I believe that UFOs are real. I think a little bit more on the nose right now, but I actually believe there's fundamentally different propulsion technology and that I think the US government has access to it.”
Sivulka: Deep religious faith has been a major driver of founder success
“One thing that I always hid was the fact that I'm deeply religious in an industry that's like very atheistic or agnostic, it was like something that I, that was very personal to me and I think it's been massively contributing to.”
Sivulka: Salesforce will survive AI disruption due to human switching costs
“I think that ultimately I don't think they are. I think that I think that Salesforce Like, has built, again, like, a very, very, very sticky network effect with people, and people are the shifting function at the end of the day. It's not a technology problem. …”