George Sivulka on Hebbia, Enterprise AI and the Future of Work

Hebbia founder and CEO George Sivulka joins Harry Stebbings to discuss his path into AI, enterprise applications, model economics and the future of work. He has a direct commercial interest in Hebbia, which sells enterprise AI software. The interview does not independently verify the company's performance claims or its predictions about employment and markets.

Sivulka recalls growing up in a family oriented toward athletics while he gravitated toward mathematics. He says the mismatch gave him a desire to prove himself. He jokingly offers categories of founder backgrounds involving difficult childhoods, sexuality or adoption; these personal generalizations are not supported by systematic evidence and should not be treated as predictors of ability.

As a teenager he wanted a NASA internship and says he was rejected repeatedly. On a snowy day he went to the office, was turned away, and called numbers until someone agreed to meet him. After a poor interview he studied the research, returned and obtained an unpaid opportunity that later became paid work, he recalls.

He credits persistence for progress and argues founders should keep going through setbacks. Stebbings asks when persistence becomes counterproductive; Sivulka makes the much stronger claim that almost any business can be forced to a large revenue level with enough determination. That is his belief, not evidence that all ventures can succeed or that product-market fit is irrelevant.

At Stanford he was studying meta-learning when GPT-3 appeared. He says the development led him to focus on building a product rather than competing to create the underlying model. Friends entering finance described painful work processing unstructured information, which suggested a market for Hebbia.

He took leave from his doctorate and rented a closet in a shared house, alternating a mattress and folding table in the small space. He describes financial strain and fundraising calls with clothes visible behind him. The circumstances are his recollection of the early company, not a prescribed route for founders.

An introduction led to a long breakfast with Peter Thiel, who invested in the early round, Sivulka says. He recalls a roughly $1 million pre-seed involving Thiel and Floodgate and a later investment from Mike Volpi at Index. His praise for these investors comes from a founder with an ongoing financial relationship with them.

Sivulka describes retrieval-augmented generation as searching for relevant material and providing it as context to a language model. He says Hebbia was early to deploy such an approach for enterprise data. His historical first-mover claims are not independently established in the interview.

He argues simple retrieval often falls short when the question requires analysis across documents rather than finding an explicit quotation. An investor asking whether a business is attractive may need customer concentration, management assessment and other criteria, not merely a statement in a pitch deck. He presents Hebbia's approach as work performed over the source material.

He recalls reaching $1 million in revenue around 2021 or 2022 and raising a $30 million Series A. He says roughly 90% of questions customers asked could not be answered by simple search alone. Neither the revenue period nor the query sample and measurement method are supplied, so these are company recollections rather than independently verified rates.

Sivulka claims much enterprise AI activity remains experimental and demonstrations often fail in practical use. He argues Hebbia focuses on specific, repeatable tasks and traceable citations, including analysis of long credit agreements. Those are a vendor's claims; the interview supplies no controlled accuracy, reliability or financial-performance comparison.

Stebbings asks whether business apps will collapse into agents. Sivulka disagrees and expects a mixture of applications, platforms and agents. He describes Hebbia Matrix as orchestrating many model calls over documents and argues tools should make their work understandable to people. This is a product vision aligned with Hebbia's business, not a demonstrated requirement for every workflow.

He predicts AI and agentic applications could create $100 trillion of additional market value over decades and eventually contribute more than half of GDP. Stebbings questions the scale of that claim. Sivulka presents it as technological optimism; no methodology or reliable timetable is provided, and it should not be treated as an established economic forecast.

On enterprise adoption, Sivulka argues finance can move quickly when a tool produces clear economic value, citing historical examples of spreadsheets and alternative data. His precise historical adoption comparison is not substantiated in the interview. The ability of a customer to deploy AI also depends on its workflow, security and organizational constraints.

Stebbings worries junior employees could lose the learning path created by doing detailed analysis themselves. Sivulka thinks AI could instead let them compare far more prior deals and challenge senior investors with richer information. His numerical investment example is hypothetical; the interview does not prove that a system's outputs yield better decisions.

Pressed on whether agents will replace people, Sivulka predicts work will change and some junior or back-office tasks will be automated, but believes firms using the tools could manage more assets and employ more people. He questions public claims about AI-driven staff cuts. His employment outlook is a company founder's prediction, not evidence that displacement will not occur.

He expects room for many competitors and describes Hebbia as a platform that can complement specialized applications. The projected $100 trillion opportunity underlying that position is his own assumption. His comparison to Excel illustrates a strategy rather than proving that competing products will necessarily reinforce Hebbia.

He says Hebbia is model-agnostic, using major providers and some of its own models. Its Matrix product breaks a question into many model calls across documents, which he describes as scaling at inference. He claims this produces strong accuracy; the interview offers no matched independent benchmark or full error and cost analysis.

He says different models perform differently across legal and other documents, with tradeoffs among speed and accuracy, and that Hebbia sometimes combines providers. He contrasts domain-specific fine-tuning with larger general models and predicts general models plus inference scaling will prevail. His examples and future ranking are not comprehensive tests.

Sivulka distinguishes pretraining larger models from using more compute when answering a question. He suspects high-quality training data may become constrained, while acknowledging that this is a gut feeling and that he does not collect data himself. Stebbings counters that video and synthetic data could expand supply; the interview does not resolve the question.

He argues orchestrating many model calls can yield better answers without waiting for a larger model. He claims inference costs for a fixed parameter count fell by seven orders of magnitude over four years and that Hebbia has leading accuracy. The cost basis, workloads and benchmark definitions are not supplied, so neither claim can be generalized from the interview.

He expects the model layer to commoditize while hardware, infrastructure and applications capture value. Comparing models with cloud services, he argues changing a model can be easier than moving cloud infrastructure, although he acknowledges that the model business could also resemble cloud. His switching-cost estimates and margin claims are commercial judgments, not verified market data.

Asked to choose among AI labs at stated valuations, he suggests xAI may overtake OpenAI and Anthropic in value within 12 to 24 months, and says he thinks all are undervalued. He cites operational and geopolitical considerations. These are speculative investment opinions from an enterprise AI founder, not verified valuations or investment advice.

Stebbings questions whether AI adoption could be so fast that established businesses are disrupted before they adapt. Sivulka replies that turning a technology into a useful product takes work, though he predicts a good interface can accelerate the transition. His comparison of Hebbia with Excel expresses the company's ambition, not evidence of equivalent adoption or impact.

He considers chat a useful feature but not necessarily the final interface for enterprise AI. As agents do more work, he says, people need ways to understand, coordinate and manage them. He views orchestration and human-readable outputs as product challenges rather than assuming an autonomous agent needs no interface.

Asked about government efficiency efforts, he says reforming a very large organization will be difficult and calls for clear rules for the AI industry. He declines to make a broader political endorsement when pressed. His comments are views on policy and implementation, not a report of results.

Sivulka sees Nvidia's developer ecosystem and CUDA knowledge as a strong human and network advantage. He predicts a shift toward inference could benefit AMD and other architectures, while remaining positive about Nvidia. This is a market forecast, and the interview does not establish relative performance or future share.

He estimates roughly 90% of the market is still using experimental AI budgets, with early repeatable value emerging. He argues business users closest to a workflow should help define uses rather than leaving every decision to IT. The percentage is his estimate without a disclosed survey or denominator, and Hebbia sells to those businesses.

On pricing, he lists consumption, seat, salary-like and flat fees as possible approaches. He says Hebbia uses per-seat pricing because people are central to its product and he wants to encourage use rather than charge for each model call. This describes the company's stated model at the time, not a guarantee about future pricing or unlimited service terms.

In quick fire, Sivulka says he would not sell Hebbia for $2 billion and discusses his relationships with investors. He also states a belief about unidentified flying objects and government-held propulsion technology, without presenting evidence. Those are personal beliefs, not established facts.

He says religious practice and an hour of morning prayer help him find purpose and make room for thought. He also exercises to manage startup pressure. These are personal practices and experiences, not evidence of a universal mental-health effect.

Although Hebbia operates in New York, he identifies it with Silicon Valley's style of technology investment and work. He says he has changed his mind about the importance of inference scaling. Discussing Salesforce, he argues customer habits and switching costs can sustain a business even when underlying software becomes easier to build.

Asked which public company could benefit from AI, he chooses AMD, based on his expectation of a shift toward inference. The choice is a speculative personal investment view, not a recommendation or a verified performance comparison. He also says he would find the role of New York mayor interesting.

He describes painting as a creative practice that gives him time away from distractions and helps connect ideas. Asked whether his parents are proud, he recalls a moment when his father's boss praised his work. He says the drive to prove himself remains, even after that recognition.