20VC: Why All AI Companies Are Under-Valued | The Future of Foundation Models: Scaling Laws, Generalised vs Specialised, Commoditised? | From Unable to Afford Rent to Raising $130M From Index and Peter Thiel with George Sivulka @ Hebbia

Hebbia's founder says AI value lies in applications, not models

Hebbia's founder argues that AI's value will settle in applications and hardware rather than in models. Models can be swapped by changing an API key, while the gains come from agents that analyze documents at scale. Search alone cannot answer most enterprise questions, so running many model calls at inference is how Hebbia delivers accuracy today.

Takeaways

• Models commoditize while hardware and applications capture value. Switching clouds costs $10 to $20m, but switching models takes a new API key. Hebbia is model agnostic, using Anthropic for dense legal documents and OpenAI models for other types. Sivulka sees value in hardware and the agent layer.
• Search-based RAG fails most enterprise questions. Retrieval augmented generation is among the most used AI architectures, yet almost 90% of customer questions could not be answered by search. They asked for analysis of documents, such as exposure to Silicon Valley Bank during the regional banking crisis.
• Scaling at inference beats waiting for larger models. Hebbia's HebbiaMatrix orchestrates hundreds or thousands of smaller model calls over every document. Sivulka says this yields more accurate results than waiting for bigger models, and it is the company's core differentiator.
• AI will create about $100tn of value. Sivulka predicts that value over sixty years, with agentic applications supplying over half of GDP within decades. He rejects the idea that business apps collapse into agents, and argues AI makes humans better and raises firms' assets under management.
• Inference growth weakens NVIDIA's training stranglehold. As inference grows, NVIDIA's hold on training loosens. Sivulka thinks AMD and custom chips may benefit and would buy AMD, while he says he does not know about Intel.
• XAI is the most undervalued AI company. Sivulka says XAI could overtake OpenAI and Anthropic in value within 12 to 24 months and praises Elon Musk's positioning. He calls Klarna's claim of firing half its staff a marketing story.

How Hebbia got off the ground

1. GPT-3's release in June 2020 convinced Sivulka the most important technology was being built by others, so he chose to build the most important product, aimed at finance analysts stuck processing unstructured data.
2. He lived in a closet bedroom in East Palo Alto, working 16 to 18 hours a day on a PhD salary of about $42,000.
3. He drove at dawn to pitch Peter Thiel, and breakfast became a four or five hour conversation. Thiel's round, with Floodgate, totaled about $1m.
4. Mike Volpe at Index led the seed, Maples also invested, and Volpe later added two to two and a half million dollars.