Korva is a scroll of the internet’s best ideas.
Moats that remain
A company with a deep pile of its own data can turn that moat into a self-improving product. It takes an open-source model, fine-tunes it on the company's data, and runs the model inside the company. That way the business owns its whole AI stack end to end, apart from the compute hosting.
Owning the stack
Picture a Coca-Cola or a Cisco: a huge number of users, nowhere near the frontier of AI research, and a massive corpus of data of its own. Using that data as hard as it can speeds up the business and holds off competitors.
AI sovereignty
The pull is stronger because of AI sovereignty, the idea of owning your entire AI supply chain. Companies care about cost and about improving their own systems, and they don't want to hand their data to an outside service that might compete with them one day.
How does fine-tuning specialize a model?
Fine-tuning continues a model's training on a narrower set of examples, which shifts its weights toward that material. A method called LoRA, introduced by Microsoft researchers in 2021, trains only a small set of added parameters and leaves the original weights frozen. That makes tuning far cheaper than training from scratch. A company can run the tuned model on its own servers with its data never leaving the building.
Why do network effects make businesses hard to copy?
A network effect means a product grows more valuable as more people use it. The telephone is the classic case: one phone is useless, and a million are a utility. Metcalfe's law, named for the Ethernet inventor Robert Metcalfe, holds that a network's value grows roughly with the square of its users, though economists debate how well it fits real products. A rival that copies the software still has to rebuild the users.
Why do countries pursue AI sovereignty?
Governments and large firms worry about depending on foreign infrastructure for systems that handle sensitive data. Europe's GDPR, in force since 2018, restricts how personal data moves out of the European Union. In 2020, the EU Court of Justice struck down the Privacy Shield framework for transatlantic data transfers in a case known as Schrems II. Several countries have since funded national compute and model projects.
More from 20VC with Harry Stebbings
