Korva is a scroll of the internet’s best ideas.
Why Law Fits Models
The reasoning started in late 2022 with next token prediction. If a model can predict language that way, a field built on documents should be a natural match: lawyers read a lot of them, retrieval could find the right ones, and the work ends in generated text, with precedent sitting inside firms and across common law and history.
Conviction weighed that against what was newly possible, what was valuable within it, and who was aligned with them. Harvey's founders, Winston and Gabe, believed AI would transform the practice of law in a very AI-pilled way. Plenty of investors would call a technology-forward approach backwards and insist on working from the customer back, and Conviction wanted to do both.
How Far It Could Go
The ambition spans a wide range. At one end, a model looks at a landlord-tenant agreement in California and answers a question, which is somewhat trivial. At the other end, it does 85% of the work on an Activision Blizzard acquisition. That is a leap, and it could come next year or five years from now.
How does next token prediction help with legal text?
A language model learns which words tend to follow others by training on enormous amounts of text. Legal writing is full of set phrases, standard clauses, and rigid structure, so its patterns are strong and predictable. Contracts and briefs also repeat across thousands of cases with small variations. That regularity lets a model draft a clause or summarize a ruling in the register lawyers expect.
Why does precedent make law a giant archive?
In common law systems, judges decide cases by looking at earlier decisions, a principle called stare decisis. The practice grew in England over the centuries after the Norman Conquest in 1066 and spread to its former colonies, including the United States. Each ruling becomes a written source for later cases, so the body of law keeps growing. Legal research has long meant searching that record for the cases that match.
Why do legal AI tools invent case citations?
A model that predicts likely text can produce a citation that looks right, with a plausible case name and court, without the case existing. In 2023, a federal court in New York sanctioned lawyers who filed a brief containing cases that a chatbot had made up. Retrieval systems reduce the problem by pulling real documents from a database and having the model work from them. Such systems are now standard in professional legal tools.
More from Invest Like the Best with Patrick O'Shaughnessy
