Korva is a scroll of the internet’s best ideas.
A custom model ends the wrapper label
The end of the wrapper
Many AI startups were dismissed as ChatGPT wrappers. Some went from founding to 50 million dollars a year in revenue, and positive cash flow, in about nine months, and it was hard to tell whether any of them would last. Open source changes that: a company can now generate data unique to its use case and train a model on it.
How Nexus works
Fireworks, an inference cloud, released a product called Nexus. For people using Claude Code, OpenAI Codex, or Grok Build, it takes three lines of code, roughly 20 words. Fireworks uses the customer's data to run reinforcement learning on a model, and a router sends each query where it fits.
Independence and its critics
Cursor, Harvey, and Legora lean into this. A company that once used two or three frontier models for everything can use them for 30 to 60 percent of its token consumption and its own model for the rest. Getting off the frontier labs' terms of service gives a company independence and maybe durability. The opposing view holds that once a frontier model reaches recursive self-improvement, it will serve every level of intelligence at far lower cost by distilling itself, leaving no place for open source.
How does reinforcement learning tune a model?
In reinforcement learning, a model attempts a task, receives a score for how well it did, and adjusts its weights to make high-scoring behavior more likely. OpenAI's 2022 InstructGPT paper used a version of this, with human preference ratings as the score, to make language models follow instructions better. A company can substitute its own signal, such as whether a customer accepted a suggestion or a test passed. The model's weights then shift toward that company's definition of a good answer.
How does a smaller model learn from a bigger one?
The technique is called distillation. Geoffrey Hinton, Oriol Vinyals, and Jeff Dean described it in a 2015 paper titled Distilling the Knowledge in a Neural Network. A small student model is trained to imitate the output probabilities of a large teacher model, which carry more information than a single right answer. The student ends up much cheaper to run while keeping much of the teacher's skill.
More from Invest Like the Best with Patrick O'Shaughnessy
