Korva is a scroll of the internet’s best ideas.
Lovable does not rely on one model. Each request is routed to the model most suitable for what the customer wants, drawn from commercial frontier models by multiple vendors and, increasingly, open-weight models. When a request goes to Lovable's own model, that model gets more intelligent for its agent harness, especially on mistakes such as which tool to call and which integration to create.
The mistakes are prioritized by what drives the most impact for customers. For the problems where frontier models are making mistakes right now, the team creates data sets or applies reinforcement learning. A strong research team in Stockholm leads this post-training work, and the company is scaling it up significantly because it considers the capability part of the European ecosystem.
The raw material is an enormous token distribution from a million new products built every week. The agent harness and its skills, including remembered facts from software engineers who know how to build really good software, are modified every single week. Lovable has never chosen a cheaper model where it is measurably worse for customers, and it can measure that.
How does reinforcement learning improve a model?
In reinforcement learning, a model attempts a task, receives a score for the result, and adjusts so that high-scoring behavior becomes more likely. OpenAI's 2022 InstructGPT paper used a version of this, with human preferences supplying the score. For software agents, the score can come from an automatic check, such as whether a tool call succeeded. The method suits problems where success is easy to verify, even when the steps to reach it are not.
Why route different requests to different models?
Models differ in speed, price, and skill at particular tasks, and no single one leads on every measure. A small, fast model can handle a simple request in a fraction of the time and cost of a large one. A router picks among them based on what the request needs. The approach is common in products that serve a wide variety of tasks.
How do models learn which tool to call?
A model picks a tool by predicting text, including the name of the tool and its arguments, from a description it has been given. Meta researchers showed in a 2023 paper called Toolformer that a model can learn to decide for itself when to call an external tool. Mistakes happen when descriptions are ambiguous or when two tools overlap. Training on examples of correct and incorrect calls reduces those errors.
More from All-In with Chamath, Jason, Sacks & Friedberg
