Blog
Thinking on model improvement, application development, and the layer between.
LLMs are best at semantic backend work
LLMs are strongest on semantic backend tasks like routing and normalization, and much weaker on deterministic execution. That is also why small models can be more practical there than many developers expect.
The right model for application logic is not always the smartest one
When an LLM is used as part of application logic, the relevant question is often not which model is smartest in general, but which model can do a specific job well enough, cheaply enough, and reliably enough.