Salmon exposed to cocaine and its main byproduct roam more widely
Summary
The article surveys methods for producing typed outputs from LLMs, contrasting post-hoc error correction with inference-time constraints. It discusses training-time strategies that encode type structures and differentiating through structure to produce well-typed outputs, and references AlphaZero-like training as a blueprint for scalable performance. The piece argues that encoding programming language structure into training can unlock gains as models scale.