ML//code generation//CriticGPT

CriticGPT is an OpenAI model trained to spot bugs in code generated by LLMs. Such code carries subtle security issues, logic errors and silent failures that pass basic tests, and CriticGPT catches errors that human reviewers miss, especially in complex codebases.

CriticGPT is an OpenAI model trained to spot bugs in code generated by LLMs. Such code carries subtle security issues, logic errors and silent failures that pass basic tests, and CriticGPT catches errors that human reviewers miss, especially in complex codebases.

Anthropic's parallel approach uses sparse autoencoders (SAEs, mechanistic interpretability) to identify the features that cause buggy generation, a direct application of interpretability to code safety.