{
  "id": 4919425,
  "title": "Constitutional Methods for LLMs: Turning Written Principles into Training Signals",
  "url": "https://urgent.news/2026/09/01/constitutional-methods-for-llms-turning-written-principles-into",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-01T18:44:09.000Z",
  "source": {
    "name": "Dev.to",
    "slug": "dev-to",
    "url": "https://dev.to/shrsv/constitutional-methods-for-llms-turning-written-principles-into-training-signals-g7b"
  },
  "original_language": "en",
  "account": "Shrijith Venkatramana, creator of LiveReview, discusses the shift in modern LLMs from simple autocomplete to more complex governance-style decision making. Traditional methods involve collecting human preference data and training reward models to optimize LLMs. However, this approach has scaling issues as humans need to inspect model behavior. Anthropic's Constitutional AI takes a different approach by providing the model with a set of principles, essentially a constitution, and using another LLM to critique and revise the model's responses based on these principles. This changes the representation of the specification from explicit examples to a broader principle, allowing the model to generate examples based on the core idea. The process involves a loop where the model prompts an initial response, the principle critiques and revises the response, which then becomes training data for an improved model. While using an LLM to critique another LLM is not novel, the key is that the constitution acts as an explicit specification connecting the two models, making the system more akin to programming than conventional preference labeling.",
  "summary": "Hello, I'm Shrijith Venkatramana, and I'm building LiveReview — a blast-radius aware AI code review built for your business-critical systems. Star us to help devs discover the project, give it a try, and share your feedback to help improve the product. There is a slightly strange thing about modern LLMs. We are increasingly asking them to make judgments that look less like autocomplete and more…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}