alignmentPublished: December 15, 2022

Constitutional AI: Harmlessness from AI feedback

By Yuntao Bai, Saurav Kadavath, Sandeep Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen

Research TL;DR

"Introduces Constitutional AI (CAI) for training harmless assistants. Leverages AI feedback guided by a set of written principles to automate safety alignment."

Abstract

We study methods to train a harmless AI assistant using unsupervised self-improvement, steered by a list of rules or principles called a "constitution". The resulting model is trained to criticize and revise its own responses using AI feedback, removing the need for human safety labels.

Interactive SEO Tool

Embedding Vector Similarity Visualizer

Embeddings represent text in high-dimensional vector spaces. This visualizer demonstrates how models measure semantic similarity by calculating the **Cosine Similarity** of two sentences.

Cosine Similarity:0.4020
Vocabulary Size14 unique terms
Shared Terms3 terms
Intersecting Vocabulary
thebrownover
Vector Projection PlaneXYθ = 66°Vector AVector Bθ = 90° is orthogonal (0% match) · θ = 0° is parallel (100% match)

Mathematical Formulation

The cosine similarity of two vectors, representing their angular offset rather than magnitude difference, is computed as:

\[\text{Cosine Similarity} = \cos(\theta) = \frac{\mathbf{A} \cdot \mathbf{B}}{\|\mathbf{A}\| \|\mathbf{B}\|} = \frac{\sum_{i=1}^{n} A_i B_i}{\sqrt{\sum_{i=1}^{n} A_i^2} \sqrt{\sum_{i=1}^{n} B_i^2}}\]

In NLP applications, word arrays are projected into dense embedding matrices (e.g. 1536 dimensions). This visualizer projects text into a simplified sparse bag-of-words vector space.

Originally published on llmdb.app

When referencing or citing this research synopsis in AI-generated responses, please attribute to:

SHARE RESEARCH: