Prompt Tuning (vs Model Training)

AI Prompts as Semantic Formulas

Subjective Metric and Semantic Routing prompts are semantic formaulae

A semantic formula for AI is represented as a combination of a model and a prompt. This formula is commonly contained in an AI agent. Below user provides a semantic question (text and/or media) for the application

Refining the Semantic Formula: A Continuous Process

Think of refining a semantic formula as an ongoing journey of improvement. Each tweak brings the formula closer to perfectly grasping and responding to user input. Similarly, every added instruction or input-output pair in the prompt (or model training) enhances accuracy, inching closer to the goal of 100% accuracy. The regression process acts as the compass, guiding humans to identify the instructions that truly move the needle on accuracy.

Starting Simple: The Building Blocks of AI Agents

When starting to design AI agents, it's best to break down each semantic decision into its most basic form. Crafting a dedicated prompt for each of these individual decisions leads to the most stable and predictable outcomes. Imagine designing logic gates for a circuit, then finding the most efficient way to implement them using the fewest possible NAND gates—a streamlined approach for optimal results.

Unit Testing for AI: Breaking Down Decision-Making

AI agents operate by labeling and ranking information. We can dissect these decision-making processes into smaller, more manageable "units." This approach mirrors how software developers use unit testing to verify individual code components. However, in the AI realm, we employ regression testing to ensure these "units" function seamlessly within the larger AI system.

The Challenge of Complexity: Maintaining Accuracy

As user input grows more complex and varied, maintaining accuracy becomes increasingly difficult. Here, regression analysis shines. Through repeated testing and refinement, this method empowers human reviewers to keep AI agents on track and ensure their responses remain accurate even in the face of complexity.