AI Periodic Table Explained: Mapping LLMs, RAG & AI Agent Frameworks

AI Periodic Table Explained: Mapping LLMs, RAG & AI Agent Frameworks

Understanding AI Through a Periodic Table Framework

Introduction to the AI Periodic Table

The speaker compares the complexity of AI terminology to chemistry, suggesting a structured approach to understanding it.

An "AI periodic table" is introduced as a conceptual framework for organizing various AI elements, although it is not an official model.

Row 1: Primitives in AI

The first element, Pr (Prompt), represents instructions given to an AI and is categorized as atomic within row 1, which signifies primitives.

Prompts are classified under group G1 (Reactive family), indicating that small changes can lead to significantly different outputs.

The second element, Em (Embeddings), refers to numerical representations of meaning used in semantic search and vector databases. It belongs to group G2 (Retrieval family).

The third element, Lg (Large Language Models or LLMs), such as ChatGPT, falls under group G5 (Models family), representing stable foundational capabilities in AI.

Row 2: Compositions in AI

The first element of row 2 is Fc (Function Calling), where LLMs invoke tools like APIs for real-time data retrieval; this remains in the reactive family.

Next is Ag (Agents), which utilize think-act-observe loops for planning and executing tasks autonomously; this marks a shift towards deployment and autonomy.

The second element of row 2 includes Vx (Vector Databases), optimized for storing embeddings and facilitating semantic searches within the retrieval family.

Another key component is Rg (RAG - Retrieval Augmented Generation). This orchestrates multiple elements by retrieving context before generating responses with LLMs, placing it in group G3 (Orchestration family).

Lastly, Gr (Guardrails) ensures safety during runtime by implementing filters and validation checks; this fits into group G4 focused on validation.

This structure provides clarity on how various components interact within the realm of artificial intelligence.

AI Periodic Table: Understanding Key Elements

Fine Tuning and Memory in AI Models

Fine Tuning (Ft): Adapting a base model to specific data or use cases, such as medical papers or company codebases. This process is categorized under retrieval as it involves adaptation.

Memory Types: Three forms of memory are discussed:

Embeddings: Encode meaning.

Vector Databases: Store information for search purposes.

Fine Tuning: Stores knowledge directly in model parameters.

Frameworks and Red Teaming

Framework (Fw): Platforms like LangChain that integrate various components to build and deploy AI systems, falling under orchestration.

Red Teaming (Rt): Involves adversarial testing aimed at identifying vulnerabilities in AI systems through methods like jailbreaks and prompt injections.

Emerging Concepts in AI

Multi-Agent Systems (Ma): Multiple AIs collaborating on tasks, each specializing in different roles such as research, writing, and critique. This represents an emerging trend in AI development.

Synthetic Data (Sy): The generation of training data using AI when real examples are insufficient. This practice is becoming more prevalent due to limitations in available data.

Interpretability and Thinking Models

Interpretability (In): Understanding the decision-making processes of models by examining their internal workings. This is crucial for ensuring safety and validation.

Thinking Models (Th): Advanced models that incorporate reasoning into their architecture, allowing them to engage in complex thought processes before generating responses.

Practical Applications of the AI Periodic Table

The speaker presents a practical application by illustrating how elements combine to create a chatbot capable of utilizing company documentation effectively.

Starts with embeddings for document vectorization, stored in vector databases.

Utilizes RAG for querying relevant information which augments prompts sent to large language models.

Agentic Loop Example

An example illustrates the agentic loop where an AI is tasked with booking a flight:

The agent breaks down the goal into actionable steps using function calling to interact with external APIs for flights, calendars, and payments.

It follows a cycle of thinking, acting, observing—demonstrating iterative problem-solving capabilities.

Mapping New Features to the Table

The speaker challenges listeners to evaluate new AI features against this periodic table framework:

Consider what elements are utilized and whether any critical safety measures are missing or if there’s over-engineering involved.