Key Takeaways

• The AI Periodic Table organises LLMs, RAG, agents, and safety tools into a coherent framework
• Understanding how elements interact matters more than memorising tool names
• Everyday AI tools already combine multiple elements from the table, highlighting the importance of recognising what is present and what is missing

Introduction

If you follow AI news, product launches, or academic discussions, you will notice a recurring problem. New terms appear constantly. Agents, retrieval augmented generation, embeddings, guardrails, thinking models. Each is presented as important, yet rarely explained in relation to the others. The result is cognitive overload rather than understanding.

A recent YouTube talk by Martin Keen at IBM proposed a useful organising metaphor: an AI Periodic Table. Just as chemistry uses structure to explain behaviour, this framework groups AI capabilities into rows and families that reveal how systems are built and how they evolve. This article explains each element clearly and shows how these elements interact in everyday AI use, with a strong emphasis on responsible and reflective learning.

It is important to clarify the nature of the analogy. In chemistry, the periodic table organises fundamental substances by their intrinsic properties, behaviour, and reactivity. In the AI context, the table does not describe irreducible substances. Instead, it organises functional capabilities that are designed for use, assistance, and application. Some so-called elements are primitives, while others are compositions built from those primitives. The value of the table therefore lies not in atomic purity, but in helping us see how AI capabilities combine, interact, and scale in real systems.

The Structure of the AI Periodic Table

The table is organised along two dimensions.

Rows represent increasing levels of complexity, similar to periods in the chemical periodic table:

• Row 1: Primitives. Fundamental capabilities that cannot be meaningfully decomposed further and form the basis of all other systems.
• Row 2: Compositions. Functional combinations of primitives that enable useful behaviours such as retrieval, tool use, and safety control.
• Row 3: Deployment. Capabilities required to operate AI systems in real-world settings, including adaptation, coordination, and testing.
• Row 4: Emerging capabilities. New and rapidly evolving patterns that extend autonomy, reasoning, and data generation.

Columns represent functional families, analogous to groups in chemistry, where elements share similar roles and behaviours:

• G1 Reactive. Elements that directly respond to instructions or trigger actions.
• G2 Retrieval. Elements concerned with representing, storing, and recalling information.
• G3 Orchestration. Elements that coordinate multiple components into coherent workflows.
• G4 Validation. Elements focused on safety, constraint, and robustness.
• G5 Models. Core generative or predictive engines that underpin all other behaviours, analogous to noble gases in chemistry: relatively stable, self-contained capabilities around which other elements interact rather than continuously transform.

Understanding both dimensions allows you to reason about any AI system, from a simple chatbot to a complex autonomous workflow.

Row 1: Primitives

Prompts (Pr)

Prompts are instructions given to an AI system. They define intent, tone, format, and constraints. Prompts are highly sensitive. A small wording change can significantly alter the output. They form the primary interface between human thinking and machine response.

Embeddings (Em)

Embeddings convert meaning into numbers. Text, images, or other data are transformed into vectors so that similarity can be measured mathematically. Embeddings allow systems to search, compare, and recall information based on meaning rather than keywords.

Large Language Models (Lg)

Large language models generate text by predicting likely sequences based on training data. On their own, they are general and stateless. They do not know your documents, your organisation, or your preferences unless other elements are connected to them.

Row 2: Compositions

Function Calling (Fc)

Function calling allows a language model to use external tools before responding. This may include querying databases, calling application programming interfaces, performing calculations, or conducting live web searches. In many recent AI systems, web search is embedded as a tool, allowing the model to retrieve up-to-date information before generating an answer. Function calling therefore enables AI to act, not just respond.

Vector Databases (Vx)

Vector databases store embeddings at scale. They allow fast retrieval of semantically relevant information from large document collections. This creates a form of external memory without altering the model itself.

Retrieval Augmented Generation (Rg)

Retrieval augmented generation, often called RAG, combines embeddings, vector databases, prompts, and language models. Relevant information is retrieved first and then injected into the prompt so that responses are grounded in specific sources.

Guardrails (Gr)

Guardrails constrain AI behaviour at runtime. They include content filters, schema validation, safety rules, and output checks. Guardrails are essential for reducing hallucinations, preventing misuse, and supporting responsible deployment.

Multimodal Models (Mm)

Multimodal models process more than text. They may accept images, audio, or video as input and generate diverse outputs. While powerful, they introduce additional complexity and ethical considerations.

Row 3: Deployment

Agents (Ag)

Agents operate in a loop of planning, acting, and observing. Instead of producing a single response, an agent iteratively works toward a goal. Agents often combine prompts, function calling, and memory.

Fine Tuning (Ft)

Fine tuning adapts a base model using domain-specific data. Knowledge becomes embedded within the model parameters. This improves performance for narrow tasks but reduces flexibility and increases maintenance demands.

Frameworks (Fw)

Frameworks coordinate multiple AI elements. They manage prompts, tools, memory, and control flow. Frameworks simplify development but do not replace understanding of the underlying components.

Red Teaming (Rt)

Red teaming involves intentionally attempting to break AI systems. This includes testing for prompt injection, bias, data leakage, and unsafe outputs. It supports validation and governance rather than optimisation.

Small Models (Sm)

Small models are compact and specialised. They are faster, cheaper, and often sufficient for focused tasks. In many real-world cases, small models are more appropriate than large ones.

Row 4: Emerging Capabilities

Multi-Agent Systems (Ma)

Multi-agent systems involve several agents working together, each with a defined role. Coordination enables complex problem-solving but introduces challenges in oversight and evaluation.

Synthetic Data (Sy)

Synthetic data is artificially generated training data. It is used when real data is limited or sensitive. While useful, overuse can reinforce existing biases.

Interpretability (In)

Interpretability focuses on understanding why models behave as they do. It supports trust, accountability, and regulatory compliance, particularly in high-stakes settings.

Thinking Models (Th)

Thinking models allocate additional computation at inference time to reason more deeply before responding. They improve performance on complex tasks but require greater resources.

Everyday AI Interactions Explained Using the Table

Example 1: Asking an AI to Help with Study Notes

When you ask an AI tool to summarise your lecture notes, several elements interact. You provide a prompt. The language model generates text. If the tool searches your uploaded documents, embeddings and a vector database are used. If the system retrieves relevant sections before answering, retrieval augmented generation is in play. Guardrails may limit unsafe or irrelevant output. What appears simple is actually a structured interaction across multiple rows of the table.

Example 2: Planning Travel with an AI Assistant

If you ask an AI to plan a trip within a budget, the system may operate as an agent. The agent breaks the task into steps, uses function calling to query travel services, observes results, and adjusts its plan. A framework manages the workflow, while guardrails prevent unsafe actions. The behaviour reflects a combination of primitives, compositions, and deployment elements.

Why This Framework Matters for Learning

The AI Periodic Table is not about memorising terminology. It is about developing judgement. When learners understand how elements combine, they can assess AI tools critically, recognise missing safeguards, and avoid unnecessary complexity. This aligns with StudyAnalyst’s commitment to reflective and responsible AI engagement.

What the Table Does Not Show Explicitly

Although the AI Periodic Table provides a powerful way to organise AI capabilities, some concepts that readers naturally look for are not shown as standalone elements. This is not because they are unimportant, but because they are implicitly distributed across multiple parts of the table and only become visible through interaction.

Context as an Emergent Property

Context does not appear as a single element because it is not a discrete capability. Context emerges from how prompts are written, how information is retrieved, how agents maintain state, and how guardrails constrain behaviour. Short-term context is shaped through prompts, while longer-term context is assembled through retrieval augmented generation and agent workflows. Effective AI use therefore depends less on possessing context and more on engineering it carefully.

Memory Across Multiple Time Scales

Memory is also implicit rather than explicit. Embeddings encode meaning, vector databases store retrievable memory, and fine tuning embeds knowledge directly into model parameters. Agents may additionally maintain episodic memory across steps. These represent different time scales of memory, each with trade-offs in flexibility, cost, and risk. The table shows where memory lives, even if it does not name it directly.

Evaluation and Validity (Ev)

The validation family includes guardrails and red teaming, which focus primarily on safety and misuse. What is not shown explicitly is systematic evaluation of relevance, factual accuracy, and appropriateness for purpose. At a basic level, this involves checking whether an AI output is on-topic and factually correct. Over longer temporal scales, particularly in agentic systems, evaluation also concerns whether output quality is maintained or gradually degrades as agents reuse prior outputs, accumulate context, or adapt behaviour over time. In educational and research contexts, this includes monitoring accuracy, drift, and alignment with learning goals. One could imagine evaluation as a primitive within the validation family, but its absence highlights that evaluation is an ongoing practice rather than a static component.

Conclusion

AI systems are becoming more capable, but also more complex. Without structure, learning becomes fragmented and superficial. The AI Periodic Table provides a shared language for understanding how modern AI works and how it should be used responsibly. By focusing on relationships rather than hype, learners can move from confusion to clarity and from passive use to informed engagement.

This article is inspired by and fully credits the original explanation by Martin Keen of IBM. The interpretation and educational framing here are provided by StudyAnalyst for AI literacy and learning purposes.This article was created with the assistance of AI tools, including ChatGPT, and Google AI, for research, structuring, and image generation.