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

I watched another interesting tutorial. video Martin Keen of IBM, where he tries to bring AI terminology into a comprehensible system. All because we have a bunch of terms like "agents," "RAG," "embeddings," "gardrails," all these terms fly around, and developers are expected to just know how it's all connected, but it's not that easy to put in your head. To do this, Martin proposed his structure for systematization as a periodic system. ~~Mendeleev~~ "AI Periodic Table." This table, strangely enough, has two dimensions.

**Lines. (period) - Etpa.

1. Primitives** - atomic blocks (Prompts, Embeddings, LLMs) 2. Compositions - primitive combinations (RAG, Vector DBs, Guardrails) 3. Deployment - production patterns (Agents, Frameworks, Fine-tuning) 4. Emerging - front edge (Multi-Agent, Thinking Models, Interpretability)

**Columns (family)

1. Reactive (jetty)** Change of entry to radical change of exit 2. Retrieval (search) - storage and retrieval of information 3. Orchestration (orchestration) binding of multiple components 4. Validation (validation) - safety and testing 5. Models (model) Stable fundamental capabilities

If you analyze the contents of the table by columns, it will be interesting

Reactive Family

• Pr (Prompt) - instructions for AI - Fc (Function Calling) Calling external APIs/tools - Ag (Agent) - Think, act, observe. - Ma (Multi-Agent) Collaboration of several AI agents

Retrieval Family - Em (Embeddings) - numerical representations of meaning - Vx (Vector Database) - storage for semantic search - Ft (Fine-tuning) Adaptation to Domain Data - Sy (Synthetic data) Synthetic data, on which new models are often learned

Orchestration Family - Rg (RAG) Retrieval Augmented Generation - Fw (Framework) Platforms like LangChain

Validation Family - Gr (Guardrails) - runtime security filters - Rt (Red Teaming) Adversarial testing with Red Teams - In (Interpretability) Understanding the “why” model works this way during inference

**Models Family

• Lg (LLM)** Large language models from OpenAI, Antrhopic, Google, Alibaba, DeepSeek and others - Mm (Multi-modal) - Multimodal models that allow processing in addition to text images, audio and so on - Sm (Small Models) Distilled models for edge and not only - Th (Thinking Models) Chain-of-thought built into new model architecture

Further in the video, Martin talks about how such a picture in his head helps to better lay down information, as well as to think about solving problems related to AI in the real world. I like the concept - I often think visually, and it looks good and fits well into my head:))

Plus, I think this scheme can be used for Design and run ideas through the prism of the table. For example, when someone pitches an AI solution, you can instantly decompose it into table elements:

• What elements are used?
• What reactions do they trigger? Is there no security element? (Gr)? Is there over-engineering in orchestration? Does the Thinking Model work where the Small Model does?

#AI #DevOps #Software #Engineering #Architecture #SystemDesign