Skip to main content

Lesson 2: Top 7 AI Terms

Topics Covered
  • Agentic AI and how it differs from chatbots.
  • Large Reasoning Models and internal chains of thought.
  • Vector Databases and Embeddings.
  • RAG (Retrieval Augmented Generation).
  • Model Context Protocol (MCP).
  • Mixture of Experts (MoE).
  • Artificial Superintelligence (ASI).

The field of AI is changing rapidly. Here are seven important terms to help you keep up with the latest progress.

1. Agentic AI​

Everybody seems to be building the next generation of AI agents. But what exactly are they?

AI Agents can reason and act autonomously to achieve goals. Unlike a standard chatbot that responds to one prompt at a time, AI agents run autonomously through a loop:

  1. Perceive the environment.
  2. Reason to determine the next best steps.
  3. Act on the plan.
  4. Observe the results.
  5. Repeat.

Agents can act as travel agents, data analysts, or even DevOps engineers spotting anomalies and deploying fixes.

2. Large Reasoning Models​

Agents are typically built using Large Reasoning Models. These are specialized LLMs that have undergone reasoning-focused fine-tuning.

Unlike regular LLMs that respond immediately, reasoning models function by "thinking" before they speak. They generate an internal chain of thought to break down a problem step-by-step. They are trained on problems with verifiably correct answers (like math or code) using reinforcement learning to prefer correct reasoning paths.

3. Vector Database​

In a vector database, we don't store raw data like text or images as simple blobs. Instead, we use an embedding model to convert that data into a vector—a long list of numbers that captures the semantic meaning of the content.

This "vector embedding" allows us to perform searches as mathematical operations. by finding vectors close to each other, we find content that is semantically similar (e.g., finding images of mountains that look like your input image, or text articles about similar concepts).

4. RAG (Retrieval Augmented Generation)​

RAG uses vector databases to enrich prompts sent to an LLM.

  1. Retrieve: The system takes a user prompt ("What is our company policy on remote work?") and converts it into a vector.
  2. Search: It performs a similarity search in the vector database to find relevant matches (e.g., the specific section of the employee handbook).
  3. Augment: It injects that retrieved context into the prompt.
  4. Generate: The LLM generates an answer based on both the user's question and the retrieved facts.

5. Model Context Protocol (MCP)​

For LLMs to be truly useful, they need to connect to external tools like databases, code repositories, or email servers.

MCP standardizes how applications provide context to LLMs. Instead of building one-off connections for every new tool, MCP provides a universal standard. An MCP Server acts as the gateway, letting the AI know exactly how to interface with your system.

6. Mixture of Experts (MoE)​

MoE models divide a Large Language Model into many specialized neural subnetworks called experts.

Instead of activating the entire massive model for every word it generates, the system uses a routing mechanism to activate only the specific experts needed for that token. It then merges their outputs.

Benefit: This allows models to scale up to huge sizes (billions of parameters) while only using a fraction of them during inference, keeping compute costs lower.

7. Artificial Superintelligence (ASI)​

This is the theoretical end-goal for many frontier AI labs.

  • AGI (Artificial General Intelligence): AI that can complete any cognitive task as well as a human expert. (We are slowly approaching this).
  • ASI (Artificial Superintelligence): AI with intellectual scope beyond human-level intelligence.

ASI would essentially be capable of recursive self-improvement—redesigning and upgrading itself to become smarter in an endless cycle. It is currently purely theoretical.