The 3 Eras of Generative AI Explained

Key Concepts

• Generative AI (GenAI): AI systems capable of generating text, code, or other content.
• LLM (Large Language Model): The foundational models (e.g., GPT, Claude) that predict the next word in a sequence.
• RAG (Retrieval-Augmented Generation): A framework that connects LLMs to private, external data sources to provide grounded, context-aware answers.
• Vector Database: A storage system that converts text into numerical vectors, allowing for semantic similarity searches.
• Agentic AI: Advanced AI systems that can perform tasks, call tools, and make autonomous decisions through iterative workflows.
• Tool Calling: The ability of an LLM to trigger external APIs, web searches, or software functions to complete a task.
• Temperature: A parameter controlling the randomness/creativity of an LLM’s output.

---

1. The Three Eras of Generative AI

The speaker categorizes the evolution of GenAI into three distinct, overlapping eras:

• The Chatbot Era (2022–2024): Characterized by direct consumer interaction with models like ChatGPT. The model functions by predicting the next word in a sequence based on a prompt.
• The RAG Era (2024–2025): Focused on grounding AI in private, organizational data. This allows businesses to provide tailored, accurate responses (e.g., patient portals answering surgery preparation questions based on specific hospital policies).
• The Agentic AI Era (2025–Present): The current frontier where AI systems move beyond text generation to active task execution, such as scheduling appointments or writing and deploying code.

---

2. Technical Methodologies

How LLMs Work (Chatbot Era)

• Next-Word Prediction: Models process input and generate the most statistically likely next word, appending it to the prompt and repeating the process.
• Statelessness: LLMs do not "remember" previous interactions; the entire conversation history is bundled and re-sent with every new query, which is computationally inefficient.
• Temperature Parameter: Controls the model's "creativity." A lower temperature yields precise, predictable results, while a higher temperature allows for more varied, creative outputs.

Retrieval-Augmented Generation (RAG)

• The Problem: LLMs cannot be trained on every private document, and context windows are limited.
• The Solution:
  1. Vectorization: Convert paragraphs of text into numerical vectors.
  2. Vector Search: When a user asks a question, the system converts the query into a vector and finds the most mathematically similar paragraphs in the database.
  3. Context Injection: The system bundles the user's question with the retrieved relevant paragraphs and sends them to the LLM as a single, comprehensive prompt.

Agentic AI Frameworks

Agentic systems rely on three core characteristics:

1. Multiple LLM Calls: Using a chain of models (e.g., one to write, one to critique, one to revise) to improve output quality.
2. Tool Calling: Empowering the LLM to interact with external systems (APIs, web search, file systems).
3. Non-Deterministic Workflows: Instead of hard-coding a sequence, the LLM decides which tool to use and when, based on the current state of the task.

---

3. Real-World Applications

• Healthcare: AI assistants that answer patient questions using specific hospital website data (RAG) or summarize patient charts.
• Marketing: Using an "orchestrator" model to draft a marketing plan, a "critique" model to review it, and a "writing" model to finalize it based on feedback.
• Software Development: Tools like Claude Code act as agents that can read/write files, run Python code, and search documentation to build entire applications autonomously over several minutes.

---

4. Key Arguments and Perspectives

• Evolution of Tool Calling: The speaker notes that tool calling originated from a "hack" where LLMs were instructed to output XML when they needed external data (like a web search). This has since evolved into a native feature supported by modern API providers.
• The "Back Button" Analogy: The speaker argues that Agentic AI is superior to simple chatbots because it mimics human writing processes—allowing for revision, self-correction, and iterative improvement.
• Governance: The speaker acknowledges that empowering AI to call tools is "scary," necessitating robust logging, security, and governance protocols.

---

5. Synthesis and Conclusion

The progression of AI has moved from simple text generation to context-aware retrieval and, finally, to autonomous agency. While the "Chatbot" era introduced the world to LLMs, the "RAG" era made them useful for enterprise data, and the "Agentic" era is transforming them into active participants in workflows. The shift from deterministic, hard-coded software to non-deterministic, LLM-driven decision-making represents the most significant change in how we build and deploy AI applications today.