A Brief History of Large Language Models (koenvangilst.nl)

In the past three years LLMs moved from a novelty to practical workplace partners. The story begins with ChatGPT’s launch, which made conversational, coherent answers feel like genuine understanding. Models then gained multimodal abilities (vision, audio) but still only knew what was in their training data. Retrieval Augmented Generation (RAG) became the crucial workaround: feed a model current documents or company data at query time so it can answer questions about things it wasn’t trained on. Researchers also discovered that “giving the model more time to think” — via chain-of-thought prompting, longer contexts, or iterative reasoning — measurably improves output quality. Those threads converged into AI agents that combine extended thinking with tool access: web search, email, databases, CRMs and APIs. Agents can now break down tasks, fetch evidence, update systems, and execute end-to-end workflows (e.g., look up an account, check inventory, process an order). For AI/ML teams this marks a shift from building ever-larger static models to composing reasoning, retrieval, and tooling pipelines. The likely next frontier is on-the-job learning: persistent, personalized models that update from interactions and map organizational dependencies — moving from temporary context to lasting, safety- and governance-aware knowledge that makes AI behave more like a long-term colleague than a one-off consultant.