TechCrunch publishes an AI jargon guide: what LLMs, hallucinations and RAG mean

The AI boom of 2024–2026 has brought not only new products and capabilities but also a whole layer of professional jargon that is easy to get lost in. LLM, RAG, hallucinations, tokens, fine-tuning, prompts: these words turn up more and more in news stories, presentations and business negotiations, yet most people's understanding of them stays vague. TechCrunch has published a detailed glossary explaining the key concepts of the AI era, from basic architecture to applied techniques for working with models.

Most modern AI systems are built on large language models, or LLMs. These are neural networks trained on huge volumes of text. They do not "understand" language in the human sense, but they can generate statistically plausible answers to any query.

The basic unit an LLM works with is the token: roughly three to four characters, part of a word or a punctuation mark. GPT-4o processes up to 128,000 tokens at a time, about 300 pages of text. The larger the context window, the more information the model can take into account when forming a response.

One of the main flaws of LLMs is hallucinations. The term describes cases where a model confidently produces factually incorrect information: invented quotes, nonexistent sources, false dates. This is not "lying" in the ethical sense. The model simply generates plausible-sounding text and has no built-in mechanism for checking facts.

The RAG (Retrieval-Augmented Generation) method was developed to combat hallucinations: before generating an answer, the system searches a real database for relevant fragments and relies on them. Many corporate AI assistants and next-generation search engines work this way today.

When a base model needs specialization, it is retrained on narrow data. This process is called fine-tuning: the model learns to answer in a particular style, on a specialized topic or within a set format. A more accessible approach is prompt engineering: phrasing requests skillfully to get the desired result without retraining the model.

AI agents are a separate and fast-growing class of systems. They do not just answer questions; they plan and carry out chains of actions: searching the internet, running code, managing files and a browser. The best-known examples are Claude Computer Use and OpenAI Operator.

Other key terms include:

• parameters: the numerical "weights" of a neural network that determine its behavior (GPT-4 is estimated to exceed a trillion parameters);
• inference: the process of getting an answer from a trained model in real time, which determines the speed and cost of AI services;
• embeddings: numerical representations of words and texts that make it possible to measure how close concepts are in meaning.

Multimodality means a model's ability to work with several types of data at once: text, images, audio and video.

Understanding the basic AI vocabulary is no longer the preserve of developers. It is becoming a necessity for managers, investors, journalists and everyone who works near this technology. The jargon keeps growing: each new class of models brings its own terms, such as multi-agent systems, synthetic data and streaming inference. But once you have mastered the core (LLMs, tokens, hallucinations, RAG, fine-tuning and agents), you can navigate most publications and conversations about AI with confidence.