EMZETT.Gen AI (Generative AI)
In short: Umbrella term for AI systems that generate new content (text, images, code, audio) instead of just classifying or predicting existing data — e.g. language models like Claude or image generators.
In more detail: Generative AI learns statistical patterns from huge amounts of data and uses them to produce new, plausible content — a language model “generates” text word by word based on probabilities, an image model produces pixels from a description. To be distinguished from classic/“discriminative” AI, which, for example, only predicts categories (spam yes/no) but doesn’t create anything new.
In Depth
Large language models (LLMs) like GPT or Claude are the best-known example of Gen AI: trained on huge amounts of text, when generating they each predict the most likely next “token” (word fragment) and append it to the text so far — repeat that thousands of times and coherent text emerges. Image generators (diffusion models like Stable Diffusion or Midjourney), by contrast, usually start from random image noise and remove it step by step towards an image matching the text input.
This technique brings its own pitfalls: “hallucinations” — the model produces confident-sounding but factually wrong statements, because it optimises for plausibility, not truth. Outputs are also not deterministic (the same input can produce slightly different results), and the training data shapes learned patterns and biases in the model.
In software development, Gen AI is increasingly integrated directly into applications via APIs (e.g. chat assistants, automated text summarisation, code completion), rather than only used interactively via a chat interface.
Prompt engineering and context
The quality of generated output depends heavily on how the input is phrased (“prompt engineering”) — precise, well-structured instructions with clear context and, where useful, examples produce considerably better results than vague requests. Modern language models also have a limited “context window” (the maximum amount of text a model can keep “in view” at once) — information outside this window simply isn’t available to the model, which can lead to noticeable limitations with very long conversations or documents.
Fine-tuning and RAG
To adapt a generic model to specific domain knowledge, there are essentially two approaches: fine-tuning (the model is further trained on domain-specific data, which changes the model weights themselves) and RAG (“Retrieval-Augmented Generation”, where relevant information is fetched from an external knowledge source at runtime and added to the prompt, without changing the model itself). RAG is often considered more flexible and cheaper, since the knowledge base can be updated at any time without doing a complete retraining.