Checkpoint
A saved 'pause point' of an AI model during the training process. Developers use this to continue working or experimenting from a specific saved point, without having to restart the entire training process.
Context window
The maximum working memory (number of words or characters) that an AI model can remember and process during a single conversation. Once the conversation exceeds this limit, the model begins to 'forget' the oldest information.
Embedding
A mathematical translation where AI converts human text into a series of numbers. This allows the computer to 'understand' the deeper meaning of words and see which texts are conceptually similar.
Fine-tuning
The process where an existing, smart AI model receives additional, targeted training on specific data. This suddenly turns a general model into an absolute expert in one particular task or niche.
Hallucination
The phenomenon where an AI model confidently generates incorrect information, fabricated facts, or total nonsense. This often happens when the model does not know the answer and tries to fill the gaps in its knowledge in a logical-sounding way.
Inference
The actual moment when a ready-to-use AI model is put to work to answer your question or perform a task. This is the 'predictive' or production phase of the software.
Parameter
The virtual brain cells or 'dials' inside an AI model that are set during training to make decisions. The more parameters a model has, the more complex the connections it can make.
Prompt
The written instruction, command, or question that you as a user give to the AI assistant. The more specific and clear your prompt is, the better and more useful the answer will be.
Quantization
A compression technique to make heavy AI models lighter and faster by slightly coarsening the rounding of internal calculations. This allows powerful AI to run on standard consumer hardware with virtually no loss in quality.
RAG (Retrieval-Augmented Generation)
A smart technique where the AI first searches a reliable external database for facts before formulating an answer. Want to build with this technology yourself? Read our explanation of RAG for beginners.
Temperature
A setting that determines the creativity of the AI when writing text. A low temperature results in professional, dry, and factual answers, while a high temperature leads to more imaginative and surprising texts.
Token
The smallest building blocks with which an AI reads and writes text; this is often a short word or part of a longer word (such as a syllable). For more information on how systems process these blocks, see our article what is a transformer.