162 articles
Browse guides
Introduced in 2017, the transformer replaced sequential processing with attention, letting the model look at every token in the input at once and weigh their relevance to each other. That parallelism is what made training on internet-scale text practical. The ‘T’ in GPT stands for transformer. In practice: In ’the trophy did not fit in the case because it was too big’, attention is how the model connects ‘it’ to ’trophy’. ...
Transparency is about disclosure rather than mechanism: publishing model cards, documenting data, and telling people they are talking to a machine. The EU AI Act makes several forms of it mandatory, including labelling synthetic content and disclosing chatbot interactions. In practice: A banner stating that the assistant is AI and a human can take over.
Trustworthy AI is the framing NIST uses in the AI Risk Management Framework, listing characteristics a system should demonstrate rather than a single score to hit. The characteristics trade off against each other, so the work is balancing them in context. It is deliberately sector-agnostic. In practice: A system can be accurate and still untrustworthy if nobody can explain or contest it.
Underfitting is the opposite failure to overfitting: the model has not learned enough. Symptoms are poor scores everywhere, and the fixes are more capacity, better features, longer training, or a less aggressive regulariser. In practice: Fitting a straight line to data that clearly curves.
Unsupervised learning looks for patterns without being told what to look for. Typical jobs are clustering similar items, reducing dimensions, and flagging anomalies. There is no accuracy score in the usual sense, so evaluating the result takes judgement. In practice: Segmenting customers into groups nobody defined in advance, based only on purchase behaviour.
Validation data sits between training and test. You use it to choose hyperparameters and to spot overfitting early, because the model never learns from it directly. Reusing it too aggressively leaks information and quietly makes it a second training set. In practice: If validation loss starts rising while training loss keeps falling, you are overfitting — stop.
Vector databases index high-dimensional vectors so that ‘find the most similar items’ runs in milliseconds across millions of records. They are the storage layer under most RAG systems. Whether you need a dedicated one or just a vector extension on your existing database depends entirely on scale. In practice: Storing every paragraph of your docs as a vector so a support bot can retrieve the right three.
Vector search embeds the query and returns the nearest stored vectors, so ‘how do I get my money back’ finds a refund policy that never uses those words. It is the retrieval engine behind RAG. It is also weaker than keyword search on exact identifiers, which is why hybrid search exists. In practice: A query with no shared vocabulary still surfaces the right document.
Watermarking biases generation so a hidden statistical or visual pattern can be detected later. It is a partial measure: image watermarks survive light editing but not determined removal, and text watermarks weaken under paraphrasing. The EU AI Act pushes providers toward machine-readable marking of synthetic content. In practice: A pattern in generated pixels that a detector can read but a viewer cannot see.
Weights determine how much each input contributes to each output at every layer. They start random and are updated by gradient descent until the loss stops improving. When people talk about ‘open weights’, they mean the trained numbers are downloadable, which is what makes local and self-hosted use possible. In practice: Downloading an open-weights model means downloading a file of billions of floating-point numbers.