162 articles
Browse guides
Modern speech-to-text is accurate enough across most accents and noisy conditions to be used unsupervised for drafts. It underpins meeting notes, subtitles, and voice interfaces. Accuracy still drops on domain jargon, overlapping speakers, and under-represented languages. In practice: A one-hour call transcribed and summarised before you leave the room.
Structured output constrains generation to valid JSON or a supplied schema, rather than hoping a prompt produces parseable text. It is the difference between a demo and a pipeline, because it removes the whole class of ’the model added a friendly sentence before the JSON’ failures. In practice: Guaranteeing every response has invoice_number, amount, and currency, or fails loudly.
In supervised learning each training example is a pair: an input and its label. The model predicts, compares against the label, and corrects itself. It is the most common and most reliable setup, and its main constraint is that labels are expensive to produce. In practice: 10,000 emails each tagged spam or not-spam, used to train a filter.
Synthetic data fills gaps where real data is scarce, sensitive, or expensive — rare edge cases, privacy-restricted records, balanced examples of a minority class. The risk is compounding: a model trained on its own kind of output can drift away from reality and amplify existing bias rather than correcting it. In practice: Generating 5,000 plausible support tickets to cover a scenario you have three real examples of.
The system prompt sits above the user turn and persists across the session, defining who the assistant is and what it must and must not do. It is where product teams put voice, scope, and refusal rules. It is influential but not a security boundary — treat it as configuration, not as a control users cannot reach. In practice: ‘You are a support agent for an EU fintech. Never give financial advice. Reply in under 120 words.’ ...
Article 3(65) ties systemic risk to GPAI models whose capabilities match or exceed the current state of the art and whose reach could significantly affect the Union market. Designation brings extra duties: adversarial testing, incident reporting, cybersecurity, and evaluation. It is the Act’s answer to frontier models. In practice: A model crossing the compute threshold triggers additional obligations for its provider. Plain-English summary of Article 3(65). The binding text is Regulation (EU) 2024/1689. ...
Article 11 and Annex IV set out what the file must contain: system description, design choices, data, training and testing, performance metrics, risk management, and post-market plan. It must exist before the system goes to market and be maintained afterwards. It is the compliance artefact — no file, no conformity. In practice: The document an authority asks for first, which teams start writing far too late. Plain-English summary of Article 11 / Annex IV. The binding text is Regulation (EU) 2024/1689. ...
Temperature reshapes the probability distribution over the next token. Near zero the model almost always picks its top choice, giving repeatable and conservative output. Higher values flatten the distribution and let unlikelier tokens through, which reads as creativity and, past a point, as nonsense. In practice: Temperature 0 for data extraction; 0.8 for brainstorming taglines.
Test data is the honest scorecard. It must stay untouched during development, because any decision made on the basis of test results turns it into validation data. When test examples leak into training — contamination — reported scores become meaningless. In practice: A model scoring 95% on a benchmark whose answers were in its training data has learned nothing.
Test-time compute is the second scaling axis. Instead of a bigger model, you let the existing model deliberate longer, sample multiple attempts, or check its own work. It gives you a live quality-versus-cost dial that used to be fixed at training time. In practice: Same model, ten times the thinking budget, materially better maths.