162 articles
Browse guides
One example is often enough to lock in format when the task itself is understood but the shape of the output is not. It is the cheapest way to stop a model from wrapping JSON in prose. In practice: Show one input-output pair, then hand over the real input.
Open weights means the parameters are published, so you can self-host, fine-tune, and inspect. It is not the same as open source: training data and code are usually withheld, and licences often restrict commercial use or scale. Read the licence before assuming freedom. In practice: Downloading weights to run a model entirely inside your own network.
Orchestration is the routing and control logic around the models: retries, fallbacks, cheap-model-first escalation, parallel steps, state. In production this layer, not the model, is usually what determines whether the system is reliable and affordable. In practice: Route easy tickets to a small model, escalate ambiguous ones to a reasoning model.
An overfitted model has learned the noise as well as the signal. It scores brilliantly on data it has seen and poorly on data it has not. The usual fixes are more data, less model capacity, regularisation, and stopping training early when validation loss turns upward. In practice: 99% on training data, 62% in production — a textbook case.
Parameters are the learnable values inside a model — mostly weights and biases. Training is the process of nudging them until the model’s outputs stop being wrong. Headline counts like ‘70B parameters’ describe capacity, not quality: architecture, data, and post-training often matter more than raw size. In practice: A 70B model has roughly 70 billion adjustable numbers; a bigger number does not automatically mean better answers.
Article 3(25) and Article 72 require providers to collect and review experience from deployed systems so problems trigger corrective action rather than accumulating. It closes the loop that shipping usually leaves open, and it pairs with serious-incident reporting duties. In practice: A documented plan for collecting field performance data and acting on it. Plain-English summary. The binding text is Regulation (EU) 2024/1689.
Precision asks: of everything the model flagged, how much was right? Recall asks: of everything it should have flagged, how much did it catch? They trade off against each other, and which one you optimise is a product decision, not a technical one. Accuracy hides both. In practice: A fraud model catching 60% of fraud (recall) with 90% of its alerts genuine (precision) may beat a 99.9%-accurate one that catches nothing. ...
Article 5 bans a defined list, including manipulative techniques that materially distort behaviour and cause harm, exploitation of vulnerabilities, social scoring by or on behalf of public authorities, untargeted scraping of facial images to build recognition databases, and emotion inference in workplaces and schools — with narrow carve-outs. These apply regardless of risk assessment: there is no compliance route. In practice: Emotion recognition on employees during work is prohibited, not merely high-risk. ...
A prompt is everything the model sees for a request, not just your last sentence. It typically bundles a system prompt, conversation history, attached documents, and your actual ask. Understanding that the whole bundle is the input is the first step to controlling output. In practice: ‘Rewrite this in 80 words for a non-technical reader’ is a prompt; ‘shorten’ is a wish.
When many requests share a long, stable prefix — a system prompt, a document, a set of examples — caching lets the provider skip recomputing it. The saving is large and mostly free: put the stable content first and the variable content last. In practice: A 20,000-token manual cached once, then queried a thousand times cheaply.