162 articles
Browse guides
Regularisation trades a little training accuracy for better generalisation. Common forms are penalising large weights, dropping random units during training, and stopping early. It is the main lever against overfitting when you cannot get more data. In practice: Dropout randomly switches off units each step so the network cannot rely on any single one.
A reinforcement learning agent acts in an environment, observes what happens, and receives a reward signal. Over many episodes it learns a policy that maximises expected reward. RL is behind game-playing systems and, in modified form, behind the alignment step in modern chat models. In practice: An agent learns to play a game with no rules explained — only a score that goes up or down.
Responsible AI is the organisational side of trustworthiness: who decides, who reviews, who is accountable when it goes wrong. It is largely a governance and process discipline rather than a technical one. Where it stays a values statement with no owner and no veto, it is decoration. In practice: A named owner who can stop a launch, not a slide about principles.
RAG splits the job in two: a retrieval step finds passages relevant to the question, then the model answers using those passages as context. It is the standard fix for knowledge cutoffs, private data, and hallucination, because the model is grounded in text it can cite. Retrieval quality sets the ceiling — the best model cannot rescue the wrong three paragraphs. In practice: A support bot that quotes your actual refund policy rather than a plausible-sounding invention. ...
RLHF collects human rankings of model outputs, trains a reward model to predict those preferences, then optimises the language model against it. This post-training step is most of the difference between a raw next-token predictor and a usable assistant. It also encodes the preferences of whoever did the ranking, which is a real and underdiscussed limitation. In practice: Two answers, a human picks the better one, repeat a few hundred thousand times. ...
Robustness is performance outside the tidy test set: typos, edge cases, distribution shift, attack. A model that is accurate on clean data and collapses on real data is not accurate in any sense that matters. It is one of the trustworthiness characteristics in the NIST AI RMF and a requirement for high-risk systems in the EU. In practice: Accuracy drops from 94% to 51% when the input photos are taken in poor light. ...
Role prompting sets a perspective — ‘you are a tax adviser reviewing this for risk’ — which shifts what the model treats as relevant. It reliably changes register and emphasis. It does not add knowledge or authority: telling a model it is a lawyer does not make its output legal advice. In practice: ‘Review this as a sceptical CFO’ surfaces different objections than ‘review this’.
An SDK saves you from hand-rolling HTTP calls, retries, and streaming. It handles auth, typing, and error handling in idiomatic code. Using one is almost always faster and safer than talking to the raw endpoint. In practice: Three lines of Python instead of thirty lines of request plumbing.
Self-supervised learning invents a prediction task out of unlabelled data, most famously ‘predict the next token’. Because the label is just the next word, the entire internet becomes training data with no human annotation. This is why LLMs could scale in a way supervised models never could. In practice: Hide the last word of a sentence, ask the model to guess it, repeat trillions of times.
Shadow AI is the AI version of shadow IT, and it is nearly universal because the tools are free and the productivity gain is immediate. The risks are data leaving the perimeter and unreviewed output entering work products. Blocking tools reliably fails; providing sanctioned ones plus training works better. In practice: Client contracts pasted into a free consumer chatbot to get a summary.