35 articles
Browse guides
Reasoning models generate intermediate steps — a private scratchpad — before producing the final response, and are trained to do this well rather than just prompted into it. They trade latency and cost for accuracy on maths, code, and multi-step analysis. On simple tasks they are slower and pricier with no benefit. In practice: Fifteen seconds of hidden work before a correct answer, versus one second and a wrong one. ...
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. ...
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.
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.
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-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.
Text-to-image systems pair a language understanding component with a diffusion generator, so a prompt steers the denoising toward matching pixels. Output quality depends heavily on prompt specificity — subject, style, composition, and lighting all need saying. Commercial use raises rights questions that vary by tool and jurisdiction. In practice: ‘Isometric illustration of a data centre, muted palette, soft rim light.’
Current text-to-speech produces natural prosody and can clone a voice from a short sample. That capability is the reason voice cloning consent and disclosure have become live legal and ethical issues rather than theoretical ones. In practice: A course narrated in a consistent voice across 40 lessons, without a studio.
Text-to-video extends diffusion across time, which adds the hard constraint that frames must stay consistent with each other. Clip length, physical plausibility, and character consistency are the current limits. The field moves fast enough that any specific capability claim ages within months. In practice: A ten-second product shot generated instead of filmed.