162 articles
Browse guides
In a GAN, a generator tries to produce convincing samples while a discriminator tries to catch them, and both improve through the contest. GANs dominated image generation before diffusion and remain useful where speed matters, though they are notoriously unstable to train. In practice: The ’this person does not exist’ faces were GAN output.
Article 3(63) covers models trained at scale, typically with self-supervision, that are broadly capable and integrable downstream. Models used purely for research or pre-market prototyping are excluded. GPAI providers carry documentation, copyright-policy, and training-data-summary obligations, with heavier duties where the model is deemed to carry systemic risk. In practice: A general text model sold via API and embedded into a thousand different products. Plain-English summary of Article 3(63). The binding text is Regulation (EU) 2024/1689. ...
Generative AI learns the shape of its training data well enough to sample new examples from it. The output is novel in the sense that it did not exist before, and derivative in the sense that it comes from patterns in the data. This is the category that took AI from a back-office tool to something everyone uses directly. In practice: Asking a model for a first draft of a launch email and getting one in four seconds. ...
GPUs perform the massive matrix multiplications behind training and inference thousands of times faster than a CPU. Their memory capacity sets the ceiling on model size, which is why quantization matters. Supply and price of GPUs shape the whole market. In practice: Whether a model fits in 24GB of VRAM decides if it runs on your desk.
Gradient descent computes which direction each parameter should move to reduce loss, then takes a small step that way. Repeat millions of times and the model converges. Nearly all modern training uses a variant of it, usually Adam. In practice: Like walking downhill in fog: you cannot see the valley, but you can feel the slope under your feet.
Guardrails are the layers around the model: input filters, output checks, allowed-topic scoping, tool permissions, rate limits. They exist because model behaviour is probabilistic and a prompt instruction is not enforcement. Good guardrails live in your code, not only in the system prompt. In practice: An output filter that blocks the reply if it contains anything resembling a card number.
A hallucination is content the model generates because it is statistically plausible, not because it is true. It is a structural consequence of next-token prediction rather than a bug to be patched: the model optimises for plausible, and plausible and true usually coincide, right up until they do not. Grounding, retrieval, and verification reduce it; nothing eliminates it. In practice: Invented case citations, invented DOIs, invented API endpoints — all classic, all confidently formatted. ...
High-risk covers AI used as a safety component of a regulated product, plus the Annex III list: biometrics, critical infrastructure, education, employment, essential services, law enforcement, migration, and justice. High-risk is not banned; it is regulated. Obligations include risk management, data governance, technical documentation, logging, human oversight, accuracy and robustness, and conformity assessment. In practice: Software that scores job applicants or grades exams sits in Annex III. Plain-English summary. The binding text is Regulation (EU) 2024/1689. ...
Article 14 requires high-risk systems to be designed so that natural persons can oversee them effectively — including understanding capacity and limits, watching for automation bias, interpreting output, deciding not to use it, and intervening or halting operation. Oversight must be designed into the system, not asserted in a policy. In practice: A visible stop control and a reviewer who is empowered and trained to use it. Plain-English summary of Article 14. The binding text is Regulation (EU) 2024/1689. ...
Human-in-the-loop puts a checkpoint at the consequential step — sending, paying, publishing, deciding. It is the main practical control on agent risk and, for high-risk systems under the EU AI Act, a version of it is a legal requirement rather than a nice-to-have. The failure mode is rubber-stamping: a review that nobody actually performs is not a control. In practice: The agent prepares the refund; a person clicks approve.