35 articles
Browse guides
AI slop is the mass-produced output that floods search results, feeds, and inboxes because generating it costs almost nothing. It is a distribution problem more than a model problem: the models can do better, but nobody asked them to. For anyone publishing, it raises the bar — generic AI text now reads as a signal of low effort. In practice: Twelve near-identical listicles on the same topic, none of which answers the question. ...
Attention computes, for every token, how much every other token should influence it. Self-attention does this within one sequence and is the core of the transformer. Its cost grows quadratically with sequence length, which is exactly why long context windows are expensive and why so much research targets this bottleneck. In practice: Reading a contract clause, the model attends heavily to the defined term three pages earlier.
The context window is the model’s working memory for a single request. Everything must fit: system prompt, chat history, attached files, and the response. Exceed it and the earliest content is dropped or the call fails. Note that a large window is not the same as good recall inside it — models still lose track of the middle. In practice: A 200K window holds roughly a 500-page book — but the answer must fit in there too. ...
Deepfakes combine generative video or voice cloning with a real person’s likeness. Detection is unreliable and getting worse, so the practical defences are provenance signals, verification channels, and organisational process. The EU AI Act imposes disclosure obligations on this kind of content. In practice: A cloned CEO voice on a call authorising a wire transfer.
A diffusion model is trained to reverse a noising process. Given pure static and a text prompt, it iteratively denoises until an image emerges that matches the prompt. It replaced GANs as the default because it trains more stably and scales better. In practice: Fifty denoising steps between random static and a finished illustration.
In distillation a large teacher model generates outputs that train a smaller student. The student ends up cheaper and faster while retaining much of the teacher’s behaviour on the target distribution. Most affordable production models are distilled from something larger. In practice: A model 10× smaller and 20× cheaper that handles 90% of your traffic just as well.
An embedding turns text, an image, or an item into a vector in a high-dimensional space arranged so that semantic similarity becomes geometric proximity. That single property powers semantic search, recommendations, clustering, and the retrieval half of RAG. In practice: ‘car’ and ‘automobile’ land near each other; ‘car’ and ‘carpet’ do not, despite sharing letters.
Fine-tuning updates a pretrained model’s weights on a smaller, targeted dataset. It is the right tool for teaching form — tone, format, a classification scheme — and the wrong tool for teaching facts, which RAG handles better and cheaper. Most teams try fine-tuning too early; prompt work and retrieval solve more problems than expected. In practice: 500 examples of your support tone, so replies stop sounding like a press release. ...
The foundation model idea inverts the old workflow: rather than training a fresh model for every problem, you train one general model at great expense and adapt it cheaply via prompting or fine-tuning. That economics is why a handful of labs train models and everyone else builds on top. The EU AI Act addresses much the same thing under the name general-purpose AI model. In practice: One base model powering a support bot, a code assistant, and a summariser — all with the same weights. ...
Frontier model has no crisp technical threshold; it means whatever is currently most capable and therefore least well understood. The term matters mainly in policy, where extra evaluation and reporting obligations attach to the leading edge. The EU AI Act’s nearest concept is a general-purpose AI model with systemic risk. In practice: Every generation, yesterday’s frontier model becomes today’s ordinary default.