162 articles
Browse guides
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.
Autonomy ranges from suggesting, to acting with approval, to acting and reporting, to acting silently. It is the single most important design decision in an agent product, because it sets who is accountable when something goes wrong. The EU AI Act’s definition of an AI system explicitly turns on operating with varying levels of autonomy. In practice: Draft the reply (low) vs. send the reply (high) — same model, very different risk. ...
Backpropagation runs the error backwards through the network, layer by layer, assigning blame to every weight using the chain rule. It is what makes training deep networks computationally feasible, and it is the reason deep learning works at all. In practice: The output was wrong by 0.3 — backprop decides which of the billions of weights are responsible.
Batch inference processes a queue of inputs where latency does not matter, using hardware far more efficiently. Providers often discount it substantially. If a human is not waiting for the answer, it is the cheaper path. In practice: Classifying a year of support tickets overnight at half price.
Batch size trades stability against memory and speed. Large batches give smoother gradient estimates but need more GPU memory; small batches are noisier but sometimes generalise better. It interacts with learning rate, so the two are usually tuned together. In practice: Batch size 32 means 32 examples in, one weight update out.
Benchmarks make models comparable on paper. Their weaknesses are structural: they leak into training data over time, they are optimised for as targets, and they rarely resemble your workload. Read them as a coarse signal and then run your own evals. In practice: A model topping a coding benchmark and still failing on your codebase.
Bias in AI has two senses that get conflated: a statistical one (a systematic offset) and a societal one (unfair treatment of groups). NIST distinguishes systemic, statistical, and human cognitive bias, because each needs a different remedy. Naming which kind you mean is half the work of addressing it. In practice: A model trained on ten years of hiring data reproduces ten years of hiring patterns.
Article 3(40) defines it, excluding uses ancillary to another service and strictly necessary for technical reasons. Categorising people by sensitive or protected characteristics is a prohibited practice; other biometric categorisation generally falls in the high-risk tier. In practice: Inferring ethnicity or political affiliation from a face is prohibited. Plain-English summary of Article 3(40). The binding text is Regulation (EU) 2024/1689.
CE marking is the familiar EU conformity mark extended to AI. Affixing it is a legal statement by the provider that the system meets the requirements and that the paperwork exists to prove it. It is a declaration, not a certificate someone hands you. In practice: The same principle that governs a kettle now applies to a CV screening system. Plain-English summary. The binding text is Regulation (EU) 2024/1689.
Chain-of-thought prompting improves accuracy on multi-step problems by forcing the model to generate intermediate steps rather than jumping to a conclusion. It also makes the reasoning inspectable, which helps you catch where it went wrong. Note that the stated reasoning is not guaranteed to be the actual computation — it is a useful artefact, not a proof. In practice: ‘Work through the calculation step by step, then state the final figure.’ ...