162 articles
Browse guides
The action space defines what an agent can reach. Too small and it cannot finish the job; too large and it makes more mistakes and expands the blast radius of each one. Scoping it deliberately is a security control, not a configuration detail. In practice: Read the calendar and send invites — but not delete events or email externally.
An activation function decides what a unit passes forward. The non-linearity is the point: stack a hundred linear layers and you still have a linear model, but insert a non-linear function between them and the network can represent curves, interactions, and structure. ReLU is the common default. In practice: ReLU: negative input becomes zero, positive input passes through unchanged.
Adversarial examples exploit the fact that a model’s decision boundaries do not match human perception. Tiny, targeted changes flip the output with high confidence. They are a reminder that high benchmark accuracy says nothing about behaviour under attack. In practice: A stop sign with a few stickers that a vision model reads as a speed limit sign.
Models are stateless, so memory is something you build: short-term state within a task, and long-term storage retrieved when relevant. Without it, an agent relearns the same facts every session. With it, you inherit questions about accuracy, staleness, and data protection. In practice: Remembering across sessions that this client wants British English and no bullet points.
Agentic AI is the adjectival form: the qualities that make a system an agent, chiefly autonomy and the ability to affect the world. The practical shift is from ’the model writes a draft’ to ’the model does the task’, which changes the risk profile entirely. Anything that can act can act wrongly, at speed. In practice: The difference between a model describing how to file the expense and one filing it. ...
The agentic loop is the control structure that turns a stateless model into something that gets work done. Each pass feeds the outcome of the last action back into the next decision, so the agent can recover from failures. Loops need explicit stopping conditions, or an agent will happily retry forever and bill you for it. In practice: Search → read result → realise it is wrong → refine query → try again. ...
An agent is given an objective, not a script. It decides what to do, calls tools to do it, observes the result, and adjusts. The word carries two senses: in reinforcement learning it is the entity acting in an environment, while in current product usage it means an LLM wired up to tools and a loop. Both are correct; check which one a document means. In practice: ‘Find the three cheapest flights and put them in the sheet’ — the agent chooses the steps. ...
Alignment covers the gap between the objective you can write down and the outcome you actually intend. Systems optimise what they are measured on, so a badly specified objective produces technically correct, practically wrong behaviour. RLHF and constitutional methods are current partial answers, not solutions. In practice: Told to maximise engagement, a recommender learns that outrage works.
AI governance is the answer to a small set of questions: who approves a use case, who reviews outputs, what data may go where, and who can stop it. It is mostly organisational design rather than technology. Under the EU AI Act, parts of it stop being optional. In practice: A register of AI use cases with an owner and a risk tier for each.
An AI incident covers real harm and near misses: discriminatory outcomes, unsafe advice, data leakage, dangerous automated action. Treating them like security incidents — logged, triaged, reviewed — is what turns them into learning rather than repetition. The EU AI Act requires reporting serious incidents for high-risk systems. In practice: A chatbot promising a refund policy that does not exist, at scale, for six days.