Category: Glossary

Hallucination is when an AI model produces confident, fluent text that is factually wrong — a known property of language models that you have to work around.

Extended thinking is Claude’s deliberate-reasoning mode — the model spends extra compute on hard problems before answering. Better on math, code, plans.

Sub-agents are smaller AI agents that a main agent delegates specific tasks to — like a project manager handing parts of a job to specialists who report back.

The Claude API is Anthropic’s developer interface for calling Claude programmatically — pay-as-you-go, used to build Claude into your own apps.

Coding agents are AI systems that read your repo, plan changes, run commands, fix their own errors, and ship features — not just suggest snippets like Copilot.

Memory cache (prompt caching / KV caching) stores the work an AI model already did on repeated prompt parts, so it doesn’t redo it on every single request.

Parallelization is doing many calculations at once across many chips — why training a frontier model needs tens of thousands of GPUs working in a cluster.

RAMageddon is industry slang for the severe shortage of DRAM and HBM memory chips caused by AI data-centre demand outstripping global supply in 2026 now.

Token throughput measures how many tokens an AI system produces per second — how fast answers feel and how many users a system can serve at the same time.

Compute is the umbrella term for the raw computational power — GPUs, TPUs, memory, networking — needed to train and run AI models. Strategic resource now.

Validation loss is the score an AI model gets on data it didn’t train on — the most reliable signal of whether the model is learning, not just memorising.

Web feeds (RSS and Atom) are machine-readable summaries of a site’s latest content — letting any software subscribe to updates without scraping the site.