What is AI?
Arificial Intelligence is a broad field, not one product. An attempt to replicate human intelligence with computers
Think of AI as the floor beneath a large library. It includes many ways for computers to spot patterns, make predictions, recognise images, translate language and support decisions. An LLM is one part of that floor: useful for working with language, but not the whole building. We are interested in worthwhile AI: specific uses that create enough real value to justify their cost, energy use and risks.
That means asking what a system can actually do, how reliably it can do it, what it costs, what energy it consumes, what ethical consequences follow from its use and whether it strengthens or diminishes human agency. We are particularly interested in the difference between centaur AI, where people and machines complement one another, and reverse-centaur AI, where humans are left checking, correcting and taking responsibility for systems that increasingly dictate the work.
These judgements are difficult to make from product launches, benchmark scores and predictions alone. Models vary enormously between tasks, and their capabilities, prices and limitations change quickly. Something that looks impressive in a demonstration may create little value in practice, while a modest capability may prove genuinely useful in the right context.
What is an LLM?
An Large Language Model is mathematical predictor or the words and sentences you most likely want to see.
If AI is the library floor, an LLM is a bookcase on it. Its shelves hold a mathematical map of patterns learned from a huge collection of writing. When you give it a prompt, it does not retrieve a single book or know an answer in the human sense. It moves from one likely piece of language to the next, using the patterns on those shelves to build a response.
Under the bookcase are the numbers: language is broken into small pieces called tokens, then represented as numbers that the model can compare and combine. The model uses those numbers to estimate what token should come next. That can produce useful drafts, summaries and ideas, but it can also produce confident mistakes. The result depends on the task, prompt, model and the person checking the work.
We know this is a complex subject, and this page will not make every part of it clear. If there is something we should explain better, or a change that would make the project more useful, we would like to hear from you.
Contribute your feedback
