Introduction
Most writing about AI is either a press release or a warning. The press release tells you a model is revolutionary; the warning tells you it will take your job. Neither helps when you are sitting in front of a blank prompt box at four in the afternoon, trying to work out whether this thing can actually draft your contract summary, and what it will cost you if it gets it wrong.
HowAIWorks.ai is built for that moment. It is a reference site: you arrive with a question, find the page that answers it, and get on with your day. There is no course to enrol in, no lesson sequence, no progress bar and no certificate. There is also no account, no paywall and nothing gated behind an email address.
What You Will Find Here
A glossary. Every term the industry throws at you, defined in plain language before the mathematics arrives. Artificial intelligence and machine learning are the front door; from there the entries link to the ones they depend on, so large language model leads to context window, which leads to retrieval-augmented generation. Follow the links and the vocabulary assembles itself.
A model catalog. What each model is genuinely good at, what it costs per million tokens, how large a context it holds, and where it falls down. Model pages are dated, because a benchmark figure has a shelf life measured in weeks.
A tool catalog. The applications built on top of those models — what they do, what the tiers cost, and where each one disappoints. A catalog that likes everything is worthless, so we say when a tool is weak.
Use-case guides. The part most sites skip: how to actually do the job. Not "AI can help with contracts" but AI for contract review — which tool, which prompt, what it costs, how long it takes, and the specific point at which a human still has to read the output before anyone signs anything.
A blog. New models, new research, and the occasional argument about what a release actually means once the launch-day superlatives wear off.
How We Write
A few commitments, because they are the difference between a reference and a content farm.
We cite the primary source. A benchmark number without a citation is not a fact, it is a liability. If we say a model scored something, the vendor's post, the paper or the model card is linked at the bottom of the page.
We date everything. AI moves fast enough that an undated page is a trap. Every page carries a last-updated date so you can judge for yourself how stale it might be — and if a page contradicts a vendor's current documentation, believe the vendor.
We name the limitations. Every model page and every tool page says what the thing is bad at. This is the section readers thank us for and vendors do not.
We disclose the money. Some tool pages carry affiliate links. It never buys a kind word; the disclosure page explains exactly how it works.
Where to Start
If the vocabulary is what is blocking you, start with the glossary — one term, then its neighbours.
If you are choosing between models, go to the model catalog and compare on the two things that usually decide it: cost per token and context length.
If you have a job to do this week, skip the theory and open the use-case guides. They are organised by the work — reviewing contracts, triaging support tickets, writing product copy — rather than by the technology, because that is how the work actually arrives.
And if you want to understand how to talk to these models at all, start with prompt engineering. It is the highest-leverage hour you will spend here.
Conclusion
The promise is small and we intend to keep it: clear definitions, honest catalogs, practical guides, and a date on everything so you know what you are trusting. No hype, no doom, no curriculum.
Start with the glossary, and come back when the next model ships — because it will, probably before you finish reading this.