Back to basics
The ideas behind the guides, in plain words
The how-to guides on this site build things with AI, servers and a few unfamiliar words. This page explains those words once, without assuming you already know them. Nothing here is aimed over your head, and nothing here is aimed beneath you either. If a guide uses a term you have not met, it links back to the right spot below.
Running AI on your own computer
Most AI you have used runs on a company's computers somewhere else. You type, it sends your words away, an answer comes back. “Running it locally” means the same kind of program runs on a machine you own, in your home or office. Nothing you type has to leave the building.
Why it matters: Two reasons people do it. Privacy: your words, your voice and your documents stay with you. Cost: once the machine is paid for, you are not paying a monthly fee per use. The trade is that you look after the machine, and a home machine is usually slower than a big company's.
A model
A “model” is the trained program that does the actual thinking. It might turn your speech into text, or answer a question. It is a large file you download once. Different models are better at different jobs: some transcribe speech, some write text, some make images.
Why it matters: You will see models described by size, like “14 billion parameters”. Bigger models are usually more capable but need a more powerful computer and are slower. Picking the right size for the machine you have is half the job.
Tokens, and “tokens per second”
A token is a small chunk of text, roughly a short word or part of a word. AI programs read and write in tokens rather than whole sentences. “Tokens per second” is simply how fast the program produces its answer.
Why it matters: It is a speed number, like miles per hour. As a rough feel: around 5 to 10 tokens per second is slower than a person reads and feels sluggish; 30 or more feels instant, faster than you can read along. So a machine doing “0.4 tokens per second” is painfully slow: an answer that should take a moment takes most of a minute. Meanwhile “165 tokens per second” arrives quicker than you can keep up with.
GPUs and VRAM (why the graphics card matters)
A GPU is a graphics card, the part of a computer originally built for games. It happens to be very good at the maths AI needs, so it is what makes local AI fast. VRAM is the memory on that card, measured in gigabytes (GB).
Why it matters: A model has to fit inside the card's VRAM to run fast. If the model is bigger than the memory, it either refuses to load or crawls. This is why guides talk about a model needing “20 GB of VRAM” and why you sometimes have to choose which program gets the card. It is like counting seats: only so much fits at once.
Self-hosting
“Self-hosting” means running a service yourself instead of paying someone else to run it for you. The dictation server in these guides is self-hosted: it is a program on your own computer that you start, rather than a website you sign into.
Why it matters: The upside is control and privacy. The cost is that you are now the person who keeps it running, updates it and locks the door. These guides try to make that as small a job as possible, usually a single command to start.
Whisper and speech-to-text
Speech-to-text turns spoken words into written ones. Whisper is a well-known, free speech-to-text model. When a guide here talks about “a Whisper server”, it means a small program running that model, waiting for audio and sending back the text.
Why it matters: Dictation is faster than typing for a lot of people, and it sidesteps spelling entirely. Running Whisper yourself means your voice is transcribed on your own machine and then thrown away, rather than kept by a company.
Ready to try one?
The build-log guides put these ideas to work, one command at a time, with the reasons for each step written out.
Read the how-to guides