Knowledge Base

Glossary

Every guide on this site has a short "Words you'll see in this guide" box. This page collects all of those words in one place, in plain English. New terms get added as new guides go up.

3-2-1 backup: Three copies of your data, on two kinds of storage, with one copy somewhere other than your house.
API: A way for one program to send requests to another, often over the internet.
ARM / x86: Two families of processor. Raspberry Pis use ARM; most PCs use x86. Some software only runs on one.
Barebones: A computer sold without RAM or storage, so you add your own.
Bay: A slot for one hard drive. A 2-bay NAS holds two drives.
BIOS / UEFI: The settings screen built into a motherboard that runs before the operating system starts.
Bootable USB: A USB stick set up so a computer can start from it instead of its own drive.
Cloudflare Tunnel: A free service that makes an app running at home reachable from anywhere, without opening ports on your router.
CMOS: A tiny bit of memory that stores your BIOS settings. Clearing it resets them.
CMOS battery: A coin battery on the motherboard that keeps the clock and settings when the PC is unplugged.
CMR / SMR: Two ways hard drives store data. CMR is the one you want in a NAS. SMR is cheaper and can be very slow when rebuilding.
Codec: The format a video is compressed in, like H.264, HEVC (H.265), or AV1.
Command line (CLI): A text window where you type commands instead of clicking buttons.
Container: A sealed box that holds an app and everything the app needs to run.
Context length: How much text the model can keep in mind at once, measured in tokens. Longer context uses more memory.
ControlNet: An add-on that lets you guide an image's layout or pose with a sketch or reference picture.
Cores: The separate workers inside a processor. More cores means more jobs at once.
CPU: The processor, the main chip that runs everything. It can run AI models too, just slower than a graphics card.
CUDA: Nvidia's software for running AI on its graphics cards. Most local AI tools support it first and best.
DDR4 / DDR5: The two common types of RAM. DDR5 is the newer one, and they don't fit each other's slots.
Debug lights: Small labeled lights on the motherboard that show which part is stopping the startup.
Desktop environment: The look and layout of the screen, like the taskbar, menus, and windows. Cinnamon and Xfce are examples.
DHCP reservation: A router setting that gives one device the same network address every time.
Direct play: Your TV or phone plays the video file as it is, with no converting needed.
Distilled model: A smaller model trained to copy how a bigger one answers. Much easier to run, a bit less capable.
Distro: Short for distribution. One packaged version of Linux, like Linux Mint or Ubuntu.
DNS: The internet's phone book. It turns website names into addresses your device can reach.
Docker: A tool that runs each app in its own sealed box, so apps don't break each other or your system.
Dongle: A small stick that plugs into a USB port and adds a feature, here a radio for Zigbee or Z-Wave.
DRM: Copy protection that streaming services use. It limits which devices can play their shows.
ECC memory: RAM that catches and fixes small memory errors. Common in business servers.
eMMC: Storage soldered onto a board. Slower than an SSD, and much more reliable than an SD card.
End-to-end encrypted: Your data is locked before it leaves your device, so the company storing it can't read it.
Fine-tuning: Training an existing model further on your own examples so it gets better at a specific job.
FP16: Full precision, the uncompressed version of a model. Best quality, most memory.
Frontier model: The newest, most capable cloud models from the big AI companies.
1GbE / 2.5GbE / 10GbE: Network speeds. Most home networking is 1GbE; 2.5GbE is about 2.5 times faster, and 10GbE about 10 times.
GPIO pins: The pins on a Raspberry Pi that connect to sensors, lights, and other electronics.
GPU: The graphics card. It does the heavy math for AI models much faster than a processor.
Hallucination: When an AI states something false as if it were true.
Hardware acceleration: Letting a dedicated chip do a job, like video conversion, instead of the main processor.
HAT: An add-on board that sits on top of a Raspberry Pi, like the one that holds an NVMe SSD.
HBA: A card that adds more drive connections to a computer.
HDMI-CEC: A TV feature that lets your normal TV remote control devices plugged in over HDMI.
HDR tone mapping: Adjusting an HDR video's colors and brightness so it looks right on a regular screen.
Headless: Running a computer or program with no screen or window, controlled from another device or app.
Home Assistant: Free software that controls smart home devices from many brands in one place, running on your own hardware.
Home lab: A computer or two at home that you use to run your own services and learn on.
Hypervisor: Software whose job is running other operating systems as virtual machines.
Idle power: How much electricity a computer uses when it's on but not busy. For an always-on server, this number drives your power bill.
iGPU: Graphics built into the main processor, with no separate graphics card. Quick Sync lives here.
Integrated graphics: The basic graphics built into many processors, as opposed to a separate graphics card. It shares system RAM.
ISO file: A single file that holds a whole operating system installer.
Kilowatt-hour (kWh): 1,000 watts used for one hour. It's the unit on your electricity bill.
Kodi: A free media center app that organizes and plays your videos and music on a TV.
KV cache: The model's working memory for your conversation. It grows the longer the conversation gets.
LibreELEC: A tiny operating system whose only job is running Kodi.
Live session: Linux running straight off a USB stick, so you can try it without installing.
llama.cpp: The free engine that actually runs the model. Both apps are built on it.
LLM (large language model): Large language model, the kind of AI behind ChatGPT and Claude. A local LLM runs on your own computer.
LoRA: A small add-on file that teaches an image model a style or subject. Each one uses a bit more memory.
LTS: Long-term support. A release that gets security updates for years, so you can install it and leave it.
LXC container: A lighter-weight alternative to a full virtual machine that Proxmox can also run.
Memory bandwidth: How fast a graphics card can read its own memory, in GB per second. For local AI it mostly decides how fast a model writes, because the card reads through the whole model for every word.
Memory training: A first-boot process where the motherboard tests and tunes new RAM. It can take several minutes.
Mesh VPN: A private network where each of your devices connects directly to the others, like Tailscale.
Mixture-of-experts (MoE): A model design that only uses part of itself for each word it writes, so it runs faster than its total size suggests. The whole model still has to be loaded.
Model: One specific version of an AI. Claude comes in several models of different sizes and speeds.
NAS: Network-attached storage. A box of drives on your home network that every device can save to.
NIC: A network card, the part that plugs into your network.
NVMe SSD: A small, fast storage drive that plugs straight into the motherboard. Much quicker than a hard drive or SD card.
OCR: Technology that reads the text in a scanned image so you can search it.
Off-lease: Business computers returned after a few years and resold used, usually cheaply.
Offloading: Running part of a model on the graphics card and the rest on the processor and system RAM. It works, but slowly.
Open source: Software whose code is public, so anyone can check it, use it, and usually run it for free.
Open-box: An item someone bought and returned, checked and resold for less. Amazon sells these in Amazon Resale with a condition grade like Like New or Very Good.
Parameters (7B, 14B, 70B): A rough measure of a model's size. "8B" means 8 billion. Bigger models are usually smarter and need more memory.
PCIe: A high-speed connection inside a computer for fast add-ons like SSDs. The Pi 5 has one; the Pi 4 doesn't.
Pi-hole: A network-wide ad blocker that runs on a small computer in your home.
Platform: The processor, motherboard, and RAM type that have to match each other.
Plug-in energy meter: A small device between the wall and a plug that shows how many watts something really uses.
Port: A numbered door on a computer that one app listens on, so several apps can share one address.
Port forwarding: Telling your router to let outside internet traffic reach a device in your home.
Proxmox: Free software that splits one computer into several virtual computers.
PSU: The power supply, the box inside the case that turns wall power into power every part can use.
Quantization: Shrinking a model by storing its numbers less precisely, so it fits in less memory with a small drop in quality. Think of saving a photo as a smaller JPEG. "4-bit" (Q4) is the common setting.
Quick Sync: A video engine built into many Intel chips that converts video fast and with little power.
RAG: Letting a model look things up in your own documents before it answers.
RAID: Storing data across several drives so one can fail without losing anything. It protects against a dead drive, and it isn't a backup.
RAM: Your computer's main working memory. More RAM lets it run more apps at once. AI models can run from it too, just slower than from VRAM.
RAM kit: Two or more RAM sticks sold together and tested to work as a pair.
Ransomware: Malicious software that locks your files and demands payment to unlock them.
Rate limit: A cap on how much you can use a service in a day or hour.
Rebuild: Copying data onto a replacement drive after one in a RAID set fails.
Refurbished: Used gear that a seller has checked, cleaned, and resold, often with a short warranty.
Renewed: A used device that's been checked and resold, often with a short warranty.
SaaS: Software as a service. An app you rent by subscription, like Dropbox or Google Photos.
Secure Boot: A security setting that only lets approved operating systems start. It sometimes blocks Linux installers.
Self-hosting: Running an app on your own hardware at home instead of paying a company to run it for you.
Serverless: A rental setup that only charges you while a request is actually running.
SFP+: A type of network port used for fast 10GbE connections, common in business gear.
Single-core speed: How fast one core works on its own. This is what keeps a Minecraft world from lagging.
Snapshot: A saved copy of a whole system at one moment, so you can roll back if something breaks.
Socket: The spot on the motherboard where the CPU sits. The CPU and board must use the same type.
SODIMM / DIMM: The two sizes of RAM stick. SODIMM is the small laptop and mini PC size; DIMM is the longer desktop size.
Spot instance: A cheaper rental that the provider can take back at short notice when someone pays full price.
SSD: A solid-state drive. Storage with no moving parts, much faster than an old spinning hard drive.
SSD wear level: How much of an SSD's rated lifespan has been used up.
SSH: A way to control another computer by typing commands from your own, over the network.
Tailscale: A free app that lets your own devices reach each other securely from anywhere, without opening your home network to the internet.
TDP: Roughly how much power and heat a processor is designed for. Lower means cooler and quieter.
Telemetry: Anonymous usage stats an app sends back to its makers, like how often it crashes.
Thermal paste: A thin layer between the processor and its cooler that helps heat move away. It dries out over the years.
Thermal throttling: A chip slowing itself down when it gets too hot.
Tick: One update of the game world. Minecraft aims for 20 per second, and lag starts when the server can't keep up.
Tokens per second: How fast a model writes. A token is roughly three-quarters of a word.
Training cutoff: The date a model's built-in knowledge stops.
Transcoding: Converting a video on the fly so it plays smoothly on a phone or TV that can't handle the original file.
Unified memory: Apple's design where the processor and graphics share one fast pool of memory, so a Mac's RAM works like VRAM.
UPS: A battery backup that keeps your server running through short power cuts and lets it shut down safely.
vCPU: A share of a processor on a rented server. More vCPUs means more processing power.
Virtual machine (VM): A complete computer simulated in software, running inside your real one.
VPN: A private, encrypted tunnel between devices over the internet.
VPS: A small rented server in a data center that runs around the clock. You control it like your own computer.
VRAM: The memory on your graphics card. AI models load into it, so more VRAM means you can run bigger models.
Watt (W): A measure of how much power something uses at a given moment.
XMP / EXPO: A one-click setting that runs your RAM at the speed printed on the box.
Zigbee / Z-Wave / Thread: Low-power wireless languages that smart sensors, bulbs, and locks use to talk to your hub.

Stuck on a word that isn't here? It probably should be. The guides in the Knowledge Base explain most of these in more depth.