// SYSTEM ONLINE

Understand Artificial Intelligence
one term at a time.

A plain-language field guide to how AI actually works — from tokens and weights to inference and transformers. Toggle between Simple and Complex depending on how deep you want to go.

terms defined
categories
2
depth levels

01 Learning Tracks

New here, or filling gaps? Take a guided, start-to-finish tour — nuts to bolts. Pick your level and step through it like a slideshow.

02 The Glossary

Search any term, or filter by category. Every entry comes in two depths.

03 The Inference Lab

When you talk to an AI, your words make a round trip: tokenization → inference → detokenization. Pick a level of detail below and watch it happen.

Level:

04 Anatomy of a Model

A model is really just layers of simple math, stacked deep. See the big picture — or open it up to the transformer blocks inside.

Level:

05 Mixture of Experts

How a model can be huge but cheap to run: a router sends each token to only a few specialist “experts” instead of the whole network.

Level:

06 Inference Engines

The software that actually runs a model to generate text. Server engines squeeze out maximum throughput for many users; local engines make models easy to run on your own machine. Model format matters — most local tools want GGUF, most servers want safetensors.

Engine Best for Model formats Pros Cons
vLLMserver High-throughput production serving on GPUs Safetensors (HF); GPTQ / AWQ / FP8; limited GGUF Top throughput (PagedAttention + continuous batching); OpenAI-compatible API; broad model support GPU-focused; heavier to set up; overkill for single-user laptops
SGLangserver High-throughput serving with fast structured output Safetensors (HF); GPTQ / AWQ / FP8 Very high throughput (RadixAttention KV reuse); excellent for agents / JSON; OpenAI-compatible Newer; GPU-focused; smaller ecosystem than vLLM
TensorRT-LLMserver Maximum performance on NVIDIA GPUs HF / safetensors compiled into TensorRT engines Fastest on NVIDIA (FP8 / INT4, fused kernels); production-grade NVIDIA-only; complex compile/build step; less flexible
Hugging Face TGIserver Production serving tied to the Hugging Face ecosystem Safetensors (HF); GPTQ / AWQ / EETQ Instant access to HF models; continuous batching; solid tooling GPU-focused; fewer bleeding-edge perf features than vLLM
llama.cpplocal Local / edge inference almost anywhere, CPU or GPU GGUF (native) Runs on nearly any hardware (CPU, Metal, CUDA, ROCm, Vulkan); tiny footprint; huge quant selection Lower multi-user throughput; mostly text; more manual setup
Ollamalocal Dead-simple local models on any OS GGUF (own registry; can import safetensors) One-command install & run; clean local API; cross-platform; built on llama.cpp Less tunable; GGUF-centric; not for large-scale serving
LM StudioGUI Running local models through a friendly desktop app GGUF; MLX (Apple Silicon) No-code GUI + chat; built-in model browser; local OpenAI-compatible server Desktop app (not for headless servers); closed-source
Lemonadelocal Local LLMs on AMD Ryzen AI PCs (CPU / GPU / NPU) GGUF (llama.cpp); ONNX / OGA (Ryzen AI) NPU + iGPU acceleration on Ryzen AI; OpenAI-compatible server; multiple backends Best on AMD hardware; newer, smaller ecosystem
MLX-LMlocal Fast local inference on Apple Silicon Macs MLX (converted from safetensors); some GGUF Optimized for M-series unified memory; supports fine-tuning; open-source Apple Silicon only; smaller model hub than HF

// Engines and formats move fast — check each project's docs for the latest support.

07 About

Clanker University exists to make the vocabulary of AI approachable for everyone — students, professionals, the curious, and the skeptical. No gatekeeping, no assumed PhD. Just clear definitions with a depth dial you control.

Every term comes in two depths — flip any card or diagram between Simple and Complex to match what you need.