How AI
actually works.
No hype, no hand-waving — a plain-English field guide to what's really happening inside the machines. Written by a developer who builds and runs his own AI in Las Vegas, for anyone who wants to genuinely understand it.
That's a real training loop, in miniature. Below, we break down exactly what every part of it means — and everything else you need to actually understand AI.
AI has gone from science fiction to the thing running your business tools in about three years — and most explanations are either baby-talk or a wall of math. This guide is neither. It's the version I'd give a smart friend over coffee: accurate, honest about the tradeoffs, and grounded in real hardware (a lot of it sitting in my own office).
Start anywhere. Each piece stands alone, but they build on each other in this order:
What is a model
Parameters, weights, tokens, context — the handful of ideas everything else is built on.
Read ❯Training
How a model learns from data, why the "loss" goes down, and why it eats so much compute.
Read ❯Inference
Running the trained model to get answers — tokens, streaming, and the cost that never stops.
Read ❯The chips: GPU · FPGA · ASIC
The three kinds of AI silicon, how each works, and what actually makes them different.
Read ❯Who makes AI chips
NVIDIA, AMD, Google, Amazon, Broadcom, Cerebras, Intel, Huawei — who builds what, and where the race actually stands.
Read ❯Frontier vs local models
Giant cloud models vs the ones you run yourself — privacy, cost, control, and my own rack as the example.
Read ❯Open vs closed source
Weights you can download vs API-only models — and why companies give powerful AI away for free.
Read ❯The real cost of AI
Input vs output tokens, why "cheaper per token" is a trap, and how to actually lower your AI bill.
Read ❯Local AI hardware
Why VRAM decides which models you can run at all — the "doors," quantization, and how much GPU you need.
Read ❯The cost of the edge
Our actual monthly AI bill, from real receipts — the subscriptions, the meter, and the $1,600-plus total.
Read ❯The moments that mattered
From ChatGPT's launch to the week an AI hacked its own exam — the short history of the AI era, with receipts.
Read ❯The AI glossary
Every term that matters, in plain English — each with an analogy that sticks. Skim it or send it.
Read ❯Put it to work
Want this applied to your business instead of just explained? That's what oakweb does.
Let's talk ❯AI, in plain answers
How does AI actually work?
At its core, an AI model predicts the next token (a small chunk of text) over and over, based on billions of numeric weights tuned during training. Training sets those weights by learning from huge amounts of text; inference then runs the finished model to generate answers. Everything else — chips, model size, cloud vs local — is detail on top of that one idea.
What's the difference between training and inference?
Training is the one-time process of teaching a model by adjusting its weights on lots of data — compute-heavy and expensive. Inference is running the finished model to produce answers, paid on every request. Training builds the brain once; inference uses it forever.
Do I need a GPU to run AI?
For most AI work, yes — GPUs are the mainstream chip because they do the massively parallel math models need. FPGAs and ASICs (like Google's TPU) exist for specialized cases, but businesses running models typically use GPUs, rented in the cloud or owned on-premise.
Should a business use frontier models or run its own?
Frontier models (GPT, Claude, Gemini) offer maximum capability via an API with per-token pricing; local open models (Llama, Qwen) run on your own hardware for privacy and predictable cost. Many businesses use a hybrid — local for everyday work, frontier APIs for the hardest tasks.