An AI Begins as a Computer
Before discussing tokens, neural networks, transformers, attention, or anything that appears intelligent, we need a machine.
For HAAW we will begin with one physical Linux server. Conceptually, it is not very different from an ordinary AlmaLinux machine. The components are familiar. What makes it capable of running a modern AI system is the specialized hardware and software added to that familiar foundation.
Large AI systems may involve many servers, accelerators, storage systems, and networks. Those complications can wait. First, we understand one box.
Inside Our AI Machine
CPU
General-purpose processor. It helps boot the machine, runs Linux and ordinary processes, and coordinates system activity.
System RAM
Working memory used by Linux, services, applications, buffers, and CPU-side AI software.
NVMe / SSD
Persistent storage containing Linux, drivers, AI software, configuration, and the trained model files.
GPU / AI Accelerator
Specialized hardware able to perform very large numbers of mathematical operations in parallel.
GPU VRAM
High-speed memory closely associated with the accelerator. Model weights and working data may be placed here during execution.
Network Interface
Provides communication with users, services, and—later in HAAW—possibly other AI machines.
Power & Cooling
Power keeps the electronics operating; cooling removes the substantial heat produced by computation.
What Makes This an AI Machine?
At this moment, nothing does.
Imagine our machine sitting on a bench with its power switched OFF. It has a CPU, RAM, NVMe storage, a GPU, VRAM, networking, Linux, AI software, and a trained model.
But nothing intelligent is happening. The GPU is not thinking. The CPU is not thinking. The model is not executing. The trained model is simply stored information on a storage device.
What Is Stored on the Machine?
Before startup, our NVMe storage may contain:
Linux + drivers + AI runtime/software + trained model
The trained model is important, but at this stage it is data at rest. Later we will examine what a model contains, what its weights are, where they go when loaded, and how the hardware uses them.
Deployment
The AI workload must be placed into an environment in which it can run. It might execute directly on the host operating system, inside a virtual machine, inside a container, or as part of a larger managed system.
For now, the exact deployment method does not matter. The important point is that the operating environment, AI software, model, and hardware have been assembled so the model can eventually be executed.
Training creates or changes the learned parameters—the model's weights. Deployment takes an already-trained model and puts it into an environment where it can be run.
Our Starting State
The machine is powered off. Nothing has received a prompt. Nothing has generated a token. Nothing has begun inference.
Now we press the power button.
MACHINE BOOTS
HAAW Working Discipline
As we move deeper into the system, HAAW will keep different kinds of statements distinct.
- Established mechanism
- Behavior we can describe from known system operation.
- Observable behavior
- Something that can actually be measured or recorded.
- Inference or interpretation
- A conclusion drawn from what we observe.
- Open question
- Something we do not yet know or cannot directly establish.
The objective is not merely to describe what an AI appears to do. The question guiding HAAW is:
What is the AI actually doing?