Field guide · Twelve layers · Edition 01

How AI is built and controlled

Start with silicon. Follow the chain through computation, learned weights, context and tools. Finish where the difficult question begins: who gets to decide what the system can do?

Each note below is a boundary. Understand what crosses it, what stays behind, and which trade-off appears at the next layer. Approximately twelve minutes.

By Yousef A. Salam

Twelve layers, one system

1. 01 · The substrate

   Silicon & transistors

   The physical layer where computation begins: purified sand patterned into billions of microscopic switches. A transistor is an on/off switch a few nanometres wide, and density is what makes everything above it possible.

2. 02 · Parallel compute

   The GPU

   A CPU does a few tasks fast, in sequence. A GPU does thousands of small operations at the same time. AI workloads are mostly multiplication, so the parallel machine wins — AI is multiplication at scale.

3. 03 · Learning loop

   Training

   Thousands of GPUs and months of compute turn text into a model. Text becomes tokens of roughly three-quarters of a word; the loop predicts the next token, adjusts parameters, and repeats billions of times.

4. 04 · What remains

   Model weights

   When training stops, parameters become weights: a very large file of numbers, not traditional code. A trillion parameters at two bytes each is a 2TB file. Mixture-of-Experts architectures activate only a fraction of it per token.

5. 05 · Compression

   Quantization

   Fewer bits per number make a model smaller, cheaper and faster with a deliberate precision trade-off. Two bytes down to half a byte drops the file to a quarter of its size — less precise, far easier to run.

6. 06 · Access boundary

   The registry & the split

   Open-weight models — Qwen, Llama, DeepSeek — can be downloaded and owned. Closed models are reached through someone else's API. This split decides whether self-hosting is even on the table, which is a procurement and data-residency question before it is a technical one.

7. 07 · Serving

   Runtime & inference

   The runtime turns weights into a usable service. Ollama and LM Studio suit one user on a laptop; vLLM and SGLang suit production load, where the throughput gap becomes an order of magnitude. Choose for the workload, not the benchmark.

8. 08 · Sampling

   Temperature

   Temperature changes the randomness of word choice — not whether an answer is accurate. Low is consistent and repeatable; high is varied and creative. A low setting will repeat the same hallucination every time: consistency is not truth.

9. 09 · Interface design

   Prompt engineering

   Inputs are an interface: examples, constraints and system instructions steer the output. Zero-shot versus few-shot is a design decision; the system prompt is where the house rules live.

10. 10 · Working memory

    Context & caching

    The context window is short-term memory. Caching lets a system reuse the expensive computed state of a large prompt instead of paying for it again, which is where latency and cost are actually won.

11. 11 · Connection layer

    Model context protocol

    MCP standardises how a model discovers and uses tools, data sources and local resources: a client-server architecture that exposes resources with explicit boundaries. In a finance estate this is the layer where segregation of duties is enforced.

12. 12 · The destination

    Chatbot → agent → agentic

    A chatbot is text in, text out. An agent adds tools plus a loop, so it can change system state. Agentic is a dial, not a thing: suggest to a human, act then seek approval, or act alone and escalate exceptions.

Autonomy is set by whoever owns the process, the budget and the accountability — not by the engineer.