The thought: the mental image of the Gibs I want. The geometry of the rosette. The proportions of the fireplace. The repeating pattern. This is me — not the machine.

The translation: the vector drawing in Inkscape. The generated G-code. Parameter configuration — cutting speed, depth of pass, tool. This is the calibration.

The execution: the RSCNC32 follows the G-code. The router carves the BA13. The rosette appears. The fireplace takes shape. The machine executed my thought on matter. It had no idea — it executed mine.

The Opération Dindon corpus: exactly the same model. The mental image of a structural study — the theses, the angle, the arguments — is Amine RAITI's. Claude receives this thought as a prompt. It processes it. It produces HTML, CSS, structured text. It had no idea. It executed mine.

The loop is closed. Hyperscalers dominate cloud because they master bare-metal. Now they dominate AI for the same reason — AI runs on bare-metal they own and their customers no longer master. AWS sold "stop worrying about servers" — then "stop worrying about AI, it's in the cloud". In both cases: the dependency is at the physical substrate level.

The Principal SRE who understands bare-metal understands AI from the inside. They know what a GPU is physically, what HBM3 memory means, why memory bandwidth is the bottleneck of a language model. The one who jumped straight to cloud without going through bare-metal understands AI as a black box. Mastery of Iron is not a nostalgic prerequisite — it is the prerequisite for AI sovereignty.

A 70-billion-parameter language model is 70 billion floating-point numbers stored in electronic memory, multiplied by electrical signals according to mathematical rules. Each multiplication is an operation in a circuit. Each circuit is made of transistors. Each transistor is an electrical switch — open or closed, 0 or 1. AI is electricity controlled by mathematics.

The chain is direct and uninterrupted:
V = R × I (year 7) → blinking LED → stepper motor → Arduino → GRBL → CNC → GPU → transformer → language model → AI response.

It is the same electron taking increasingly complex paths. The student who understood that electricity "flows like water in a pipe" has understood the foundation of everything that follows — including artificial intelligence. The Anti-Amputation Foundation does not only prepare for the CNC. It prepares to understand AI from the inside.

Level 1 — Basic calibration:
CNC: steps/mm. 200 steps = 1 motor turn. If the lead screw moves 2mm/turn, steps/mm = 100. If this parameter is wrong, all dimensions will be too. Inevitably.
AI: the base prompt. "You are an infrastructure SRE architect. You write in English. You never quote copyrighted texts. You structure responses in imperial HTML." Without this base calibration, the AI produces generic content.

Level 2 — Tool calibration:
CNC: choosing the right router bit — diameter, number of flutes, cutting angle — for this material. A wood bit in aluminium: guaranteed breakage.
AI: fine-tuning. Training the model on a specific corpus. A model calibrated on legal texts becomes a legal tool. The architecture is the same — the tool is different.

Level 3 — Project calibration:
CNC: the complete G-code file. Sequence of operations, tool changes, depths of pass, speeds. The complete translation of the designer's thought into machine instructions.
AI: the system prompt. Role instructions, format constraints, examples, exclusion rules, tone, length. The complete translation of the user's will into machine constraints. The system prompt of Claude for the Opération Dindon corpus is the equivalent of the G-code of the RSCNC32.

Each line tells the machine: what to do, where to go, at what speed, in what order. No ambiguity. No interpretation. The machine executes exactly what is written.

Each line of the system prompt tells the AI: who it is, what to produce, what it must not do, how to structure, in what order. The machine executes. It does not interpret beyond.

The formula is simple and brutal:

Well-calibrated AI + excellent thought = production of excellence at scale.
The Opération Dindon corpus is the demonstration. Twenty years of fieldwork, teaching, production incidents and reflection by Amine RAITI — calibrated in a precise system prompt — produce structural studies that neither AI alone nor the human alone would produce at this speed and with this coherence.

Poorly calibrated AI + mediocre thought = industrial-scale mediocrity.
This is what happens in 90% of current AI usage in organisations. Industrial quantities of generic text, empty PowerPoint presentations, poorly documented code, interchangeable reports. AI multiplied volume. It did not multiply quality — because quality comes from calibration, and calibration comes from the thought of the one who holds the G-code.

When AI is removed from a process that depends on it, what remains reveals what the human thought behind it was capable of producing alone. If the output collapses, it was substitution — AI was replacing thought. If the output holds, it was amplification — AI was executing thought.

The amplification / substitution distinction is the only one that matters. AI as amplification takes ideas forged over years of fieldwork and executes them at a speed impossible by hand. AI as substitution produces text in place of the human, who merely validates. The result may be stylistically correct — but it has no lived experience. It will be indistinguishable from thousands of other generated texts. It will not build lasting trust.

The student who blinked an LED at 12 understands AI at 18 — not because they studied transformers or neural networks, but because they understand the fundamental model: a machine receives instructions, processes them, and produces a result. The rest is a matter of scale and processed matter. Deamputation at the Source — the Anti-Amputation Foundation — prepares for AI far better than any decontextualised AI course.

The engineer who knows AI is a machine retains primacy over it. They calibrate. They verify. They validate. They do not delegate their thought — they delegate the execution of their thought. The one who believes AI "thinks" cedes their technical primacy. And as with the CNC: AI does not replace the craftsman. It executes what the craftsman cannot do with their hands alone.

Just as the RSCNC32 is in my space and under my physical control, a sovereign AI is an AI whose physical substrate one controls. Locally hostable open source LLMs exist:

Llama 3 (Meta): up to 70 billion parameters, hostable on a local server with 2-3 NVIDIA RTX 4090 GPUs.
Mistral / Mixtral (Mistral AI — French): MoE (Mixture of Experts) architecture, very efficient, hostable on modest hardware.
Qwen (Alibaba): multilingual, including Arabic, performant on technical tasks.
Ollama: local LLM hosting tool, one command to run a model locally.

Not the same performance as a frontier model like Claude or GPT-4 — but a thought machine whose G-code you hold. Whose physical address you know. That you can repair if it fails.