Local and sovereign AI

But transposing "open source" to AI poses an unprecedented problem. A classic piece of software is code. An AI model is three things: code, weights (the billions of learned parameters), and training data. What must be opened to be truly "open source"? The debate raged. In October 2024, the Open Source Initiative (the reference organization) settled the matter with its open source AI definition (OSAID 1.0): a model is open source if it can be freely used, studied, modified, and shared, which requires publishing the code and the weights, and providing sufficient information about the data so that anyone could, in principle, recreate an equivalent model (without necessarily publishing the raw data itself, which is sometimes sensitive or copyrighted).

This definition reveals a spectrum of openness that must be carefully distinguished:

Diagram 9.1. The spectrum of openness. Most so-called "open" models are in reality open weights: you can download and run them, but you don't know with what data and what exact code they were trained. That's useful, but it isn't quite "open source" in the strict sense.

Benefits and risks. We sketched this debate in Chapter 7; the open source angle sheds more light on it. The benefits are transparency (the code and weights can be inspected, which aids accountability and explainability), democratization (countries and companies without their own model gain access to AI cheaply), sovereignty, and the fact that an open model is harder to censor or pull overnight. The risks are symmetrical: an open model can no longer be "recalled", its guardrails can be removed through retraining, which raises concerns about malicious uses (a 2024 White House report nevertheless found no sufficient reason to restrict the publication of weights "for now"). This is the whole dilemma between openness and control, at the heart of AI governance (Chapters 24 and 25).

---

Key takeaways (Chapter 9)

• Running AI locally addresses needs for privacy, sovereignty, cost, offline operation, and control.
• Open source AI inherits from the free software movement. The OSI's definition (OSAID 1.0, 2024) requires code, weights, and sufficient information about the data.
• Beware the spectrum: most "open" models are open weights (downloadable but without data or code), which critics call "openwashing"; fully open models (like Apertus) are rare.
• Openness has become a geopolitical stake (the "DeepSeek moment", China's strategic weapon, European sovereignty).
• Quantization compresses models to run them on commonplace hardware; tools that are largely open source (Hugging Face, PyTorch, Ollama, LM Studio, llama.cpp, vLLM) make local AI accessible.
• Mac clusters (unified memory, energy efficiency) offer a good price/performance ratio for modest organizations; NVIDIA remains king for raw power.
• Digital sovereignty (sovereign models, European cloud, open source) is the "European way", ambitious but costly.

This power, local or in the cloud, has a price increasingly measured in megawatts. Chapter 10 confronts the question that could slow everything down: energy.

---