The European regulation for artificial intelligence, the AI Act, approved in March 2024, could prevent Meta from implementing the Llama 3.1 AI models, as by the law definition they could be considered "a systemic risk" .
The AI Act was designed to protect the EU consumers and citizens, but it seems that it is instead blocking EU from accessing to the full power of the AI revolution.
In the technical documentation for the Llama 3 family of models - with the 3.1 models included - Meta states what the "scale" of these models is:
"We trained a model at a much larger scale than previous Llama models: our flagship language model was pre-trained using 3.8 × 10^25 FLOPs, almost 50 times more than the largest version of Llama 2. Specifically, we pre-trained a flagship model with 405B trainable parameters on 15.6T text tokens."
And precisely that computational power used to train these models is higher than what the AI Act specifies to not consider them a systemic risk.
With these technical limitations it is impossible to train a complex AI model. Now the EU authorities will have to make a choice: to enforce the law and thus have a considerable competitive disadvantage against the rest of the world, or to change the law to allow the use of more computational power in AI training.
