A General-purpose AI(GPAI) model is defined by Article 3(63) of the AI Act.
The Commission gives a concrete approach: it is the amount of computational resources that are used to train the model measured in FLOP, and the modalities of the model that will define whether a model is a GPAI or not.
A model is likely to be considered a GPAI if:
- The training compute of the model is greater than 10^23 FLOP. As explained in the Guidelines, the "amount of compute used to train a model is typically proportional to the number obtained by multiplyiing the number of its parameters with the number of its trainig examples".
- It is capable of performing a wide range of distinct tasks like generate language (text or audio), text-to-image or text-to-video. The model's training on a broad range of natural language, ability to use language to communicative, store knowledge & reason is an indicator of its significant generality.
If the first threshold is met, but the model cannot perform a wide range of distinct tasks, then it is not a GPAI.
*> For example, if it uses 10^24 FLOP but can only transcribe speech to text, it is not a GPAI because it can perform a narrow set of tasks.
If the model is general enough in its capabilities without meeting the threshold, it is still a GPAI.
While this single threshold seems easier to hold as a criterion, it is not set in stone, as the European Commission indicates that it continues to investigate other criteria.