The Act’s broad exemption for

[kexej28769@nongnue]

In this regard, after intense debate, EU policymakers finally decided to adopt a two-tier regulatory approach to regulate FM. The first tier requires all FM providers to undertake a series of transparency obligations and prove their compliance with copyright law, except for those that are used only for research or released under open source licenses.



The exception does not apply to the second tier, which covers models classified as high impact (or carrying systemic risk , Article 52a) according to the amount of computation used for their training ( sri lanka number data in floating point operations, or failures). According to the current text, today only models such as GPT-4 and Meta LLama-2 would find themselves falling into the second tier. While the rationale for the tiering has been criticized by parts of the scientific community, the EU legislator seems to have accepted the proportionality approach advocated by the OS ecosystem (explicit treatment of different uses and development models), and the OS community considers the compromise reached to be promising.

A free and open source AI models, as well as the adoption of the proportionality principle for SMEs (Article 60), appear to be a reasonable compromise at this stage. The latter principle provides that in cases involving model modifications or fine-tuning, the provider’s obligations should be limited to these specific changes. This could, for example, involve updating existing technical documentation to include information on modifications, including new sources of training data. This approach could successfully regulate the potential risks associated with AI technologies without stifling innovation.