training measured in floating point

There was also agreement on a “two-tier” system of guardrails to be applied to “general” AI systems, such as the so-called foundational models underpinning the viral boom in generative AI applications like ChatGPT.

As we reported earlier, the deal reached on foundational models/general purpose AIs (GPAIs) includes some transparency requirements for what co-legislators referred to as “low tier” AIs — meaning model makers must draw up technical documentation and produce (and publish) detailed summaries about the content used for training in order to support compliance with EU copyright law.

For “high-impact” GPAIs (defined as the cumulative amount of compute used for their training measured in floating point operations is greater than 10^25) with so-called “systemic risk” there are more stringent obligations.