I've noticed a trend in AI policy (e.g. the EU AI act or the USA's AI executive order) to include references to quantities of Floating -Point operations to differentiate between AI models but I don't really understand the rationale behind it or what the language actually means. For example:<p>```<p>Until such technical conditions are defined, the Secretary shall require compliance with these reporting requirements for:<p><pre><code> (i) any model that was trained using a quantity of computing power greater than 10^26 integer or floating-point operations, or using primarily biological sequence data and using a quantity of computing power greater than 10^23 integer or floating-point operations;
</code></pre>
```<p>What does this mean to me as an engineer? I normally see FLOPS (which I think of as a measurement of operations per second) but this seems to be about total operations without a time dimension? How did this language get into legislation and why does it seem so widespread?
The rationale: it's a manageable number to "easily" differentiate regulated parties.<p>How they got to this number, I've got no idea, but having reviewed training, it's easy to demonstrate whether or not your model should be regulated. It lets companies "self regulate" and as an engineer, if you build a model exceeding these parameters, you should be complying with the rules.<p>How do you know: take your hardware specs and multiply them by how long it takes you to train your model on that hardware. It's straightforward and only enforceable in hindsight.
This article has some information. <a href="https://jack-clark.net/2024/03/28/what-does-1025-versus-1026-mean/" rel="nofollow">https://jack-clark.net/2024/03/28/what-does-1025-versus-1026...</a>