Summary:
Les nouvelles directives sur les obligations des modèles d’IA à usage général (GPAI) de la Commission mettent l’accent sur la nécessité de se conformer aux exigences, sous peine de sanctions financières. Le GPAI est défini comme un modèle capable de réaliser une large gamme de tâches distinctes. Les modèles en open source doivent respecter certaines obligations de transparence, tandis que la mise en application des dispositions GPAI débutera en août. Il est important de se préparer pour rester conforme au règlement.
Original Link:
Original Article:
🤖🧮 Remember how I had to recall my book draft because the Code of Practice dropped two minutes after hitting “Send”? Well, guess what.
The 34 pages of Guidelines on General-Purpose AI model obligations carry a little more weight than the other ones, because on GPAI models the Commission is the enforcer (art. 88). Thus, deviate from this guidance at your financial peril.
🔍 First: What counts as a “general-purpose” AI model?
The AI Act defines GPAI as “display[ing] significant generality and is capable of competently performing a wide range of distinct tasks”. But what is “generality” and “wide range”? Is lyrics-to-music narrow or general, would that depend on the number of genres & instruments you support? I get what the law intends but it’s fundamentally unclear.
The Guidelines take the same simplified approach as the Act does for “systemic risk” GPAI. If your total training compute exceeds 10²³ FLOPs – and you’re capable of generating language, images, or video – you are GPAI. (10²³ FLOPs is approximately what you need for a 1 billion parameter model, recital 98).
🧪 Second: What if you’re fine-tuning someone else’s model?
Few have the capacity to build a 10²³ FLOP model themselves. Usually you’d fine-tune an existing model. If that took a third of the original compute, congrats, you’re a GPAI provider. Note that your transparency requirements are limited to your modifications (recital 109).
You can use the original provider’s FLOP reporting (mandatory, Annex XI.1(2)(d) AIA) or base yourself on 10²³ for non-systemic-risk or 10²⁵ if the base model has systemic risk.
🧵 Third: What if your model is open source?
OSS models are exempt from the technical and API docs requirement, although they still must provide copyright policy & training disclosure (art. 53(2)). The Commission understands what “OSS” means – free as in speech, not free as in beer. No license restrictions on downstream usage (except as necessary to protect fundamental rights). Asking money or personal data for access to a model means it can’t be OSS.
🗓️ Fourth and final: when does this start to apply?
GPAI provisions start to apply August 2, but the Commission’s powers aren’t available until next year August. Also, models released prior to Aug 2 have until Aug 27 (art. 111(2)). Grumblingly, the Guidelines note the AI Office is “dedicated to supporting providers in taking the necessary steps to comply” until August 26, after which it will “enforce with fines”.
If you’re already working on a GPAI model, you can claim “disproportionate burden” for training work done prior to August 2 (with docs & discussion with AI Office, of course). And if it’s your first model (aww!), you will get “particular consideration to [your] challenging situation”.
🏆 Guidelines & Code of Practice
The CoP is the real prize: comply & be presumed AIA compliant (53(4)). The Guidelines are helpful to determine your starting position. So let’s get ready.
(And yes, updating my book as I type this.)