From our other sites
The story
EUROPA, an AI model developed by French AI company Domyn with EU backing, is reportedly aiming for nearly 1 trillion parameters, sparking discussion on the overseas forum Lemmy’s c/europe community. While attention focused on the EU touting this as an “open” AI model, the thread’s first reaction was suspicion about where the training data actually came from — commenters suggested it likely relies on scraped or pirated data, just like other major models. The weights (the model’s actual data) haven’t been released either, and Domyn hasn’t given a firm release date; a June report said Domyn was aiming for release “within the year.” Discussion also turned to just how hard it would be to run a trillion-parameter-class model on personal hardware.
EUROPA, the EU’s “open AI” model, is aiming for nearly 1 trillion parameters
Source: maccus.fr / Read the original article here
What people said
Background and Key Points of This Discussion
To start, the thread never really settled on what “open” even means here. “Open weight” — releasing only the weights (the parameters themselves) — is different from “open source,” which includes the training code and data too, and many so-called “open” AI models, including Meta’s Llama series, are in practice only open-weight. EUROPA is no different: developer Domyn still hasn’t released the weights and hasn’t even committed to a release date. On top of that, a trillion-parameter-class model simply isn’t something you can run on a personal PC — and even if you could, it would take specialized, expensive hardware. The rule of thumb raised in the thread — “70–80B parameters on 64GB of RAM” — already assumes compression via quantization, and that’s still nowhere close to 1 trillion parameters. As for where the training data comes from, lawsuits over scraping web data without copyright holders’ permission have been piling up across multiple countries, and the underlying point being made is this: even if the EU builds its own “open” model, if it’s using the same methods as everyone else, it’s hard to see how it’s really any different from the major models that came before it.
*This article is translated and excerpted by our editorial team from the overseas forum thread on Lemmy c/europe, “EUROPA, the EU’s open AI model, is aiming for nearly 1 trillion parameters.” Our site takes responsibility for the accuracy of this translation.
Leave a Reply