Since the full training process is documented and reproducible, researchers and watchdogs can audit the data sources, verify compliance with data protection laws, and inspect how the model was trained. Apertus’ development explicitly adhered to Swiss data protection and copyright laws, and incorporated retroactive opt-out mechanisms to respect data source preferences.
From a privacy perspective, Apertus represents a compelling shift in the AI landscape. The model only uses publicly available data, filtered to exclude personal information and to honor opt-out signals from content sources. This not only aligns with emerging regulatory frameworks like the EU AI Act, but also provides a tangible example of how AI can be both powerful and privacy-respecting.
I’m still very skeptical about AI, even beyond its privacy issues, but I wonder if Apertus is more ethical than Proton’s Lumo.
There are a lot of writers whose prose I admire. Many of them have blogs or wrote for free online publications. If AI is trained on their public work, isn’t it stealing? If AI can write in the style of a specific blogger, don’t they deserve to be compensated?
Lumo is essentially Mistral – a French startup, but with a better data retention and privacy policy.
Apertus is worlds apart from Mistral as it is built entirely from scratch and supposedly is trained from ethical sources. I think the main issue going forward is whether online bloggers would need to implement opt-out signals in the future. Not sure how existing work can be protected though
LLMs aren’t really going away but there is an opportunity to see how online bloggers can better protect their work even if it’s public-facing.
Yeah, that’s the question, and it applies to basically any writing you can find on the Internet for free. It’s one thing to use AI to find out facts about the world. It’s another when you ask it to write an essay or an article in the style of your favorite writer.
I think its highly risky and too fast. There are websites similar to fightchatcontrol for contacting politicians, but about AI: ControlAI
But the difference is that fightchatcontrol has clear actions to take - vote against, while ControlAI one can express grave concern, but unclear what politicians can do exactly.
Transparency of “Open Source AI” seems to be a more positive thing, as AI developed in this way would grow slower and more ‘ethical’ because of transparency.
But some politicians in the world have a lot more influence. This is great documentary of political battles that are happening in California state https://inv.nadeko.net/watch?v=JQ8zhrsLxhI (“SB-1047: The Battle For The Future Of AI - Full Documentary“)
If you watch this documentary, notice how supposedly “open source” strongly oppose regulation. But they don’t say what that “open source” is exactly.
There are some like Meta corporation or venture capitalist Marc Andreessen - Wikipedia who talk in open source name…
#aisafety (testing if tags on here appear on mastodon:)
Thanks for this. Just started following them on socials. For some reason, the link to their YouTube channel doesn’t work. Has it been nuked?
I still think there’s a lot they can do with the right guidance. I agree with Cory Doctorow when he said the following in a recent article:
CORY DOCTOROW:I don’t think that it’s impossible for politicians, even nontechnical politicians, to make good tech policy. After all, the fact that no one in Congress is a microbiologist doesn’t stop federal standards from delivering potable water
The problem is, they are too seduced by the deceptive promises of capitalism.
Yes, it is better, but I suspect that most businesses including respectable privacy companies like Proton subscribe to Facebook’s motto of move fast and break things. At least to some degree.
If doing things right means arriving to the market 5+ years after the big players settle themselves and take over the world, some privacy companies are not going to be ok with that. The fact that privacy company like Skiff allowed themselves to be bought by a non-privacy like Notion, tells you that that temptations is very real, and applies to everyone, even if they haven’t given into it yet.
As in this documentary, they are advised by experts, but are lobbied by other powerful interests (for money and power), and still law failed to be adopted in a very weird way:
We’ll look into possibly integrating this model as well at some point, thanks for the feedback.
For comparison, current Lumo model which is OpenAI’s (I guess) open model https://openai.com/open-models/ which is 120B in size and trained with proprietary state-of-the-art quality dataset.
Compared to smaller Apertus 70B trained with new open dataset.
So, current Lumo has highest quality, but Apertus open model should be given weight for openness reasons somehow as well.
P. S. And Proton CEO comment
It will depend on the performance, which still has to be assessed. If it’s good, we will use it.
Spoiler: quality won’t be good considering the size and dataset.
Yeah, it’s really hard to fight lobbies. In a recent episode of Firewalls Don’t Stop Dragons, Carey interviewed Monique Priestley, a Vermont House Representative who is a privacy advocate. She goes into great detail about how hard it actually is to pass privacy legislation in the US, one of the issues being powerful lobbies who go against her. It’s very inside baseball.
Thanks. Don’t know if you’re aware of Ed Zitron, but he’s another prominent AI skeptic. He’s also got a podcast called Better Offline. A good introduction to his positions is his first interview on Adam Conover’s show from last year.
How democracy is run have to change, so citizens would have influence, and business interests less
OpenAI released their open model in August, and after Proton released next Lumo version 1.1 Introducing Lumo 1.1 for faster, advanced reasoning | Proton with vastly better quality (I tested both versions), but information on this page about models did not change (I checked archives). This page is outdated.
FWIW, in countries with no real rule of law or real intellectual property laws to enforce, this is already understood. So if you make something where you list ingredients, you leave 1 or 2 out and add in 1 red herring.
