nostril
Joined 3 years ago
Karma
66
Submissions
24
Member since
Aug 25
Recent submissions
Comment · 3 years ago
This is exactly what I was hoping for. Thanks.
I wish there was some way for us to keep in touch. There are a few things I was hoping for some thoughts on, and most of the people here don't have emails in their profiles.
Comment · 3 years ago
smspva is perfect. Thanks! Do you know of one for gift cards? Particularly ones that could be used to pay for services like server rental or Tarsnap.
Comment · 3 years ago
> What's your game? Are you profiting enough to make sense - financially-wise - to break the law?
Not at all. Hosting costs $130/mo, and I feel the sting each month. I'm not sure we'll even get enough donations to cover that, let alone have some kind of profit motive. But we wouldn't want to profit off the works anyway, or else we'd be no better than the corporations.
My game is to help people like you be able to train your own models. If I don't help you, who will? Companies will have the final say in what you're allowed to do on your own hardware, because they control the data. No data, no training.
The hard part is to balance this with doing the right thing. I'd like to figure out the right thing from first principles and by asking thoughtful people like you, rather than from fear of consequences.
As for consequences, we're being careful enough that it seems worth the risk. (You can read more about our precautions at https://news.ycombinator.com/item?id=37346620.) But I agree that staying out of jail is preferable to being in one.
Comment · 3 years ago
Could you elaborate a little more on how you came to this belief? I'm interested in the process of deciding whether to agree or disagree. A good way to get better at that is to get perspectives from thoughtful people.
Comment · 3 years ago
We're not worried. Our operation is anonymous, and as long as we don't slip up, we'll be fine. Though saying "don't slip up" is very "draw the rest of the owl"; it's most of the work: https://news.ycombinator.com/item?id=37346620
But we'd like to do the right thing ethically, which is hard to figure out.
Hypothetically, if you were going to set up a process for yourself outside of the law, what criteria would you use?
Comment · 3 years ago
This is an excellent point. Thanks. Do you mind if I quote your comment in our official policies?
It's interesting because libgen also provides most fiction titles, but everyone is rooting for them.
For example, one of the books they want taken offline is from 1954, republished in 2008. So in this case they operate closer to the textbook model than the author model.
Comment · 3 years ago
As far as I can tell, Books3 seems to be training data for language models. I'm not sure how it was created, but it contains a lot of books. I got it by torrenting The Pile after it was forced offline by the Danes.
They're asking to remove 109 books from the dataset, which I can do. But I'm not sure whether to. Once you set aside the question of law, it becomes a matter of ethics, and these questions aren't so easy.
Comment · 3 years ago
One other question: Is there legal basis in Denmark for asking for proof that they control the copyright for the listed works? DMCAs operate on "good-faith belief, under penalty of perjury" but in international situations it becomes trickier.
If anyone knows of a Danish lawyer I could consult with, or someone versed in international affairs, please let me know. (Or if you care to contribute funding. Hosting costs around $140/mo right now, which isn't free, but paying for consultation is costlier.)
Comment · 3 years ago
Hi all. I would like your help with an ethics problem.
I run a site called The Nose, a safe haven for AI training data. It operates overseas in a region out of reach of DMCAs. (Past info: https://news.ycombinator.com/item?id=37512147)
This was necessary because I felt it was unacceptable for entire datasets to be forced offline by one lawyer.
The ethical problem is that I'm sympathetic with people who want to remove their content from AI training data.
I received an email from the Danish Rights Alliance about Books3: https://pastebin.com/6qw3yMWZ
They point out that this is illegal in Denmark and elsewhere, and threaten to ban thenose.cc from Denmark.
Obviously, the threats are meaningless. But I'm interested in your views on whether we should comply with the request by removing the specific titles they list.
I was thinking of saying "If you say 'please', I will remove the listed titles." There are 109 entries, so it wouldn't be too much hassle to just remove those from the tarball, and it would be amusing to force a lawyer to ask nicely.
For now, I asked for a complete list of the full filenames they want to be removed, along with proof that they represent the listed rightsholders.
I'm more interested in how you feel. It seems reasonable to let people opt out of training. We could formalize this process by setting up a way to do this. We could also just ignore takedown demands.
What do you think?
Comment · 3 years ago
I think if you're interfacing with your server without going through Whonix, you're asking for trouble. Not only do you need to pay for the server using BTC that can't be traced back to your identity, but anything that touches the server (such as your server you're proxying with) needs to take the same precautions, which means no DigitalOcean, unless you can somehow pay them without that also being tied to your identity.
If you're not actually worried that DMCA people will follow through on their threat to sue you, or you really want to risk losing your property in the event of a lawsuit, then perhaps this might work.
Feel free to email me for more advice or to keep in touch. Your project sounds interesting.
Comment · 3 years ago
Indeed. For anyone who isn't convinced, I wrote up some details on our use case (creating a training data DMCA safe haven) in the Tails thread: https://news.ycombinator.com/item?id=37512147
If you're serious about protecting yourself, Whonix is a requirement.
Comment · 3 years ago
Other thoughts:
Day to day browsing is a pain. I use a VNC client to remote into our server, which is running a desktop environment with a regular browser. That way you can use apps (gmail, discord, etc) from outside the Tor network. But since you're tunneling through Tor, this is painfully slow. You'll likely want to type out long messages in Whonix, then copy-paste into your remote session. Each keystroke can sometimes take a full second to appear when animations are heavy.
Transferring large amounts of data is also painful. If you try to start Litecoin Core on Whonix, you'll need to sync more than 30 GB, which can take a very long time.
Patience is your weapon. You have all the time in the world not to make a mistake, and moments to make a fatal one. Think carefully about everything you do.
Stylometry scares me. AI can help here: run an assistant locally, and ask it to reword everything you write. You won't be able to use ChatGPT for this, obviously because OpenAI retains a history of everything you submit, but also because they require a real phone number to sign up. And you can't get a real number through any means I've found so far.
Payment is also a pain. I'm hoping to ask the community to donate Vanilla gift cards so that I can sign up for Tarsnap or spin up a droplet.
By applying the discipline normally found in aeronautics, I think it's possible to do this safely. But you'll still be risking jail time, and the intersection of people who want to do something for altruistic reasons and willing to risk prison is pretty small. I'll be documenting everything I do so that you can learn from my example, or perhaps from my mistakes.
Comment · 3 years ago
Hi. We're building The Nose (https://thenose.cc), a safe haven for training data that can't be taken down with DMCA. Since this involves copyright infringement, strong anonymity is a requirement.
I wrote up our security procedures here: https://news.ycombinator.com/item?id=37346620
The reason Tails isn't an option is because, as others have mentioned, there have been Tor browser exploits which reveal the IP address of the Tails user. While this is unlikely for our case, it's important to approach security from first principles with threat modeling. An attack from the FBI may seem unlikely today, but both Silk Road and one of its successors were taken down by mistakes they made when setting up their site. Learning from history, if you're not careful early, you're in for a surprise later.
Case in point: When I started Whonix Workstation to post this comment, the Whonix Gateway VM failed to boot. So when I tried to start Tor Browser and go to https://news.ycombinator.com, all I saw was a connection error. This kind of layered defense is essential if you're serious about staying out of jail.
Realistically, you'll likely dox yourself through some other means: sending Bitcoin to your pseudonym from your real identity, admitting to someone you know that you control your pseudonym (this work gets lonely, so this is a real temptation), or even accidentally signing off an email with "Thanks, [your real name]". And once you make a single mistake, you can never recover.
Comment · 3 years ago
If anyone wants to download Books3, it seems to be available here: https://thenose.cc/public/AI/EleutherAI_ThePile_v1/pile_prel...
(We noticed The Pile was recently taken offline, so we hosted it: https://thenose.cc. Apparently Books3 was also a part of The Pile, so feel free to download.)
Comment · 3 years ago
Oh, thank you. Basically AI training datasets have been knocked offline recently by DMCAs, and the goal is to bring them back online in a place that can't be knocked offline. The most popular training dataset was The Pile, hosted by The Eye: https://pile.eleuther.ai/
Notice the links now 404. We tried to make a drop-in replacement for those links. All they have to do is change the-eye.eu to thenose.cc in the urls.
Unfortunately there's not a lot of ways to get their attention to let them know this exists now. I'll try emailing the contact address but I imagine they receive lots of spam, so I was hoping to try to get noticed by people like yourself first. Maybe a direct email is still the best way, but there's no guarantee they'll even be willing to change the urls due to legal risks. For all they know I could be logging the IP address of everyone who downloads it and forwarding it to authorities. But I'm not, and it's a frustrating problem to try to solve. I just want to help AI flourish.
This also serves as a template for someone else to do the same thing, so at least there can be multiple mirrors.
Thank you again. The fact that you even took the time to look it over meant a lot. If you have any other ideas, I'd be interested to hear.
about: https://news.ycombinator.com/item?id=37346620
email: [email protected]
twitter: @TheNoseCC