• 2 Posts
Joined 1 year ago
Cake day: June 15th, 2023


  • For someone to work it out, they would have to be targeting you specifically. I would imagine that is not as common as, eg, using a database of leaked passwords to automatically try as many username-password combinations as possible. I don’t think it’s a great pattern either, but it’s probably better than what most people would do to get easy-to-remember passwords. If you string it with other patterns that are easy for you to memorize you could get a password that is decently safe in total.

    Don’t complicate it. Use a password manager. I know none of my passwords and that’s how it should be.

    A password manager isn’t really any less complicated. You’ve just out-sourced the complexity to someone else. How have you actually vetted your password manager and what’s your backup plan for when they fuck up?

  • Imagine you were asked to start speaking a new language, eg Chinese. Your brain happens to work quite differently to the rest of us. You have immense capabilities for memorization and computation but not much else. You can’t really learn Chinese with this kind of mind, but you have an idea that plays right into your strengths. You will listen to millions of conversations by real Chinese speakers and mimic their patterns. You make notes like “when one person says A, the most common response by the other person is B”, or “most often after someone says X, they follow it up with Y”. So you go into conversations with Chinese speakers and just perform these patterns. It’s all just sounds to you. You don’t recognize words and you can’t even tell from context what’s happening. If you do that well enough you are technically speaking Chinese but you will never have any intent or understanding behind what you say. That’s basically LLMs.

  • Just because something is available to view online does not mean you can do anything you want with it. Most content is automatically protected by copyright. You can use it in ways that would otherwise by illegal only if you are explicitly granted permission to do so.

    Specifically, Stack Overflow licenses any content you contribute under the CC-BY-SA 4.0 (older content is covered by other licenses that I omit for simplicity). If you read the license you will note two restrictions: attribution and “share-alike”. So if you take someone’s answer, including the code snippets, and include it in something you make, even if you change it to an extent, you have to attribute it to the original source and you have to share it with the same license. You could theoretically mirror the entire SO site’s content, as long as you used the same licenses for all of it.

    So far AI companies have simply scraped everything and argued that they don’t have to respect the original license. They argue that it is “fair use” because AI is “transformative use”. If you look at the historical usage of “transformative use” in copyright cases, their case is kind of bullshit actually. But regardless of whether it will hold up in court (and whether it should hold up in court), the reality is that AI companies are going to use everybody’s content in ways that they have not been given permission to do so.

    So for now it doesn’t matter whether our content is centralized or federated. It doesn’t matter whether SO has a deal with OpeanAI or not. SO content was almost certainly already used for ChatGPT. If you split it into 100s of small sites on the fediverse it would still be part of ChatGPT. As long as it’s easy to access, they will use it. Allegedly they also use torrents for input data so even if it’s not publicly viewable it’s not safe. If/when AI data sourcing is regulated and the “transformative use” argument fails in court and if the fines are big enough for the regulation to actually work, then sure the situation described in the OP will matter. But we’ll have to see if that ever happens. I’m not holding my breath, honestly.

  • No, the intent and the consequences of an action are generally taken into consideration in discussions of ethins and in legislation. Additionally, this is not just a matter of ToS. What OpenAI does is create and distribute illegitimate derivative works. They are relying on the argument that what they do is transformative use, which is not really congruent with what “transformative use” has meant historically. We will see in time what the courts have to say about this. But in any case, it will not be judged the same way as a person using a tool just to skip ads. And Revanced is different to both the above because it is a non-commercial service.

  • It’s definitely not “draconian” to make enshittification illegal. But you don’t regulate the turning-to-shit part. You regulate the part where they offer a service for free or too cheap so that they kill the competition. This is called anti-competitive and we supposedly address it already. You also regulate what an EULA can enforce and the ability of companies to change the EULA after a user has agreed to it. Again, these concepts already exist in law.

    We’ve essentially already identified these problems and we have decided that we need to address them, but we been ineffective in doing so for various reasons.

  • Humans are not generally allowed to do what AI is doing! You talk about copying someone else’s “style” because you know that “style” is not protected by copyright, but that is a false equivalence. An AI is not copying “style”, but rather every discernible pattern of its input. It is just as likely to copy Walt Disney’s drawing style as it is to copy the design of Mickey Mouse. We’ve seen countless examples of AI’s copying characters, verbatim passages of texts and snippets of code. Imagine if a person copied Mickey Mouse’s character design and they got sued for copyright infringement. Then they go to court and their defense was that they downloaded copies of the original works without permission and studied them for the sole purpose of imitating them. They would be admitting that every perceived similarity is intentional. Do you think they would not be found guilty of copyright infringement? And AI is this example taken to the extreme. It’s not just creating something similar, it is by design trying to maximize the similarity of its output to its training data. It is being the least creative that is mathematically possible. The AI’s only trick is that it threw so many stuff into its mixer of training data that you can’t generally trace the output to a specific input. But the math is clear. And while its obvious that no sane person will use a copy of Mickey Mouse just because an AI produced it, the same cannot be said for characters of lesser known works, passages from obscure books, and code snippets from small free software projects.

    In addition to the above, we allow humans to engage in potentially harmful behavior for various reasons that do not apply to AIs.

    • “Innocent until proven guilty” is fundamental to our justice systems. The same does not apply to inanimate objects. Eg a firearm is restricted because of the danger it poses even if it has not been used to shoot someone. A person is only liable for the damage they have caused, never their potential to cause it.
    • We care about peoples’ well-being. We would not ban people from enjoying art just because they might copy it because that would be sacrificing too much. However, no harm is done to an AI when it is prevented from being trained, because an AI is not a person with feelings.
    • Human behavior is complex and hard to control. A person might unintentionally copy protected elements of works when being influenced by them, but that’s hard to tell in most cases. An AI has the sole purpose of copying patterns with no other input.

    For all of the above reasons, we choose to err on the side of caution when restricting human behavior, but we have no reason to do the same for AIs, or anything inanimate.

    In summary, we do not allow humans to do what AIs are doing now and even if we did, that would not be a good argument against AI regulation.

  • I have my own backup of the git repo and I downloaded this to compare and make sure it’s not some modified (potentially malicious) copy. The most recent commit on my copy of master was dc94882c9062ab88d3d5de35dcb8731111baaea2 (4 commits behind OP’s copy). I can verify:

    • that the history up to that commit is identical in both copies
    • after that commit, OP’s copy only has changes to translation files which are functionally insignificant

    So this does look to be a legitimate copy of the source code as it appeared on github!


    • This was just a random check, I do not have any reason to be suspicious of OP personally
    • I did not check branches other than master (yet?)
    • I did not (and cannot) check the validity of anything beyond the git repo
    • You don’t have a reason to trust me more than you trust OP… It would be nice if more people independently checked and verified against their own copies.

    I will be seeding this for the foreseeable future.