Potential partners are one thing. Replacing mothers, sisters, and female family members tends to be more difficult to manage.
503 karma · joined March 4, 2017
By definition, half the code out there is of below median quality. Whether or not half the code out there is of quality below the arithmetic mean depends on assumptions about the distribution of code quality. I would suspect that much more than half the code out there is "below average", so to speak.
The first was among the most brilliant people I've ever met. Sure, this person was a full-ride honor student with straight As all around, but they never found the challenge they needed, even when sitting in on grad-level classes for fun. They hardly ever studied, spent most of their time playing video games, exclusively took classes in their areas of interest, and transferred to another school when those classes ran out. Repeat for at least three colleges, taking over a decade to finally cave in and fulfill their undergrad requirements. Then after school, their career never really took off.
The second was, to put it politely, not the sharpest tool in the shed. They studied furiously to master material that seemed trivial, and whenever I tried to discuss advanced topics with them, they were completely lost. But they worked hard, harder than anyone else I knew, to maintain a perfect GPA. Went on to have a very successful career.
I suppose I bring them up to point out that even academic success doesn't paint a full picture.
My own academic performance in college was wildly inconsistent. Started off with scholarships, Dean's list, etc., then plunged into failures and academic warnings (barely avoided probation). I recovered, hit Dean's list again, and then back into failure mode. Then another last string of As during my senior-level courses brought me back into a low 3.x GPA at graduation. Regardless of the recoveries, damage was done, costing me intern opportunities that might have helped out a bit when graduating into a recession.
Depending on my mood, I could place the blame for my failures on a lot of things. A few different medical conditions ultimately requiring surgeries and months of prescribed opiates. Lack of parental support. Depression preceding my downturns. Undiagnosed (at the time) ADHD. But in retrospect, I also could have been smarter in how I dealt with all those things instead of embracing the self-destructive aspects.
No, not talking about the copyright holder, I'm talking about the hypothetical individual(s) creating infringing copies. If those people need to already have a copy of the image to extract a copy of the image from the generative image model, then I'm saying the argument that the model itself is infringing seems weak. Or it's at least not an open-and-shut case.
>> A sufficiently advanced model could, in theory, generate any image. You could then, again in theory, find an embedding for any image. Does said model then infringe on all copyrighted images?
> Without the slightest doubt.
I'll continue to argue otherwise. This proposed model is not a compressed archive that reproduces a set of infringing works when decompressed. Instead, you already have to have a copy of an image to find an embedding. Otherwise, the chances of the model spitting out copies of infringing works is exceedingly improbable. (A program that outputs random noise also has a vastly improbable chance of spitting out a copyrighted work, but that's hardly keeping copyright holders up a night.) Furthermore, in being able to produce any image, the model is not going to contain every image, and provided a copyrighted image produced after the creation of the model, you could still find an embedding. From a copyright perspective, suing the creator of this model would be like suing someone over distributing an image of random noise, claiming that because you can find an "embedding" which produces your copyrighted work (really just the difference between the two images), the noise is infringing.
Now, if you want to sue someone for distributing an embedding into this model for infringement, that's another matter entirely. That makes perfect sense.
In reality, I acknowledge that models like Stable Diffusion are going to be a bit more muddy. There definitely is some overfitting going on, so some images are literally present. However, it's a case-by-case thing. Given the requirements (framed as an "attack" no less) for extracting those images, a particular release of SD might or might not be found to infringe. Other models, with better training and better datasets, could avoid the overfitting problem.
> The method of storing the information is pretty much irrelevant to copyright. Your link argument has been tried by pirates and it's not working too well, although it depends on the country and legislation.
Unless you agree that Pi itself is a copyright violation, I think you misunderstand. I'm not making the same "link" argument made by pirates. The index and length needed to find copyrighted embeddings in Pi is just a different encoding for the same data, similar to a compressed version of the same data, though I'm sure the Pi embedding would in fact tend to be absurdly larger than the original. Again, I'm saying that Pi isn't infringing here, but the Pi-rates with their "links" would be where the infringement happens.
A sufficiently advanced model could, in theory, generate any image. You could then, again in theory, find an embedding for any image. Does said model then infringe on all copyrighted images? A program that creates Fourier epicycle drawings could be given input that causes trademarked output. An evolutionary algorithm iterating on noise could, given metrics for an image and the right fitness function, generate infringing images. Hypothetically, and admittedly absurdly, you could extract any image in the binary expansion of Pi and share it by "just" providing an index and length. If you have to know exactly what you're looking for and have to perform a substantial amount of computation to get it, it could be argued that the act of infringement is in the effort made by the person seeking infringing content (and distributing the results) rather than whatever it is they're attempting to extract the content from.
But hey, courts don't always make sensible rulings, so who knows.
Not in the general case, no. For the study done against Stable Diffusion [1], researchers were only able to reproduce about 0.03 percent of the images tested. Those were also believed to be cases of overfitting on images which were over-represented in the training data and they're not something you'd hit upon by accident.
Generative text models seen to be more problematic, depending on the subject. Code seems especially prone to overfitting, probably due to insufficient amounts of it compared to other text sources as well as lots of copying going on between the repos the models were trained on.
[1](https://arstechnica.com/information-technology/2023/02/resea...)
Or hell, the extracted potentially infringing and over-represented material might all be pop music set to variations of Pachelbel's Canon, and I'd pay to see that lawsuit.
> …but, it turns out it is possible (1) and it does, in fact, have embedded full copies of the source images.
The _not possible_ assertion is in response to people arguing that generative image models literally copy and thus "steal" the images in their training set or that such generative models "work" by infringing on said images. The assertion is not that the models cannot possibly contain _any_ copies of training data, the assertion is that the models are generalizing and can not contain a majority of the training data.
Now, let's look at some lines from your linked article titled "Paper: Stable Diffusion 'memorizes' some images, sparking privacy concerns", which happens to lead with scare quotes around "memorizes" and already hedges with "some":
> Researchers only extracted 94 direct matches and 109 perceptual near-matches out of 350,000 high-probability-of-memorization images they tested . . . resulting in a roughly 0.03 percent memorization rate in this particular scenario.
I do not consider a 0.03% rate of full copies particularly damning. Instead, it sounds like a flaw in training which could be addressed, which the article does in fact also mention being as a matter of overfitting to images that are over-represented in the training data.
> the 160 million-image training dataset is many orders of magnitude larger than the 2GB Stable Diffusion AI model. That means any memorization that exists in the model is small, rare, and very difficult to accidentally extract.
Which side's claims do you feel this line supports?
Generative text models should, in theory, also be generalizing and not memorizing. In practice, however, it does appear far easier to get Copilot and the like to spit out exact copies of code, complete with license text. Presumably, there's just not enough code out there to sufficiently generalize from, so we're seeing overfitting as a matter of course.
It took me years and a lot of luck to claw my way into developer roles. And you know what? It wasn't me. I didn't suck. Life had just dealt me a bad hand (one of several).
Not to say everything is great now. I missed out on years' worth of promotions, investments, and retirement savings. I'll probably never achieve fu money or so much as retire early. But I'm hardly hurting financially nowadays.
Which is all to say, hey, I hear ya. Hang in there.
Doing what the company needs done can also mean not taking orders from your PO. Depending on your situation, that could manifest as ignoring them, working around them, convincing them it was all their idea, directly challenging them, or going over their head. Skullduggery. Ass-kissing if necessary. Be a hero or a martyr in the end, but at least end each day with a clear conscience.
None of that being my preference, given the choice. All I expect from a PO is a collaborative relationship where I'm treated as an equal and not some member of the worker underclass who's just there to close tickets.
When the project collapses under its own weight as a consequence of the PO's own choices, is the PO going to take the blame? Is the PO going to help when production goes down? Will they working overtime to fix things? Who suffers? The PO? Not so much.
I'm not so sure the possibility should be dismissed so easily. It's hardly a secret that CEOs gather periodically gather at various retreats. Is it really talking conspiracy theories to recognize they actually do talk business with each other and at least informally coordinate their actions? They don't all need to be at the same location just to get a few massages.