![](/static/253f0d9b/assets/icons/icon-96x96.png)
![](https://lemmy.ml/pictrs/image/d3d059e3-fa3d-45af-ac93-ac894beba378.png)
13647/F/a weird anime
Cryptography nerd
13647/F/a weird anime
Because few people know what’s realistic for LLMs
Humans learn a lot through repetition, no reason to believe that LLMs wouldn’t benefit from reinforcement of higher quality information. Especially because seeing the same information in different contexts helps mapping the links between the different contexts and helps dispel incorrect assumptions. But like I said, the only viable method they have for this kind of emphasis at scale is incidental replication of more popular works in its samples. And when something is duplicated too much it overfits instead.
They need to fundamentally change big parts of how learning happens and how the algorithm learns to fix this conflict. In particular it will need a lot more “introspective” training stages to refine what it has learned, and pretty much nobody does anything even slightly similar on large models because they don’t know how, and it would be insanely expensive anyway.
Yes, but should big companies with business models designed to be exploitative be allowed to act hypocritically?
My problem isn’t with ML as such, or with learning over such large sets of works, etc, but these companies are designing their services specifically to push the people who’s works they rely on out of work.
The irony of overfitting is that both having numerous copies of common works is a problem AND removing the duplicates would be a problem. They need an understanding of what’s representative for language, etc, but the training algorithms can’t learn that on their own and it’s not feasible go have humans teach it that and also the training algorithm can’t effectively detect duplicates and “tune down” their influence to stop replicating them exactly. Also, trying to do that latter thing algorithmically will ALSO break things as it would break its understanding of stuff like standard legalese and boilerplate language, etc.
The current generation of generative ML doesn’t do what it says on the box, AND the companies running them deserve to get screwed over.
And yes I understand the risk of screwing up fair use, which is why my suggestion is not to hinder learning, but to require the companies to track copyright status of samples and inform ends users of licensing status when the system detects a sample is substantially replicated in the output. This will not hurt anybody training on public domain or fairly licensed works, nor hurt anybody who tracks authorship when crawling for samples, and will also not hurt anybody who has designed their ML system to be sufficiently transformative that it never replicates copyrighted samples. It just hurts exploitative companies.
Remember when media companies tried to sue switch manufacturers because their routers held copies of packets in RAM and argued they needed licensing for that?
https://www.eff.org/deeplinks/2006/06/yes-slashdotters-sira-really-bad
Training an AI can end up leaving copies of copyrightable segments of the originals, look up sample recover attacks. If it had worked as advertised then it would be transformative derivative works with fair use protection, but in reality it often doesn’t work that way
See also
Math and formal logic are effectively equivalent and philosophy without conditional logic is useless. Scientifically useful philosophy is just “explorative logic” or something like it
Wine/Proton on Linux occasionally beats Windows on the same hardware in gaming, because there’s inefficiencies in the original environment which isn’t getting replicated unnecessarily.
It’s not quite the same with CPU instruction translation, but the main efficiency gain from ARM is being designed to idle everything it can idle while this hasn’t been a design goal of x86 for ages. A substantial factor to efficiency is figuring out what you don’t have to do, and ARM is better suited for that.
It’s not that uncommon in specialty hardware with CPU instructions extensions for a different architecture made available specifically for translation. Some stuff can be quite efficiently translated on a normal CPU of a different architecture, some stuff needs hardware acceleration. I think Microsoft has done this on some Surface devices.
Quantum mechanics still have endless ratios which aren’t discrete. Especially ratios between stuff like wavelengths, particle states, and more
Complex numbers, and a bunch more things too
But you can’t detect such things without either server side scanning (kills E2EE dead) or client side scanning (will always be limited in what it can detect, and it’s easy to patch out of clients, AND there’s still the risk of govs maliciously pushing detection of banned media)
Not fully encrypted unless you enable lockdown mode (and losing various features)
The perceptual hash algorithm was broken in hours, then so fully broken that modified images were visually indistinguishable from unmodified images, so you could send people images with hash values that match flagged photos.
Also, then there’s the thing of the risk of various jurisdictions pushing for adding detection of other banned content.
But once a process is running its trivial to get weeks of extremely detailed history and lots of secrets you thought were ephemeral
Recall was set to be default on for everybody and to record everything in a database which is trivial to extract data from.
There’s a lot of nonsense Apple is doing too (like the chatgpt integration) but they didn’t put keylogger into the system.
And that scene where she can’t pull in the non-accelerated astronaut colleague while still being in atmosphere thin enough that he wouldn’t fall behind, so he just drifts away through magic
It doesn’t usually need to go to court if the lawyer can remind them of what laws they’re breaking
As a cryptography nerd, +100000 to that
Orange is the new black