Machines are a little different than humans (you don't say), so it's easy for them to end up doing "commercial-scale infringement" if you run your scraper for more than a minute… Liability in those cases is even bigger, including higher damages: https://
law.cornell.edu/uscode/text/17
/504
…
@alexjc
-
Web Scrapers and Copyright Infringement Liability Scale
By
–
-
LLM Training on Pirated Content: Copyright Implications
By
–
Since these models are widely known to be trained on pirated books, and Tim is so eager to compare LLMs to people, keep in mind that people learning from pirated content is infringement (books or videos): https://
quora.com/Is-it-copyrigh
t-violation-that-downloading-pirated-movies-on-the-Internet-for-learning-purposes
… -
Fair Use Legal Implications for AI in Gaming Applications
By
–
It's not just @TimSweeneyEpic personal opinion, given legal risk, he has a fiduciary duty to ensure shareholders are onboard too. Even more impressive, Fair Use is assessed on a case-by-case basis, for each in-game application it may turn out differently: https://
guides.library.cornell.edu/copyright/fair
-use
… -
Copyright Chains and AI Training Data Legal Complexity
By
–
This risk is the reason why all movies and games have clear Chains Of Title involving the Copyrights of tens of thousands of people and many terabytes of data… (Almost as much data as goes into a model, though admittedly created by fewer people than a web-scale datasets.)
-
Epic Games Liability Safe Harbor Provisions Online Store Infringement
By
–
It's worth commending Tim's decision here, regardless whether he's wrong or not, because Epic will be liable for damages in all cases where infringement is found. Why? Safe Harbor provisions don't apply to online stores "receiving financial benefits":
-
LLaMa Inference Speedup 1.33x to 1.91x on GPUs
By
–
Speeding up LLaMa inference end-to-end by 1.33x on A6000 (for 13B model) and 1.91x on A100 (for 34b model). https://
arxiv.org/abs/2308.16369 -
AI Activism: Preventing Exploitation Without Consent
By
–
If you read Karla's thread, you'll see the ongoing activism aims to prevent exploitation. The demands are not about stopping AI. If you're advocating violating people's economic and moral rights, be aware that without consent there are no contracts and thus no markets.
-
Ant Colony Resilience: Individual Deaths, Collective Survival
By
–
Yes, if you zoom in there's death everywhere (individual ants) but zooming out the ant colony can be around for a very long time…
-
Is Stitching Multiple AI Parts Together Truly Transformative?
By
–
Have some respect for the work they did! They had to stitch multiple parts together. That's transformative… surely?
-
Data Augmentation Techniques for Texture-based Model Training
By
–
Since it's a domain-specific dataset, and in particular textures, you can do many augmentations to get more out of it… Training at 256×256, that's approximately 27,6k unique patches and with flip/rot augmentations about 221k! (Training random patches works even better.)