A real search engine shows no sign of the "inference optimal" phrase; I agree with your assessment that you posted in your own thread FWIW. Train to convergence if you want inference optimal.
@alexjc
-
Optimal in Relative Terms: Chinchilla and Model Training
By
–
They use "optimal" in a relative sense. I read it in a reply to my post, and then subsequently looked it up in LLM search engine—which happily told me why "Chinchilla is inference optimal because […]". But that's only relatively to undertrained models.
-
Getty lawsuit challenges: Stable Diffusion watermark hallucination analysis
By
–
Getty is going to have a tough case… images from their lawsuit don't really have any similarities. The only problem that #StableDiffusion hallucinated a watermark from the noise, or maybe because it was prompted? https://
stableattribution.com/?image=1e3bdf0
2-8e7d-489c-828d-137f789c99f6
… -
AI Models Embed Copyrighted Character Expression as Infringement
By
–
Some images are kinda similar to copyrighted works, but the underlying expression of the characters is always there. Thus it must also be in the model, and it's infringement beyond individual works.
-
Generative Models as Expression Copying Machines in Copyright Context
By
–
FWIW, the core of Mathew's argument can be dismantled in this case by the fact that generative models are expression copying machines. For instance, in the context of character copyrights:
-
Generative Models: How They Copy Expressions and Compete with Authors
By
–
The PDF you linked is fascinating! Only written three years ago but feels very dated already… With generative models you could argue the purpose is Expressive. The generative models copy the expressions and compete with the original author.
-
Learning from Data and Generating: Self-Consistent Engineering Principles
By
–
You're lost in limbo with these arguments. It's just learning from data and generating from what's learned. It's backed by self-consistent engineering principles.
-
Copyright Resolution and Artist Control in AI-Generated Content
By
–
(Assuming the copyright issue is fixed…) If people buy the results are OK with that, or if the artists know how to control or correct it, I don't see any problem.
-
Illusion versus Delusion: Technology’s Useful Appeal
By
–
Whether it's illusion or delusion is a matter of perspective. ***lusion is useful in many ways, as shown by the millions of people who have flocked to the technology. Don't let your history with copyright cloud your judgment.
-
Consistency and Measurability in Research Override Concerns
By
–
It doesn't matter if it's delusion or hallucination. It works as long as it's shown to be consistent with the theory, and in this case it's measurable. Non-profit research is allowed worldwide even if the data is copyrighted.