"Inference Optimal" isn't a thing… yet? The point still stands and is worth pursuing! A nicer way to phrase the same thing:
@alexjc
-
Clarification on Article 4 scope and directive text interpretation
By
–
Also, could you clarify this from p. 19? I don't see how Article 4 applies to "any purpose", when the directive text says only "reproductions and extractions".
-
Clarification needed on datasets versus dataset indices and copyright infringement
By
–
Interesting read so far! The bits about datasets vs. datasets indices could use clarifying. LAION's dataset *index* is non infringing, but if you include the scraped works of course it is (as per Berne).
-
Training to Convergence and Inference Time Optimality Trade-offs
By
–
The original question was about optimality, so if you train until convergence you have better performance per inference time — thus it's more optimal?
-
Overtraining of 13B Model vs Suboptimal 65B Production Deployment
By
–
So to answer the original question, they overtrained the 13B model but not the 65B model — likely because they decided on the budget beforehand. Thus, it's suboptimal to run the 65B in production.
-
AI Models as Pattern Extraction, Not Derivative Works
By
–
They're arguing that it's not a decomposition of the original expressions, but patterns extracted from the non-copyrightable data that's there. Thus, the model is not derivative and outputs too. The copying is a triviality they hope is waived by Fair Use or equivalent exception.
-
Legal argument for machine learning as expressive use
By
–
I wonder if someone with sufficient knowledge of the technology and of Copyright could come up with a legal argument to show that ML is actually an expressive use?
-
Model Performance Limited to Benchmarks, Real Version Unavailable
By
–
The disclaimer is that its on benchmarks only and nobody has access to the real model yet…
-
Machine Learning Training and Copyright Bypass Legal Arguments
By
–
By saying "Machine Learning and training is a non-expressive use" that just uses the data, you basically try to bypass Copyright internationally. Of course, it's bullshit that high-priced lawyers make up to land Big Tech contracts — or get funding for their University.
-
Non-expressive Use: Legal Workaround for Copyright Avoidance
By
–
The legal buzzword to use is "non-expressive use" here! It's a workaround for Copyright. Here's how it works: Since Copyright only applies to expressions and not ideas/data, you use semantic & "ontological games with legal implications" to avoid Copyright.