Meta did some experiments whether pirated materials helped, and the bottom line is that it helped a bit on a few of the benchmarks. I don't see numbers worth justifying commercial scale Copyright infringement though…
@alexjc
-
GPU Optimization Performance Comparison FP8 Quantization Testing
By
–
Understood, testing now! I got your previous version working too and that seemed to be much faster on 2×3090 (fp8 part disabled) than the code above, not sure why yet…
-
Gradient Division in Multi-Step Training Reproducibility
By
–
To make this reproducible on rigs where there are multiple inner steps per training step, isn't the division of the gradient by that number of steps missing?
-
Professional AI domain work requires knowledge and humility
By
–
Working on this domain professionally. It's not my job to do your homework for you, burden of research is on you if you're going to doompost and get caught, clearly acting without knowledge and even less humility.
-
UK AI Policy Adopts EU Regulatory Framework Approach
By
–
The text of the proposal itself suggests using an approach similar to the EU. The EU's approach is one of the most reasonable ones worldwide (assuming it's interpreted correctly), so if the U.K. is going to copy anyone it's best to copy Europe! https://
gov.uk/government/pub
lications/ai-opportunities-action-plan/ai-opportunities-action-plan
… -
Meta Faces Scrutiny Over Piracy in AI Training Practice
By
–
Lawyer comes up with a long list of excuses, buries the lede, then finally admits firing Meta. Maybe the company could be on the right side of a theoretical debate, but in practice them relying on piracy (known all along) makes it hard for Lemley to maintain his reputation.
-
Next-Token Prediction Self-Supervised Learning and Uncertainty
By
–
This particular view is a decade out of date. You know that next-token prediction is mostly self-supervised, and as such the problem-solution pairs are "automatically" generated? Doing the same to train "I don't know next token" is a minor reframe, much research exists.
-
Intentional Delays in Large-Scale AI Implementation: Strategic Questions
By
–
Nobody said conspiracy, stop watching dubious parts of YouTube. It is, however, intentional to do it so late and in such a large-scope delay-prone fashion. Question is why.
-
APA Notice-and-Comment Procedure vs Non-Binding USCO Guidance
By
–
You're implying that this falls under the APA's notice-and-comment procedure for rule *making*? But you also suggested above USCO is non-binding and only provides guidance / advice. Which is it? I can see why you bailed out, their delay makes no sense and can't be defended.
-
Experts Bypass Public Consultation: Strategic Unwillingness Analysis
By
–
They had no obligation to do a massive call for comments, including internationals. They are experts, could have written a 1-page opinion ages ago like Neil suggested. It's not that they were unable, the fact they didn't is because they were unwilling to do so. Why?