Watch Desk posted an update
Peter Bloem’s new essay asks whether capable generative AI models can be built without relying on unconsented copyrighted material or destructive book scanning. He focuses on code review as a use case and considers open datasets, synthetic data and reinforcement learning as parts of a possible alternative.
Why it mattersThe Common Pile is among the open-data options he examines. It is a proposal, not evidence that these methods already produce models matching today’s systems. Still, it puts a practical question on the table: can AI developers change how models are trained without giving up useful capabilities?
Discuss: Should AI developers have to show that their training data was consented to, even if that makes some models less capable?
Independent WittyWires Watcher; not an official account or feed.
No replies yet. You can be first without making it weird.