Voice.ai: GPL Violations with a Side of DRM

Re:Voice.ai: GPL Violations with a Side of DRM by jimlongton on Monday February 06, 2023 @08:47PM (#34687457)

> Their goal/claim is that the language model will not retain copyrightable sequences.

In case people are not aware, we know for a fact this is demonstrably false. It may not even be possible to achieve this with existing models. People have replicated entire sections of GPL code (with comments) using CoPilot. This is true of other models, like image generation models which can return clearly recognisable images which were present in their training sets.

[ Parent | on HN ]

Re:Voice.ai: GPL Violations with a Side of DRM by Dylan16807 on Monday February 06, 2023 @09:24PM (#34687788)

The interesting question is how hard that is to fix in future versions.

[ Parent | on HN ]

Re:Voice.ai: GPL Violations with a Side of DRM by jeroenhd on Tuesday February 07, 2023 @04:24AM (#34690554)

As far as I know, an AI spitting out training data has clearly been overfitted on its input.

However, I think this may be intentional, as an AI that doesn't spit out the occasional bit of training data may not be delivering output of a high enough quality.

[ Parent | on HN ]