Reply to thread

Message: [QUOTE="Bombastus, post: 92251, member: 2178"] Suno launched v6 on September 9, 2026, and said the new model family was trained from scratch on data different from prior versions. Suno also said its new training process used preference signals gathered from users, which is where Universal and Sony see a connection to the older systems. Suno's position is narrower than saying nothing from its past influenced v6. Chief Product Officer Jack Brody said the model was built from the ground up on different training data, including licensed material, and that the dataset did not include Universal or Sony recordings. He also described user learnings, tuning, technical breakthroughs, and model architecture as important parts of the improvement. Universal and Sony attack that distinction in [B][URL='https://goldmidi.com/community/threads/universal-and-sony-have-sued-suno-again-over-60-202-tracks.77641/']the dispute over Suno v6's inherited training[/URL][/B]. Their complaint says a new dataset does not create an independent model if older Suno systems still supplied outputs, preference information, or learned behavior during development. [HEADING=2]A clean start can still carry older model behavior[/HEADING] Training from scratch says something important about the new model's construction, but it does not describe every source of information used during training. A student model can begin with newly initialized parameters and still learn from another model's responses rather than inheriting its weights. The original [B][URL='https://arxiv.org/abs/1503.02531']knowledge-distillation framework[/URL][/B] describes transferring useful behavior from a larger teacher into a smaller student by training the student against information produced by the teacher. The student does not need a literal copy of the teacher's parameters for the teacher to influence what it learns. Universal and Sony allege Suno used this kind of model-transfer process when developing v6. Their complaint names earlier systems including v4, v4.5, v4.5+, v5, v4.5-all, and v5.5 as predecessors whose learned behavior allegedly fed forward. A technical inheritance claim is therefore possible even when old training files never enter the new corpus. The harder issue is legal. Showing that one model helped teach another does not establish that protected expression from a particular recording survives inside the student in a form copyright law treats as an infringing copy. [HEADING=2]Preference data is a different kind of connection[/HEADING] Suno generates two songs for a normal creation and can learn from signals about which result a user prefers. Suno has said this preference information, rather than users' uploaded audio, was part of the user data involved in developing v6. Those signals are not the same as feeding an old recording into a training set. A preference can show that users repeatedly choose one type of output over another, giving Suno information about arrangement, prompt adherence, vocals, pacing, or other qualities people respond to. The labels focus on where those choices came from. If users were judging pairs generated by earlier Suno models, Universal and Sony argue the preferences were built on outputs produced by systems trained on their recordings. In their theory, v6 benefited from information created through the older models even if the original masters were excluded from the new dataset. Suno can contest how far that chain reaches. A user's choice between two generated songs is a judgment about the outputs presented, and a ranking signal does not automatically identify which protected recording, if any, contributed something legally recognizable to the preferred result. Preference tuning and distillation should not be collapsed into one process. One captures human reactions to model outputs. The other transfers behavior from a teacher model into a student through training targets or related techniques, and the factual record still has to show what Suno actually used. [HEADING=2]Copyright has to follow more than technical influence[/HEADING] The lawsuit pushes model lineage into territory copyright cases have barely mapped. Technical influence can travel through outputs, rankings, and teacher signals without carrying the original audio file along for the ride, but copyright liability still requires a legally relevant act involving protected expression. Universal and Sony want the court to treat v6 as part of a continuous exploitation chain. Their argument becomes stronger if discovery shows the earlier models produced training material or teacher signals that preserved recognizable elements derived from the labels' recordings rather than broad musical capabilities. Suno can press the opposite distinction. General improvements in composition, arrangement, prompt following, or musical quality are not automatically copies of individual sound recordings merely because earlier systems helped produce the training signals used to reach them. A direct reuse finding would change the shape of the fight. If evidence eventually shows Universal or Sony recordings themselves entered the v6 corpus, the labels would not need to rely as heavily on preference data or model inheritance to connect old training conduct with the new system. Without that direct link, the case turns on something more specific than whether v6 learned from Suno's past. The difficult question is whether anything protected by copyright traveled through those intermediate steps in a form the law recognizes as reproduction or infringement, rather than technical know-how, user taste, or generalized model behavior. [/QUOTE]

Name