Sony and Warner Chappell sue Anthropic, naming Amodei and Mann personally

Sony Music Publishing and Warner Chappell sued Anthropic, Dario Amodei and Benjamin Mann over the pirated acquisition of training data, the third major-publisher action against the company. Tencent released and open-sourced Hy4 Preview, a 770B-parameter model with 49B active and a context window past one million tokens, 2.6 times the total size of the Hy3 it replaces after seven weeks. A four-step distillation of MiniMax H3 published by a university lab cut 15-second video generation from 678.7 seconds to 47.2 on a single B200.

Regulatory & Policy #

Sony Music Publishing and Warner Chappell sue Anthropic over torrented training data #

TechCrunch / Music Business Worldwide

The complaint, filed in the Northern District of California, alleges a “brazen campaign of illegally torrenting, scraping and downloading copyrighted works” — specifically that Benjamin Mann used BitTorrent in June 2021 to pull at least five million pirated books from Library Genesis, that employees torrented at least two million more from Pirate Library Mirror in July 2022, and that Anthropic scraped lyrics from MusixMatch and LyricFind, both of which had paid to license that content. It brings four counts: direct infringement by torrenting against all three defendants, contributory infringement against Amodei and Mann individually, and direct infringement plus removal of copyright management information against Anthropic alone, seeking up to $150,000 per infringed work and $25,000 per instance of stripped CMI. Anthropic says it disagrees with the claims and will defend itself in court. The structure is the thing to read: the suit attacks how the corpus was obtained rather than what Claude produces, which tracks the July judgment that ordered Anthropic to pay $1.5 billion on the finding that training on copyrighted work could be lawful while acquiring it through piracy was not. With Universal, Concord and ABKCO already litigating from 2023 and again in January, all three major music publishers are now suing the same defendant on the same theory.

Model Releases #

Tencent open-sources Hy4 Preview at 770B parameters with 49B active #

Tencent / Simon Willison / Hacker News

Hy4 Preview is a text-only mixture-of-experts model — 770B total, 49B active, a context window over one million tokens, and 1.56TB of weights on Hugging Face. The jump from Hy3, released in July at 295B total, 21B active, a 256,000-token context and 598GB, is 2.6x on total parameters in about seven weeks. The chat template exposes exactly two reasoning settings, high as the default and no_think to disable it, rather than a graded effort dial. API pricing is $0.834 per million input tokens, $2.501 per million output and $0.042 per million on cache hits, with access through Tencent Cloud TokenHub and OpenRouter and free use in WorkBuddy and CodeBuddy for two weeks. The quality claim needs more scepticism than the specifications do: it rests on Tencent’s own internal evaluation of 203 engineering tasks judged by 163 experts, scoring 2.99 out of 4.00 against GLM-5.3’s 2.92 and Kimi K3’s 2.94 — margins of 0.07 and 0.05 on an unpublished rubric run by the vendor. The announcement also does not name the licence the weights ship under, which matters more than the score for anyone planning to deploy them.

Open Source #

A four-step distillation of MiniMax H3 generates 15 seconds of video in 47.2 seconds #

Hao AI Lab (UCSD) / FastVideo

FastH3 Preview v1 distils MiniMax H3’s base transformer for text-to-video-and-audio, replacing 49 transformer calls with four diffusion steps and running video attention at 90% sparsity on tile-64 blocks while text and audio stay dense; the H3-Base text encoder, video and audio VAEs, tokenizers and schedulers are reused unchanged. At 1344x768 on B200 hardware, five-second clips take 16.2s on one GPU and 6.1s on four against a 132.5s base, and 15-second clips take 47.2s on one and 15.5s on four against 678.7s — 8.16x and 14.38x. It was built with Nuva Lab and NVIDIA’s FastGen team on over a thousand B200 training hours, and the weights are on Hugging Face. Two limits are load-bearing rather than cosmetic: the reported speeds require FastVideo’s VSA-H3 backend and kernel, and dense attention is explicitly not a drop-in substitute for the sparse checkpoint, so the number does not travel to an arbitrary serving stack. Coverage is text-to-video-and-audio only — the first-frame and reference-image conditioning paths were not distilled — and the lab notes difficult motion, fine detail and some audio fall below the base model.

Infrastructure #

Purdue validates a distributed-GPU simulator against silicon at 99% correlation on H100 #

Semiconductor Engineering

The paper presents a cycle-level simulation framework spanning Ampere, Hopper and Blackwell and reports validation against physical silicon at a 99% Pearson correlation coefficient on H100. That figure is the useful part, because architecture exploration for multi-GPU AI systems runs almost entirely on simulators whose fidelity to shipping hardware is asserted rather than measured, and a stated correlation against real parts is what makes a projected speedup worth arguing about. The publisher’s page would not serve to this run, so only the abstract excerpt was available: the paper’s actual conclusions about asynchronous distributed GPU architecture are not covered here, and the validation number should not be read as a claim about them.

Threads to Watch #

The capability frontier and the serving-cost frontier are moving on separate tracks, and only one of them belongs to the labs. Hy4 is 2.6x the total parameters of the model it replaces after seven weeks, which is the first track. FastH3 is the second: a university lab took someone else’s published checkpoint and cut a 15-second generation by 14x by changing the sampling schedule and the attention pattern, training nothing new of consequence. The second kind of work is cheap, fast and reproducible, but it exists strictly downstream of the first — a distillation is only publishable if the base weights were.

The legal exposure has settled on acquisition rather than output. The publishers’ complaint is built on torrenting and stripped copyright management information, not on what Claude generates, which follows the July finding that training on copyrighted work could be lawful while obtaining it through piracy was not. Naming Amodei and Mann individually for contributory infringement moves that from corporate liability toward personal, and the conduct alleged is from 2021 and 2022. Any lab that assembled a pretraining corpus in that window faces the same theory regardless of what its current models do.

Every number in today’s digest is produced by a party with a stake in it, or is conditional on a harness. Tencent’s 2.99 out of 4.00 comes from Tencent’s own experts on Tencent’s own rubric with a 0.05-point margin over the nearest competitor. FastH3’s 14.38x is real and carefully reported, but it holds only on a specific backend and kernel that the authors say cannot be swapped out. The Purdue correlation is the exception that shows the shape of the problem — it is a claim about how well the measuring instrument matches reality, which is the thing the other two leave unstated.