The Signal — August 30, 2026

Two music publishers want a full accounting of everything that went into Claude. On the same day, two Chinese labs put frontier open weights on the internet with opposite ideas about what a license is for, and a group of researchers spent the week on the least glamorous problem in the stack, which is that the processor in your machine mostly sits idle while the graphics card waits on the bus.

Sony and Warner Chappell sued Anthropic, and named Dario Amodei personally

Sony Music Publishing and Warner Chappell Music filed a 48-page copyright complaint against Anthropic on Friday in the Northern District of California, case 5:26-cv-09217. The unusual part is the caption: alongside the company, the publishers named CEO Dario Amodei and co-founder Benjamin Mann as individual defendants, with a contributory infringement count aimed at the two of them personally.

The factual spine of the complaint is borrowed. It leans on findings already established in Bartz v. Anthropic, the authors' case in the same district, which produced the allegation that Mann used BitTorrent in June 2021 to pull at least five million pirated books from Library Genesis and that employees torrented at least two million more from Pirate Library Mirror a year later. The filing quotes unsealed internal material, including Mann calling LibGen "sketchy AF" and Anthropic's own Archive Team describing it as a blatant violation of copyright. Beyond the torrenting counts, the publishers allege scraping of licensed lyric sites in violation of their terms and a destructive-scanning operation that digitized second-hand physical books before disposing of them.

The relief sought includes statutory damages up to $150,000 per work willfully infringed, up to $25,000 for each alleged removal of copyright management information, destruction of infringing copies, and an accounting of Claude's training data. Those are ceilings written into the statute, not a forecast of what a jury would award, and nothing here has been adjudicated.

What makes this filing different from the four that preceded it is arithmetic. With Sony and Warner Chappell in, the publishing arms of all three major music companies are now litigating against Anthropic, alongside BMG and Round Hill, whose suit is thirteen days old. The complaint is explicit that the September 2025 book-authors settlement did not work as a deterrent, arguing that $1.5 billion reads as a cost of doing business to a company the plaintiffs describe, citing a Forbes report on a projected October IPO, as carrying a $2 trillion valuation. That framing is the publishers' own and the valuation is not independently confirmed, but it tells you what the next round of these cases is arguing about. Not whether the training data was pirated, which one judge in this district has already characterized as straightforward piracy at massive scale, but whether any settlement number is large enough to change behavior. Anthropic had not issued a public response in any of the coverage.

Sources: Music Business Worldwide · The Verge · Engadget · PacerMonitor docket

Tencent open-sourced 770 billion parameters and says the model helped build itself

Tencent released Hy4 preview on Friday under Apache 2.0, and the weights went up the same day as the press release. The model card confirms the shape: 770 billion total parameters with 49 billion active per token, 78 layers where the first is a dense feed-forward and the remaining 77 are mixture-of-experts with 256 routed experts and one shared, top-8 routing, a context window past a million tokens. The checkpoint runs 1.56 terabytes on Hugging Face, with an FP8 variant alongside it. Its predecessor Hy3, from July, was 295 billion total and 21 billion active in a 598-gigabyte package. The jump is roughly two and a half times the parameters and four times the context.

Apache 2.0 is about as permissive as open weights get, with no revenue thresholds, no security review, no acceptable-use rider. Z.ai shipped its own flagship weights the same day under a custom license replacing MIT, requiring any hosting company above $10 billion in revenue to pass a security review before commercial use. Two Chinese labs, both releasing at the frontier of open weights, on the same afternoon, having made opposite bets about what those weights are for.

Tencent's evidence for the model being good is thinner than its evidence for the model being large. The headline number is not a public benchmark at all but an internal blind evaluation, 163 Tencent experts rating outputs across 203 engineering tasks, where Hy4 preview averaged 2.99 out of 4.00 against GLM-5.3 at 2.92 and Kimi K3 at 2.94. Vendor raters, vendor tasks, and a spread narrow enough that a different rubric could reorder it.

Tencent also says Hy4 preview participated in optimizing its own training methods, data strategies, and evaluation frameworks, and that it autonomously analyzed its own inference stack and tuned operator fusion and communication for a 31.8 percent throughput gain. The company calls this an early-stage recursive self-improvement loop. Read plainly, it describes a model doing automated engineering work inside a pipeline its engineers built and supervised, which is a real and useful thing that a number of labs are now doing without attaching a phrase to it. The phrase will travel further than the result.

Sources: Tencent · Hugging Face model card · Simon Willison · GitHub

An inference engine that stops the graphics card from waiting around

Sparse models are supposed to be cheap because they only compute a fraction of their parameters per token. On consumer hardware that promise mostly does not survive contact with the bus. The full expert pool has to be reachable somewhere, which means system memory, and a PCIe link moving 16 to 64 gigabytes a second is slow enough that when a token routes to an expert that is not already resident on the card, the graphics card simply stops and waits for the weights to arrive.

FreeToken, released this week by a group from UC Berkeley and MIT that includes Matei Zaharia, Ion Stoica, Song Han, and Kurt Keutzer, takes the position that stopping is the mistake. When an expert misses the cache, the engine computes it on the CPU while streaming the weights to the GPU in parallel, and it decides how to split that work using a closed-form policy the authors call q-star, calibrated against the machine's actual measured bandwidth rather than a fixed offloading rule. Around that sit a fast weight format, full-layer double buffering so streaming overlaps with compute, and a memory manager that reallocates video memory between the KV cache and resident experts at runtime without reloading the model.

Coding agents rewrite their own context constantly, deleting thinking blocks and editing tool arguments, and a conventional engine treats any change to the prefix as a reason to discard the cache and re-prefill thousands of tokens. FreeToken checkpoints attention state at semantic boundaries instead, so an edit invalidates what came after it and nothing before. On a consumer card, where re-prefill costs far more than it does on datacenter silicon, that gap decides whether an agent loop is usable at all.

The reported numbers are 39 tokens per second for a 35-billion-parameter model on an 8GB laptop card, a 284-billion-parameter model at interactive speed on an RTX 5090, and a 753-billion-parameter model on a single workstation GPU, with decode three to four times faster than llama.cpp and Ollama on equivalent models. All of it is author-reported in a preprint with no peer review and no outside replication, and the discussion threads are already arguing about whether the baselines were tuned as carefully as the system under test. The code is public, which means the argument is at least settleable.

Sources: InfoQ · arXiv preprint · GitHub

On the Editor's Desk

Z.ai's new license for the GLM-5.3 weights, the one that replaces MIT with a revenue-threshold security review, is a real change and we flagged it yesterday. Running it again a day later would be the same paragraph twice, so it appears above as context for the Tencent release, where the contrast is the point. If that security-review clause ever produces an actual refusal, that is worth its own story.

An arXiv preprint on adversarial capture in agent populations argues that safety evaluation still treats the individual model as the unit while agents increasingly read and write each other's decisions. Good question, but it is one preprint of simulations, and simulated agent populations are not deployed ones. We are watching it rather than reporting it. The Apex Logistics chip-smuggling probe has decent sourcing, with the company and its parent both confirming they are cooperating, but the shipments in question are from 2024, the Bureau of Industry and Security declined to comment, and there is no finding of wrongdoing. That is an investigation, not a result. And MIT's committee report on AI in teaching and research is a real document with real recommendations, but it was published on August 13 about one university, and nothing happened this week to make it news.