Six stories from the last day: a landmark music-industry lawsuit, a very large open-weight model, and the usual reminder that the AI boom runs on debt.
- Sony Music and Warner Chappell sue Anthropic over "brazen" copyright theft
- Tencent ships Hy4, a 770B-parameter open weight model
- Nvidia's AI advantage is moving beyond the GPU
- Neocloud Lambda borrows $1B to buy more chips
- An Anthropic researcher shows a peek at self-improving AI
- Musicians are turning detective to catch AI grifters
Sony Music and Warner Chappell sue Anthropic over "brazen" copyright theft
Sony Music and Warner Chappell filed suit against Anthropic in the US District Court for the Northern District of California.
The complaint covers "tens of thousands" of copyrighted works and asks for up to $150,000 per work, plus up to $25,000 for each instance where copyright metadata was stripped — a damages ceiling that could run into the billions if the labels win. It's the latest and broadest suit yet in the music industry's fight over training data, and it specifically alleges piracy, not just unlicensed use, which raises the legal stakes for Anthropic considerably.
Read the full story → The Verge
Tencent ships Hy4, a 770B-parameter open weight model
Tencent released Hy4 Preview, a text-only open-weight LLM: 770B total parameters, 49B active, 1M token context, 1.56TB of weights on Hugging Face.
That's a sharp jump from Hy3 in July (295B total, 21B active, 256K context, 598GB) — Tencent is scaling both size and context window fast, and doing it in the open. For anyone tracking the open-weight frontier outside the US labs, Hy4 is now one of the largest models freely downloadable, and its chat template hints at built-in reasoning-mode toggles worth digging into.
Read the full story → Simon Willison
Nvidia's AI advantage is moving beyond the GPU
The newest generation of Nvidia-powered data center systems is gaining efficiency from smarter traffic control between chips, not just faster processors.
This matters because the GPU-count arms race has a ceiling — power, cooling, and interconnect bandwidth are now the binding constraints for large training runs. If Nvidia's edge increasingly comes from system-level orchestration rather than raw silicon, that's a moat that's harder for chip challengers to copy with a faster GPU alone.
Read the full story → TechCrunch
Neocloud Lambda borrows $1B to buy more chips
Lambda raised $1 billion in private debt to buy Nvidia AI chips, which it will lease to Microsoft.
It's the latest in a string of debt-financed GPU buildouts, and it's a clean illustration of how much of the AI infrastructure boom is running on leverage rather than equity. When a hyperscaler like Microsoft would rather rent compute through a leveraged intermediary than buy outright, it's a signal about how capital-intensive — and how uncertain the depreciation curve — this hardware still is.
Read the full story → TechCrunch
An Anthropic researcher shows a peek at self-improving AI
Given ten benchmarks targeting specific misaligned behaviors, an automated system improved performance on every single one — without degrading the model's overall performance elsewhere.
Self-improvement loops that don't trade off general capability for safety fixes are the harder engineering problem, and this is early evidence Anthropic can do it without the usual regressions. Worth watching for how it changes the pace of alignment work versus capability work going forward.
Read the full story → TechCrunch
Musicians are turning detective to catch AI grifters
As AI-generated music tools get more sophisticated, musicians and fans are increasingly doing the forensic work themselves — tracing algorithmically-derived melodies and vocals back to human artists whose work trained the tools.
Some creators disclose AI use upfront; others have denied it until public scrutiny forced an admission. The pattern says something Nerdic cares about directly: disclosure works better as policy than as a fallback after getting caught.
Read the full story → The Verge
Tools worth your time
- Hy4 (Tencent) — A 770B-parameter open-weight text model with a 1M token context window, free to download on Hugging Face. For teams that want frontier-scale reasoning without an API bill, at the cost of needing serious hardware to run it. huggingface.co
- OCaml security disclosure practices (community tooling) — Not a product, but a live case study: patches shared for community review are seeing exploit attempts within minutes. Worth a look for anyone maintaining open-source infrastructure who still treats a "responsible disclosure window" as days, not minutes. ocaml.org
This piece was drafted autonomously and reviewed by a human editor before publishing. Every machine-written article on Nerdic carries this mark.
