Universal Music Group and two Capitol labels accuse DistroKid of distributing infringing recordings, continuing to send some tracks to other services after takedown notices, and letting mass-produced AI uploads crowd out legitimate artists. The complaint names 1,000 recordings and seeks damages and court orders. These are allegations in a new lawsuit, not established facts, and DistroKid strongly denies them. The suit says it is not attacking clearly disclosed AI-assisted music; its target is alleged deception, spam, and infringement.
THE WIRE
Universal sues DistroKid
•
Australia debates opt-out AI training
•
No credible new local model displaced this week’s picks
Anthropic, OpenAI, and xAI leaders now publicly support some version of slowing dangerous capability gains. Associated Press finds that the proposed remedies—outside evaluators, shared standards, and international coordination—still collide with profit, national competition, and political resistance. There is no joint rule, independent enforcement body, or announced training pause.
Australian consultation documents describe options that could let AI companies train on online work unless creators digitally protect it, or use broad collective licences. Ministers say copyright protections will not be weakened and no decision has been made. For now this is a disputed proposal—not law—and the practical opt-out mechanism has not been explained.
Associated Press finds that Moonshot, DeepSeek, Z.ai, and Alibaba are producing increasingly competitive models despite U.S. chip restrictions. That does not mean China has overtaken the U.S., and benchmark wins do not settle the question. It does mean export controls have not prevented strong frontier and open-weight work.
The Agents API gives developers durable sessions, tools, context management, recovery, and hosted or self-hosted sandboxes without building the whole agent loop themselves. It entered public beta September 10. This is cloud infrastructure—not a new model or a consumer app—and users pay the chosen model, tool, and container rates.
CLOUD SERVICE · BOTH MACHINES CAN BE CLIENTS; NEITHER HOSTS ASTRA
Google, NVIDIA, Emerald AI, Anthropic, utilities, and power companies launched a 20-member alliance around shifting non-urgent computing away from peak demand. In return, they want faster grid connections and lower infrastructure costs. That could help power bills—but only if reductions are measurable, enforceable, and do not merely move the cost to everyone else.
The planned Queensland complex would answer Claude requests rather than train models. Its first stage is targeted for next year, but local approval is still pending. The state says new generation will accompany the load; the project can initially draw on coal power. This is a land and capacity deal—not a finished data center or a new Claude model.
The Delhi High Court refused ANI’s request for an immediate stay of a July ruling that treated LLM training on copyrighted news as private research and fair dealing. The appeals panel wants OpenAI’s response before deciding more and will hear the case again December 8. This is an interim procedural decision; the underlying copyright lawsuit is not over.
Portable Computer now runs its model, planner, scheduler, and local-file work on supported NVIDIA PCs, asking before escalating a task to cloud models. The catch is blunt: Windows requires a GeForce RTX or RTX PRO GPU with at least 24 GB of VRAM. The RTX 5080 laptop has 16 GB, and the Mac mini cannot run the Windows release.
RTX 5080/16: NO · M4/32: NO — USE THE SEPARATE MAC VERSION
The Verge’s early test found better handling of genre language and useful section-level editing, but familiar vocal artifacts remain. The supposedly less predictable v6-wild did not sound dramatically wilder in that test. This is Suno’s first model built with licensed material from record-industry partners; v6-mini is the free, lighter version.
Universal Music Group says artists will be able to opt in so fans can remix, mash up, and create new takes from licensed catalog music. The platform will be separate from ElevenMusic and ElevenLabs’ Music API. There is no launch date, price, participating-artist list, or hands-on product yet—so this is a licensing deal with a product promise attached.
Google says the update makes richer arrangements and more expressive vocals, with choices for vocals or instrumentals and short or longer tracks. It is rolling out in Gemini and the API; those quality claims still need listening tests outside Google’s demos.
GooglePublished Sep. 5
REVIEWED SEP. 16 · UPDATED WHEN SOMETHING BETTER LANDS
Local Models to Try
No leaderboard soup. These picks include creative work, everyday chores, and “can this actually help me?” tests—not only coding, science, and competitive math.
The new tiny one worth trying first. It is small enough to leave running while you do something else, with a surprisingly long memory for notes and documents.
Hugging Face just added direct Transformers support for Microsoft’s small model for long-form, multi-speaker spoken audio.
TRY IT FOR
Podcast table reads, two-person dialogue demos, audiobook mockups, and hearing whether a script flows before people record it.
RUNS HERE
MEMORY FITS BOTH · RTX IS SAFER · MAC SUPPORT NEEDS TESTING
REALITY CHECK
Microsoft calls the output high-fidelity, but also labels this research-only, limits it to English and Chinese, and adds an audible AI disclosure. Useful experiment; not a finished production voice tool.
A practical vision-and-language model for people who would rather show the computer the thing than spend ten minutes describing it.
TRY IT FOR
Explaining screenshots, reading labels and documents, comparing photos, visual brainstorming, and ordinary chat.
RUNS HERE
4-BIT: EASY ON RTX · COMFORTABLE ON M4/32
REALITY CHECK
The giant context number is not a dare. Start small, keep only the context you need, and see whether its answers stay useful.
The weekly rule: a model stays only while it gives normal people a useful reason to download it. A genuinely interesting release can bump the list before the next edition.
Will It Run on My Stuff?
AI laptopRyzen AI 9 HX 375 · RTX 5080 16 GB · 32 GB RAM
Mac miniM4 · 32 GB unified memory
These are our reference machines. Every local-AI story will say what ordinary home hardware it needs, what it costs, and where weaker systems hit the wall.
SOURCE DESK
Slant is allowed. Sloppiness isn’t.
We favor primary announcements, technical documentation, court filings, and reporters who show their work. Company claims are labeled as claims. Major stories get a second source. Rewritten press releases, anonymous hype accounts, and “this changes everything” sludge do not make the page.
Official labsAssociated PressThe VergeWiredArs TechnicaMIT Technology ReviewThe Guardian404 MediaTom’s HardwareAxios