Banned From The Charts, Funded By The Labels | Daily AI News Brief (Aug 25)
Two things happened in music today that look like a contradiction and aren't. Australia's official charts banned wholly AI-generated tracks. And the three major labels — the same rights holders you'd expect to be cheering that ban — wrote a $76M check to an AI music company. Both moves were deliberate, and together they tell you exactly where this is heading. Then, in a completely different direction, two companies spent the day making capable AI something you run on hardware you already own.
Australia Bans Fully AI-Generated Music From The Charts
ARIA — the body behind Australia's official music charts — announced that wholly AI-generated tracks can't chart anymore, and they're out of ARIA Award eligibility too. To qualify, a song has to be "substantially human made."
What I respect here is that they actually defined it. Humans wrote the song. A human sang the lead vocal. Humans played the primary instruments. AI in a supporting role is still fine — mastering, drum machines, autotune — as long as the tools themselves are legal, licensed services.
What pushed them? Last month, an AI-heavy cover of Madonna's Like a Prayer became the most-played song on Australian radio and hit number four on two ARIA charts. Nobody is pretending this rule isn't a response to that.
Now — I wouldn't call this an anti-AI rule. What I would say is it's an anti-anonymous-AI rule. Look at what's still allowed: assistance is fine. Generation with no human in it is what got banned. And Australia isn't doing this alone — more than twenty official chart programs worldwide are moving on the same principles.
One note on precision: outlets disagree on exactly when the rule takes effect, so I'm not going to state a date I can't stand behind. The announcement is what's confirmed.
Creator takeaway: If you make music with AI anywhere in your chain, start keeping receipts today. Session files, stems, writing credits, which tools you used and where. The question is shifting from "did you use AI" to "can you prove which parts are you" — and the artists with documentation are going to sail through while everybody else argues with a chart administrator.
Sources: ABC News (AU) · Variety · Fortune · AP via US News
The Same Labels Drawing That Line Just Wrote A $76M Check
Stability AI announced a $76M Series B, and the new investor list is the whole story: Universal Music Group. Sony Music. Warner Music. Electronic Arts. Plus AMD's venture arm. That takes Stability's total funding to about $232M.
This didn't come out of nowhere. Universal signed a deal with Stability last October to co-develop AI music tools. Warner followed a month later. And in May, Stability shipped Stable Audio 3.0 — music models trained, they say, entirely on licensed data. That's the company's claim, to be clear — but the labels putting their own money in tells you they've seen the paperwork.
So here's the verdict on the day's big question. The rights holders aren't fighting AI. They're building the sanctioned lane. Fully AI tracks get banned from the charts — and licensed, artist-compensated AI tools get funded by the labels themselves. Ban the outside, own the inside. That's not hypocrisy; that's the whole playbook, and it's going to spread past music.
Creator takeaway: Which lane your tools sit in is about to matter commercially. ARIA's rules already require legal AI services. So the next time you're picking a generative audio or image tool for client work, ask the training-data question before you ask the quality question. Licensed or scraped isn't a philosophy debate anymore — it's turning into an eligibility requirement.
Sources: Stability AI (primary) · Variety · TechCrunch · Music Business Worldwide
IBM's Granite 4.2 — Free Reasoning Models You Can Actually Run
IBM released Granite 4.2 — reasoning models in three sizes, 3B, 8B and 30B parameters, all open weights under Apache 2.0: commercial use, fine-tuning, no gatekeeper.
The headline feature is a thinking switch — full chain-of-thought when the problem is hard, off when it isn't, plus a low-effort mode for the easy stuff. You also get native tool calling, a 512K token context window on the 30B, and the bigger models were trained with reinforcement learning inside real coding, terminal and web-search environments.
That last detail is the tell: these are agent engines, not chatbots. Weights are on Hugging Face and Ollama right now.
Creator takeaway: The 8B is the sweet spot for hardware normal people own. Pick one high-volume task you currently pay per-token for, run ten real samples through it with thinking on, and compare. Zero dollars a token is hard to argue with — but only if the output holds up. Test it. Don't trust it.
Sources: IBM Research (primary) · Hugging Face model card · Unite.AI
Perplexity Puts An AI Agent On Your Own Hardware
Perplexity launched Portable Computer with Nvidia — a local-first version of its agentic Computer platform. And it's actually local: the whole agent stack runs on your machine — Nvidia's DGX Spark desktop, or a Linux box with enough RTX video memory. Windows is "coming soon" — Linux first, so don't plan around Windows yet. The model is Qwen 3.8 27B, or Perplexity's own tuned version.
Two things stand out. Local work costs zero tokens — no metering. And it's private by default: tasks start on your device, and it asks permission before sending any step to the cloud. Live now for Pro and Max subscribers.
Creator takeaway: Care about this even if you never buy the hardware. If your work touches things you can't ship to somebody's API — client contracts, financials, unreleased creative — agents just became something you can run without having that conversation. The honest catch: "runs locally" still means "runs on hardware most people don't have yet." So don't go buy a DGX. Just know, before you sign an annual cloud-agent contract, that the local option exists and it's maturing fast.
Sources: Perplexity (primary) · VentureBeat · MarkTechPost
Quick Hits
OpenAI published the first Jalapeño benchmarks. The inference chip it built with Broadcom is a 700-watt part they say delivers up to nearly 2× the throughput per kilowatt of Nvidia's flagship racks, with up to 3.6× lower latency. Keep this in perspective: those are OpenAI's numbers about OpenAI's chip, not independent testing. Deployment starts in their own data centers later this year, and efficiency gains like that eventually show up in what you pay for hosted AI.
ChatGPT Business is getting Premium seats — 5× the usage, no more five-hour cap. One correction to what's circulating: it's $125/month, and $100 only if you bill annually — and it's still a waitlist, not generally available. Don't budget around it yet.
Nvidia announced the Jetson Orin Nano 2, an entry-level edge computer for robots, drones and vision projects. 78 trillion operations a second, double its predecessor's inference, same form factor. It ships first half of 2027 — so that's a heads-up, not a purchase order.
And two smaller moves that show where the money's going. Gamma — the AI presentation app — acquired design startup Lica to spin up a design research lab. And a startup called Keenable came out of stealth with $26M to build a 100-billion-document web index just for AI agents. The picks and shovels around agents are getting serious.
Sources: OpenAI Jalapeño (primary) · Tom's Hardware · Premium seats (OpenAI) · Jetson Orin Nano 2 (Nvidia) · Keenable
Actionable Takeaways for Creators & Solos
- If AI touches your music anywhere, start a receipts folder today — session files, stems, writing credits, tool log. Proof of which parts are yours is becoming an eligibility requirement, not a nice-to-have.
- Ask the training-data question before the quality question when you pick a generative tool for client work. Licensed vs scraped just became commercial, not philosophical.
- Pull Granite 4.2 8B and run ten real samples from a task you currently pay per-token for. Zero-cost tokens only count if the output holds up.
- Don't buy a DGX — but don't sign a multi-year cloud-agent contract either. The local option exists now and it's maturing fast.
- Treat vendor benchmarks as marketing until someone neutral runs them. Today's Jalapeño numbers are the case study.
- Don't budget for ChatGPT Business Premium yet — $125/mo, and it's still waitlist-only.