← Back to archive
Issue #33 · Monday, September 7, 2026 · 7 stories

🎬 Danny Boyle Used 30 Seconds of AI. Roku Wants Infinite Hours.

Plus a fake rep who brokered a Mary J. Blige ad, and a 6B editor you can run on your own GPU.

Three unrelated stories this week make the same point: the paperwork is now the hard part of AI video. The free AI channel on your Roku is run by a Los Angeles company called Fairground, and its creator agreement lets you keep copyright while taking a worldwide sublicensable licence, paying 50% of net receipts after five deduction categories, and taking up to 210 days to get there. Minnesota's nudification law survived xAI's challenge with penalties reaching $500,000 per access, and Suno pulled a Mary J. Blige campaign after finding the approval chain ran through someone impersonating her representative, because a real file and a signed-looking deal are not clearance. Generation itself keeps getting cheaper, with a local 6B editor, a step-by-step microdrama build and video models now priced from about thirteen cents a second, so the limit on your next project is contractual rather than technical.
NEWS

Suno Pulls Mary J. Blige Ad After Fake Rep Brokered Deal · 2 min read

Suno pulled a Mary J. Blige ad campaign after discovering the deal was brokered by someone impersonating her official representative. The clip itself was real, but the approval chain was fake, and Suno terminated the campaign once Blige expressed discomfort. The incident shows that celebrity likeness deals require verified rights, and a real recording doesn't equal legal clearance.

Minnesota's $500K-Per-Deepfake Law Survives xAI Challenge · 3 min read

A Minnesota court denied xAI's preliminary injunction against the state's nudification law, which fines providers up to $500,000 per deepfake access. The law stays in effect while xAI's constitutional challenge proceeds, with xAI arguing its broad definitions could criminalize harmless AI generations like swimwear images. The company also cites its enforcement record: 52,222 accounts suspended and 73,604 reports to NCMEC in 2026, and for AI-media creators, the financial risk of nudification content in Minnesota is now severe.

Oscar Winner Uses AI for 30 Seconds of New Film 'Ink' · 2 min read

Danny Boyle used AI for roughly 30 seconds of his Rupert Murdoch drama 'Ink', animating old photographs and generating digital mice for newsroom scenes. He framed it as a practical choice, noting AI is very good at creating animals and referencing his past PETA controversy over real chimpanzees. The cultural signal for AI-media makers: high-profile directors are normalizing small, tasteful AI use in prestige productions, which could ease audience and critic resistance.

ANALYSIS

The AI Channel on Your Roku Pays 50% of Net, Up to 210 Days Later · 18 min read

The free AI channel is not a Roku production. A Los Angeles company called Fairground supplies and curates it, and its creator contract is the real story: you keep copyright, but the licence is worldwide, sublicensable and transferable, payment is half of net receipts after five deduction categories and can take 210 days, liability is capped at $500 and the indemnity runs one way. Roku monetises at roughly 3.2 cents per streamed hour, which is exactly why near-zero-cost content is structurally attractive here.

Video Model Race: Pricing, Prompt Caps, and PDF References · 7 min read

Four new AI video models shipped this week, each with a different angle. ByteDance's Seedance 2.5 takes up to 50 multi-modal references and native audio at $0.47/s. MiniMax H3 offers a 7,000-character prompt cap and 4K upscaling at $0.13/s. Black Forest Labs' Flux 3 adds a draft-first workflow and strong realism at $0.29/s. Alibaba's Wan 3.0 is the value pick at $0.20/s, and its standout feature is accepting decks, spreadsheets, and PDFs as references. Full example prompts and platform availability are in the post.

TOOLS

Artlist's AI Toolkit: A Step-by-Step Microdrama Workflow · 22 min read

A practical guide walks through generating a short AI microdrama using Artlist's AI Toolkit and Studio: write the story, create consistent characters with @ tags, build locations with style references, prepare dialogue and voices, frame shots with detailed prompts, and assemble in Studio. Key techniques include using start frames, maintaining a consistent color palette, and leveraging Style Kits. Caveats: credit costs, external editors for sound and color, and story-first planning.

LLaDA-Image Generates and Edits From One Model, On Your Own Card · 1 min read

One open model that makes an image and then edits it from a plain instruction, so fixing a frame does not mean exporting to a second tool. The fast version needs 4 steps instead of 50, so you can try ten variations rather than wait on one. Weights and inference code are public, though the training code they advertised is not out yet. No VRAM figure is published, so test it on your own card first.