Issue #37 · Monday, September 21, 2026 · 7 stories

⚖️ California Makes You Label Your AI Actors 🎬 The AI Short That Cost $29,575

Plus: Qwen's open weights you are not licensed to sell, and Seedance in half the generations.

I was away for a week, so this one covers a fortnight of ground. The through line is that the free-and-instant story about AI video is quietly falling apart: PJ Ace put a real number on a ten-day AI short, $29,575, and led with the line most vendors avoid, that AI is not cheap. California now makes you label synthetic performers in ads, Qwen shipped 7B open weights you are not licensed to sell, and the labels want $9 billion from Suno over what v6 was trained on. The tools got better and the bill arrived with them. In the back half: the $3.75 account that puts an AI track on a real band's Spotify, a Claude loop that halves your Seedance generations, and Runway pushing any clip to 120fps.

NEWS

California Requires Labels on AI Performers in Ads · 2 min read

Governor Gavin Newsom signed SB 1050, making California the second state after New York to require disclosure when ads use AI-generated synthetic performers. Backed by SAG-AFTRA, it mandates visible labelling of AI actors in video or audio ads and bars continued distribution of non-compliant ones. If you ship ads with AI avatars into California, that is a direct compliance rule: label the output or face removal. It changes how you present the result, not what you can make.

Labels Sue Suno for $9B Over v6 Models Trained on Pirated Outputs · 6 min read

Universal and Sony filed a second lawsuit against Suno, alleging its v6 models were trained on outputs of earlier models that used unlicensed copyrighted recordings. The labels seek $9 billion in statutory damages for 60,202 recordings plus $150 million for bypassing YouTube's anti-downloading tech. The suit argues that training on infringing model outputs doesn't launder the original infringement, and points to Suno's recent licensing deals with Warner, BMG, and Believe as proof a legitimate market exists.

Any Video to 120fps: Runway's Frame Rate Model Claims 7x Speedup · 2 min read

Runway's Enhance Frame Rate model converts any video to broadcast-standard frame rates (25, 30, 48, 60, 120 fps, plus NTSC's 59.94). Runway claims it runs up to 7x faster and 3x cheaper than competing interpolation models, preserving 4K resolution and clips up to five minutes. Pricing is flat: one credit per two seconds of video, regardless of resolution or target frame rate.

TECHNICAL

Claude Feedback Loop Cuts Seedance Generations from 10 to 5 · 3 min read

A builder documented a workflow where Claude writes prompts from voice memos, pre-flags likely failures (weapon discharges, brand-name cars turning into cop cars), and reviews tiled frame stills to rewrite prompts before spending credits. This reduced a 25-second one-take clip from 10 generations to 5. A $30 mistake from dropping a raw MP4 into Claude was fixed with a single FFmpeg command that tiles frames into contact sheets for pennies.

ANALYSIS

This AI Short Cost $29,575 and Took 10 Days. AI Is Not Cheap · 1 min read

PJ Ace published an itemised budget for Nexus Episode 4, a 10 day AI production that came to $29,575, and led with the line most tool vendors avoid: AI is not cheap. The money goes on compute, tooling and people, chiefly the animators and editors doing hundreds of image references before a frame moves. It is the clearest public benchmark yet for costing an AI short, and worth holding against any pitch that an AI film is a weekend and a subscription. He also says the Nexus feature will still cost millions next year, so treat sub-$30k as the short-form number, not the ceiling. The limit on the receipt: the thread gives the total and the process rather than a line by line split, so use it to size a budget, not to model one.

A $3.75 Distributor Account Puts an AI Track on Any Band's Spotify · 4 min read

404 Media put an AI track made in Udio onto the verified Spotify, Apple Music, Tidal and Amazon pages of Brooklyn band Lathe of Heaven. The cost was a $3.75 monthly DistroKid account and five minutes. Nobody verifies identity: you tick a box saying you are authorised and the pipeline trusts it. Spotify conceded that profiles which are not actively managed are especially vulnerable, which describes most independent artists.

TOOLS

Qwen-Image-2.1 Runs 7B on a Consumer GPU. The Licence Stops You Selling It · 1 min read

Alibaba's Qwen team put out Qwen-Image-2.1 on 20 September with a 7B visual generation component, native 2K output, RGBA transparency straight from the model, and composition across up to 10 reference images. Transparency without a rotoscope pass and ten-reference identity holding are the two features that change a real workflow, and it runs in bfloat16 on a capable consumer card. Day zero support landed in ComfyUI, Diffusers, vLLM-Omni and SGLang, so there is no waiting on integrations. The catch is the licence, not the model: it ships under the Qwen Research License, which covers research and evaluation only, and commercial use needs a separate agreement from Qwen. Open weights you can study, not open weights you can put in client work, so test it before you quote it into a job.