← Back to archive
Issue #59 · Monday, November 16, 2026 · 8 stories

🎥 Grok's Video Team Is Down To A Handful 🎨 Seedream 5.5 Says It Can Edit One Image 100 Times

Firefly's safe label stops at Adobe's own models, and DeepL's new voice model keeps your pauses.

This week's news lives in the fine print. Meta's new terms would have made you answer for its AI, and they are paused, not gone. Firefly's commercially safe promise covers Adobe's models, not the partner models in the same menu. Even Seedream's 100 rounds of edits come from ByteDance's own tests. Read the terms before you bill a client for the result.
NEWS

Grok Imagine loses about 60 staff, leaving a handful to keep the video tool running · 4 min read

SpaceX cut roughly 60 people from the Grok Imagine team last week, The Information reported. Only a handful of staff remain to keep the image and video tool running. SpaceX has not said it will shut Grok Imagine down, and it still works today. The cuts follow a shift towards AI coding and agents. If Grok Imagine is in your workflow, keep a second model ready.

Manchester Animation Festival will stop taking generative AI films from 2027 · 3 min read

Manchester Animation Festival will not accept films made with generative AI from its 2027 edition. Director Steve Henderson says the new policy puts artists first. The festival has not said how it defines generative AI or how it will check entries. If you make AI animation for the festival circuit, read each festival's rules before you start a project, not after you finish it.

Meta pauses the ad terms that made advertisers solely responsible for its AI output · 4 min read

Meta has paused its new business and ad terms, which were due to start on 20 November. They would have made advertisers solely responsible for what Meta's AI features produce, including features Meta switches on by default. Courts in Brazil and Norway issued injunctions, and the UK's competition regulator is still investigating. The old terms still let Meta crawl advertiser sites for AI training. Blocking Meta-ExternalAds and Meta-ExternalAgent in robots.txt is the control you have.

ANALYSIS

Half of October's new Steam games disclosed AI, but few of them found players · 8 min read

49.9% of new Steam games in October disclosed AI in player-facing content. That was 1.4% when Valve's disclosure rule began in January 2024. The count comes from a Reddit user, Ploz, working from Steam's public disclosure data. Art was the most common use at 63%, then text at 30% and music at 21%. Among new games with more than 1,000 reviews, only 17% disclosed AI. The disclosed games are piling up at the bottom of the store.

Adobe's commercially safe promise covers Firefly's own models, not the partner models beside them · 5 min read

Firefly now hosts models from ByteDance, Google, Kling, OpenAI and ElevenLabs next to Adobe's own. Creative Bloq's Ian Dean points out that Adobe's licensed-training promise only ever applied to Firefly models. Partner models carry their own terms and training histories, and the app does not make that clear when you generate. If a client is paying for commercial safety, use a Firefly model and note which one you used.

TOOLS

ByteDance's Seedream 5.5 promises 100 rounds of edits without wrecking the rest of the image · 1 min read

ByteDance's Seed team put out Seedream 5.5 on 9 November. The pitch is repeat editing: change one part of an image and the untouched areas stay the same, pixel for pixel, across 100 rounds. It also outputs transparent RGBA layers, so cut-out assets need no background removal. ByteDance says severe text errors fall to 0.4%. Those figures are from ByteDance's own tests. The post gives no price, and a Pro version is still to come.

DeepL's Aure voice model now carries pace, emphasis and sarcasm into the translated voice · 1 min read

DeepL's Aure model arrived on 11 November. It translates speech in real time and keeps the speaker's own voice, plus shifts in emotion, emphasis and pace. A pause before an announcement stays a pause. It works in eight languages at launch: English, German, Spanish, French, Italian, Japanese, Portuguese and Korean. It runs in DeepL Voice and the DeepL API for Voice with no setup. DeepL gives no price or latency figure, and it is built for live speech, not finished dubs.