Practical notes on AI film and video production, creative workflow systems, agents, and the AGI era. Written from running this work daily, not from a distance.
We reviewed more than thirty festivals for our AI short film Last One and entered twelve, in seven countries. The five questions that sorted them, the patterns that ruled the others out, and the full calendar.
Our short film Last One follows one character through one rainy night, with no set and no camera. What the film is about, the five-step method behind it, and why continuity is decided before anything is generated.
Rogue One recreated Peter Cushing with his estate’s permission and still ended up in court. The Irishman hired a posture coach. What Hollywood’s digital doubles teach brands about AI avatars.
Fleischer’s 1917 patent, Snow White’s live model, AI-assisted roto on Everything Everywhere All at Once and the Academy’s 2025 rule on generative AI: what a century of tracing real footage teaches AI video.
Sound fixed film at 24 frames per second in 1927, and audiences balked at 48 and 120. AI video models set their own speed and paint their own blur, so motion has to be decided shot by shot.
$2 and $10 per million tokens against $10 and $50, and one point apart on an independent index. Where GPT-6.1 Sol keeps up with GPT-6 Astra, where it does not, and a simple rule for choosing.
Disney pinned story sketches to boards in the early 1930s; Lucas cut wartime footage into his Star Wars rough cut. What testing a film before the shoot teaches about AI drafts, first frames and motion reference.
30-second clips, 4K, 10 keyframes, 15 references: what Kling 4.0 really changes for a brand, and the one thing a longer clip does not fix.
One neutral face, three emotions: the Kuleshov effect. Why a string of beautiful AI shots can still say nothing, and where the meaning of a film really comes from.
AI video is judged on the image and fails on the sound. What Star Wars, Jaws and an MIT study teach about why sound decides whether a video feels expensive.
In 2023 the internet copied Wes Anderson by the million, with phones and with AI. Why a style anyone can reproduce stops being a signature, and how a brand builds its own.
Rope hides five of its ten cuts, and 1917 plays as two shots. How the one-take illusion works, and why the same hidden cuts now stitch AI-generated clips together.
Apple “1984”, Guinness “Surfer” and Honda “Cog” look nothing alike and share the same skeleton: a hook, a turn and a payoff. Why it matters more now that images are cheap.
Warm skin, cool shadows: the look behind most trailers. Why orange and teal works, what The Matrix, Mad Max and Amélie did instead, and why AI video defaults to it.
Two AI-produced holiday ads: one renewed, one pulled within days. The difference was not the tool. It was direction.
Nearly two in three video ad buyers use generative AI, yet 96% of small advertisers are not satisfied with their use. What is missing, and how to fix it.
Almond Breeze, Equinox and Dollar Shave Club mock sloppy AI content, using AI. What audiences actually accept, and what they reject.
A product image, a budget, and Meta’s AI does the rest by the end of 2026. Why brands are pushing back, and why creative volume is the real story.
Coca-Cola and McDonald’s hired AI studios, with teams and weeks of work. What that says about the craft, and five questions to ask before hiring a studio.
Cutaways, exploded views and a plant that does not exist yet. What generated video reaches for manufacturers, the method it takes, and the three things that still belong on camera.
Article 50 of the EU AI Act applies since 2 August 2026. Who counts as the deployer, what a compliant disclosure looks like on a video, and why Canada having no rule is not the same as no risk.
Six tells a viewer spots in about two seconds, from faces that drift between shots to motion with no weight, and why each one happens.
Every consistency trick in AI video points back to one reference. What goes on the sheet, why it lives on grey, and the checks that stop a face drifting before production starts.
ByteDance pulled its Vision Pro competitor's launch days before it happened. The official reason is five vague words. The unconfirmed reason is more interesting.
30-second single-shot clips and 50 reference inputs are real. The pricing, the 4K claim and the copyright exposure are not settled yet.
SB 1050 makes AI-generated performers in ads a disclosure duty. It names no penalty, which reads as weakness until you see which statutes it routes into.
A frontier video product went from a dated model ID to a dead endpoint in under a year, with a billion-dollar partner lost inside that window. What it proves, and what to own instead of an endpoint.
Output tokens cost 5 to 6x input tokens everywhere, and a Chinese model can undercut both by two orders of magnitude. The mechanism behind both numbers.
The number everyone quoted came from the test harness rather than the model. A read of the benchmark record, and what it means for production work.
Not a ranking. A dated read of the model landscape, including the shutdown eleven days out that a lot of pipelines have not noticed yet.
Most popular prompting advice has now been measured at scale, and a surprising amount of it does nothing. What the studies actually found.
Asking whether to use a CLI agent or MCP is a category error. What each is, when to reach for which, and the security part nobody reads.
Only two platforms publish official safe zones. Every other guide is reverse-engineered, and they disagree by more than twice.
The model does not inherit the previous shot's setup. What actually holds lighting across separately generated shots, and what is folklore.
Model rankings are obsolete on publication. Here is the decision method that outlives them, and the metric that actually predicts what a project costs.
The most common failure in AI video is a face that drifts. Seeds will not fix it. References, frame chaining and a fixed identity block will.
These models already speak film grammar. Most prompts never use it. The vocabulary, the prompt order that works, and why one move per shot beats three.
Adoption is up, marketer satisfaction is down, and most consumers still trust a real face more. The cost argument won. The credibility argument did not.
Six definitions are in active use, and changing the maths moves the score by thirty points. What the term actually means in 2026.
A system can be superhuman and narrow at the same time. Why generality is an axis rather than a line you cross.
Four measured gaps separate producing fluent output from general intelligence, and each one changes what you build around the model.
The same model scored 62.7% and 99.9% on one benchmark in a single day. Autonomy is a decision someone makes, not a capability tier.
One benchmark fell from 83.5% to 20.6% when the tasks got longer. Where agents work, where they break, and what to ask before buying one.
Adoption keeps climbing while reported earnings impact sits flat at 37%. What separates the companies getting value from the ones that are not.
AI raises individual creative output and makes everyone's work more alike, by a measured amount. What that means for a studio.
AI creative wins the click and loses the sale at high consideration. Plus zero-click search at 68% and what disclosure really costs.
One study finds early-career employment 19% below trend in exposed jobs. Another finds almost none of it is job losses. Both are serious.
There is no arrival date to plan around. Six things the evidence says to do, each of which pays off whether or not anything called AGI shows up.
No articles published yet. They will appear here as they go live.
If any of this maps to what your team is trying to do, the fastest next step is a conversation.