🔍 Read the full analysis: One Evening, One AI Video, And A Few Hard-Won Rules on ThorstenMeyerAI.com
Get audio and creator gear delivered free with Prime
- Fast, free delivery on millions of items
- Prime Video, Amazon Music and more included
- Member-only deals all year
TL;DR
Thorsten Meyer says he built a hand-drawn AI video engine in one evening, but its first full video was slow and visually sparse. After measuring frame coverage, still time and camera movement, he added checks to the production process; the account is a first-person report, not an independent evaluation.
Thorsten Meyer says he built a new hand-drawn AI video engine in one evening, but the first full video it produced the next morning was slow and visually sparse. In an account published November 6 on ThorstenMeyerAI.com, he describes how he measured the shortcomings, revised the system and introduced checks before rendering and publication—a case study in the gap between a working AI demo and a reviewed production result.
Meyer writes that the visual idea came from a video he admired, but says he aimed to capture only broad style traits rather than reproduce its drawings, assets or layouts. He describes the resulting look as a continuous whiteboard-style world, with colors from his own design system and an openly licensed handwriting font. He says Claude built the engine from scratch.
In his workflow, OpenAI’s Codex reviews the code but does not build it. Meyer says Codex identified five issues in the first review round, which were fixed that evening. Each fix, he writes, received a test designed to fail against the old code. He also says a human review remains a separate decision: after watching a roughly one-minute test video, he approved it for publication inside an article about local votes against data centers.
The first full video, produced the next morning, did not meet his expectations. Meyer describes excessive white space, icons that did not support the narration and images that stopped moving after they were drawn. He measured frame coverage and time spent completely still, then changed the engine. Across the first three scenes of the rebuilt video, he reports that filled frame area rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. These are figures from his own build notes, not an independent assessment.
Infographic 01 · The new video look
One evening, one boring first video, one measured fix
Speed was not the point. The steps in between were: a second AI that reviews, a human who says yes, and numbers that became gates.
Share of the frame filled with content on the first three scenes after the fix. Before: 15%.
From an evening idea to a measured fix
Relative times, this week
- EveningA video I likedClean room: only a list of style traits, no drawings, assets or layouts
- Same eveningOwn engine, first reviewClaude builds; OpenAI’s Codex reviews; 5 findings fixed, each with a test that fails on the old code
- Same eveningA human says yesOne-minute test video approved and published
- Next morningFirst full video: slow and boringToo much white space, icons that did not fit, a picture that stood still
- Hours laterMeasured fix, then gatesThe first video in the improved look goes live
Source: my own build notes for the new video look, this week (engine history, review round, approval, critique, before/after measurement of the first three scenes).
Powered by Thorsten Meyer AI
From Demo to Production Checks
The account’s practical point is that generating a video is only one part of putting an AI system into use. Meyer distinguishes between software that runs and work that has been reviewed by a responsible person. He also describes turning subjective criticism—such as a scene feeling empty—into measurable checks that can stop a render before it is published.
That approach matters to teams using generative tools because a polished-looking output can still have defects, and a successful test suite can miss them. Meyer gives a separate example from a model-editing experiment: a model completed 12 of 15 answers on a legal benchmark, compared with four before the edit, while its reported legal score remained zero. The example cautions against treating a higher completion count as proof of better performance. The source does not provide enough benchmark detail to judge the experiment independently.
Meyer also recounts a monitoring failure from August: a guard generated about 1,300 warnings in a log that nobody read, while an alert email did not go out. The episode supports his broader argument that safeguards matter only if they reach someone able to act. It is a reported incident from his own system, not a general measure of AI reliability.
Products Worth Considering
Samsung Galaxy Tab A11 8.7-Inch Wi-Fi Tablet, 64GB, 4GB RAM, Silver
Samsung Galaxy Tab A11 Wi-Fi Tablet, 8.7-Inch, 64GB, 4GB RAM, Gray (SM-X133, International Version)
Samsung Galaxy Tab A11+ 11-Inch Tablet, 6GB RAM, 128GB Storage, Expandable, Gray
As an affiliate, we earn on qualifying purchases.
How the Video Was Rebuilt
Meyer says the project began with a video whose style he liked. To avoid copying its specific creative work, he says he recorded a short list of general traits—hand-drawn strokes, a continuous setting and large numbers used as color accents—and used no source drawings, assets or layouts. He describes using his own design system for color and building the drawing engine from scratch.
The first short test and the first full video had different outcomes. Meyer approved the roughly one-minute test for publication the same evening it was made. The longer video, generated the following morning, was left online unchanged as an example of what he considered unsuccessful. After writing down his criticisms, he measured frame coverage and still time, then altered the engine so camera movement and continuing motion were part of the output. He says content also has to illustrate what the narration is saying at that moment.
The account links to a roughly four-and-a-half-minute explainer about the process. Meyer says its narration uses an AI voice clone of himself. He also describes a separate experiment in which Claude edited a local model’s weights for a legal task; the reported increase in completed answers came without a change in the legal score. The source presents these as personal build notes and experiments, rather than a controlled comparison of video or model systems.
“It was slow and boring.”
— Thorsten Meyer, describing the first full video
What the Report Does Not Establish
The performance figures come from Meyer’s own notes. The account does not include independent measurements, a complete description of how frame coverage and still time were calculated, or comparisons with other video-generation systems. It also does not report how viewers responded to the revised videos or whether the new checks reduced errors over a longer period.
Details of the legal benchmark and the weight-edit experiment are limited in the supplied account. The 12-of-15 and four-of-15 completion counts do not explain the scoring method, the full test set or what the unchanged legal score represents. The report therefore does not establish that the model became more capable at legal reasoning. It also does not identify the original video Meyer admired, so readers cannot independently compare the works or assess the creative relationship from this account alone.
Checks in Future Video Builds
Meyer says the revised production rules are now built into the video engine: a scene judged too empty should be stopped before rendering, and a build should flag an object that does not illustrate the spoken content. The next step described in his account is applying those checks to subsequent videos, including the explainer linked with the article.
The source does not announce a release schedule, a broader product launch or an independent audit of the workflow. Readers can expect further evidence only if Meyer publishes additional build notes, test results or examples showing how the gates perform across more videos. For now, the report documents one creator’s production process, its first failure and the safeguards he says he added afterward.
Key Questions
What is the news in Meyer’s account?
Meyer published a first-person report describing how he built and revised an AI-assisted, hand-drawn video engine, then added measurements and checks to address problems in its first full video.
What changed in the revised video?
Across the first three scenes, Meyer reports that frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. These are his own measurements.
How does the review process work?
Meyer says Claude builds the system and OpenAI’s Codex reviews its code. He also describes writing tests for review findings and requiring a human to watch and approve a video before publication.
Does the legal-model experiment show improved legal ability?
Not on the evidence provided. Meyer reports more completed answers—12 of 15 rather than four—but says the model’s legal score remained zero. The account does not provide enough benchmark detail to assess the result independently.
Has the video workflow been independently evaluated?
The source does not report an independent evaluation. Its figures and examples are described as Meyer’s own build notes, incident notes and experiments.
Source: ThorstenMeyerAI.com
