🔍 Read the full analysis: My First AI Video Was Boring. The Building Lessons Weren’t. on ThorstenMeyerAI.com
Get the latest gadgets 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 his first full video from a new AI-assisted drawing engine was slow and boring, prompting changes based on frame coverage, stillness and camera movement. The revised checks became production gates, while a separate model-editing experiment showed that completing more legal benchmark tasks did not improve the model’s legal score.
Thorsten Meyer says his first full video made with a new AI-assisted drawing engine was slow and boring, leading him to revise the production process and add checks that can stop scenes before they are rendered. In an account published on ThorstenMeyerAI.com and updated November 9, he reports that the first three rebuilt scenes used 66% of the frame, up from 15%, and had no completely still time, down from 29 seconds.
Meyer says the engine began with a style idea inspired by a hand-drawn explainer he had watched. To avoid reproducing that work, he describes using a clean-room approach: he wrote down broad traits, including hand-drawn strokes and a continuous visual world, and says he did not reuse the original’s drawings, assets, layouts or code. Claude built the engine, while OpenAI’s Codex reviewed it. Meyer says Codex identified five issues in its first review and that each was fixed with a test designed to fail on the earlier code.
A roughly one-minute test video received Meyer’s approval and was published that evening in an article about local votes against data centers. The next morning, he watched the first full video and found it too empty, static and poorly matched to the narration. He says that video remains online unchanged as an example of the initial result.
Rather than relying only on subjective adjustments, Meyer says he measured frame coverage and the time images stood still. On the first three scenes after the rebuild, he reports that camera moves increased from three to 27, while complete stillness fell from 29 seconds to zero. He says those measurements became gates: checks can stop an overly empty scene or flag an object that does not illustrate what is being said. A revised video about the Nasdaq’s worst day in 19 months went live a few hours after his critique, according to the account.
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
Turning Video Critique Into Build Checks
The account’s central point is that generating a video is not the same as establishing that it is good enough to publish. Meyer describes a workflow combining AI code generation, independent review, tests and human approval. The review system is intended to find code problems, while the human decision remains a separate step: a working demo does not itself authorize publication.
The reported measurements also show how a subjective reaction can be translated into criteria that are checked repeatedly. That may help a production team catch certain known problems before rendering, rather than discovering them only after watching the finished video. The limits matter: the figures concern the first three scenes, and the account does not provide an independent evaluation of viewer response or overall video quality.
Meyer draws a related lesson from a separate experiment involving a local model and legal benchmark tasks. He says the edited model answered all 10 published tasks, compared with four before, but its legal score remained zero. The result is a reminder that a higher task-completion count does not necessarily mean better performance on the measure that matters.
Products Worth Considering
Bose QuietComfort True Wireless Earbuds with Active Noise Cancellation, Bluetooth 5.3, USB-C Charging, IPX4, Black
Sony WH-1000XM5 Premium Noise Cancelling Wireless Headphones – Midnight Blue
Apple AirPods Pro (2nd Generation) Wireless Earbuds with USB‑C Charging, Up to 2X More Active Noise Cancelling, Transparency Mode, Adaptive
As an affiliate, we earn on qualifying purchases.
From One Approved Demo to a Full Video
Meyer presents the account as a personal description of taking AI ideas into production, not as an independently audited study. He says the style engine was built from scratch after he recorded a short list of visual characteristics from another creator’s video. He also says the engine used his own design system, with blue for the main imagery and orange for a key point, and an openly licensed handwriting font.
The chronology distinguishes two stages that can otherwise sound like one success. First came a short test video that Meyer watched and approved. Then came the first full video the following morning, which he criticized and left online unchanged. The improved version followed after he measured specific shortcomings and added checks to the build. Meyer’s update says the figures now refer to 10 published Lexassay tasks, with runs available on lexassay.com; those benchmark details relate to the separate model experiment, not the video measurements.
“It was slow and boring.”
— Thorsten Meyer, describing the first full video
Limits of the Reported Results
The video figures come from Meyer’s own build notes and concern three scenes; the account does not say whether the same changes were measured across the entire video or across multiple productions. It also does not report viewer testing, an external review of the finished videos, or a comparison against other video-generation systems. The measurements describe specific production properties, not a general quality score.
The legal experiment is also reported by Meyer. His update identifies 10 published tasks and public runs, but the account does not supply enough information here to assess the benchmark’s scope, scoring method or how well the result generalizes beyond those tasks. The reported increase in completed answers should not be read as evidence that the model became legally reliable: Meyer says its legal score stayed at zero.
The Gates Face Future Builds
Meyer says the new checks are now part of the video build process, so scenes that fail the frame-coverage or illustration checks can be stopped before rendering. The account does not give a schedule for further releases or say whether the thresholds will change as the engine is used on more videos.
The next useful evidence would be results across additional scenes and productions, alongside a clear account of how the checks are scored and how human review affects publication decisions. For the legal-model experiment, the published runs offer a way to inspect the reported task results, but further testing would be needed to show whether any edit improves performance beyond the 10 tasks.
Key Questions
What happened to Meyer’s first full AI-generated video?
Meyer says it was slow and boring because of empty space, static images and weak alignment between icons and narration. He says it remains online unchanged as an example of the initial result.
What changed in the revised video process?
Meyer says he measured frame coverage and stillness, then made checks for those issues part of the production build. He also reports more camera moves in the first three rebuilt scenes.
Did the model-editing experiment improve legal performance?
Not on the reported legal score. Meyer says the model answered 10 published tasks instead of four, but its legal score remained zero.
Were the videos or benchmark findings independently evaluated?
The source account presents Meyer’s own build notes, incident notes and experiments. It does not report an independent evaluation of the videos or the benchmark results.
Source: ThorstenMeyerAI.com
