🔍 Read the full analysis: The One-Evening AI Build: What Worked, What Didn’t, And Why on ThorstenMeyerAI.com
Prime made for students and young adults
- Fast, free delivery for dorm and study essentials
- Prime Video and Amazon Music included
- Member-only deals
TL;DR
Thorsten Meyer says he built a hand-drawn video engine in one evening, but its first full video was slow and visually dull. He later measured specific flaws, changed the engine and made those measurements build gates; he also reported a separate AI legal-model experiment where task completion rose but the legal score remained zero.
Thorsten Meyer says he built a hand-drawn video engine in one evening, but the first full video it produced the next morning was slow and visually dull. In a November 6 account updated November 9, Meyer described how he used a second AI to review code, human approval and measured checks to revise the engine—an example of the gap between a working AI demo and a result he considered ready to publish.
Meyer said the initial idea came from a video style he admired. To avoid copying it, he wrote down a limited set of visual traits—hand-drawn strokes, a continuous world and large numbers used as color accents—and said he used no drawings, assets or layouts from the reference. Claude built the engine from scratch, using colors from Meyer’s design system and an openly licensed handwriting font.
That evening, OpenAI’s Codex reviewed the code. Meyer said it was assigned to check and challenge the work, not build it. He reported that five findings from the first review were fixed, with a test for each that failed against the earlier code. Meyer then watched a roughly one-minute test video and approved it for publication. The next morning’s first full video did not meet his expectations; he left it online as an example of the problems.
After identifying too much empty space, unhelpful icons and static finished drawings, Meyer said he measured frame coverage and stillness. Across the first three scenes of the rebuilt video, filled-frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from 3 to 27. He said those measures are now build gates: checks can stop a scene deemed too empty, and flag an object that does not illustrate the narration.
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 Video Metrics to Build Gates
The account’s practical point is that completion is not quality. A video can render and a demo can run while still failing a human review. Meyer’s process turns subjective criticism into measurements, then uses those measurements to block repeat problems before rendering. That approach may help readers distinguish an impressive prototype from a production workflow with review and safeguards.
The numbers describe a limited sample, not a general test of AI video tools: Meyer reports results for the first three scenes of one rebuilt video. The article does not establish that the same changes will improve every video or that its gates assess every dimension of quality. Human judgment remains part of the process, according to Meyer’s account.
A separate experiment in the same account reinforces the warning about choosing metrics. Meyer said a model edited for legal tasks answered all 10 published benchmark tasks, up from four, while its legal score remained zero. The higher completion count alone did not show that the model had become legally capable.
Popular right now
Roborock Q7 L5 Robot Vacuum and Mop Combo
Ninja CREAMi 7-in-1 Ice Cream Maker, NC301
TCL 65-Inch 4K UHD Smart LED TV (65S425)
As an affiliate, we earn on qualifying purchases.
The Build, Review and Revision
Meyer presents the video work as a personal account based on his build notes, incident notes and published experiments. The first short test video appeared in an article about local votes against data centers, he wrote. The first full video in the new style came the following morning and remained online unchanged. A later video, about the Nasdaq’s worst day in 19 months, used the revised approach.
The article also describes a past monitoring failure from August: a guard generated about 1,300 warnings in a log that nobody read, and an alert email did not go out. Meyer uses the incident to argue that monitoring only helps when its signals reach a person. Separately, he reported a review round in which all 12 of his own tests passed but four of a reviewer’s six checks failed. These are examples from his own work, not independently assessed results.
Meyer’s November 9 update says figures for the legal experiment now refer to 10 published Lexassay tasks, with runs publicly available on lexassay.com. The update clarifies the task count and access to the runs; it does not provide independent validation of the reported results.
““A demo that runs is not a decision.””
— Thorsten Meyer
Limits of the Reported Results
The measurements are Meyer’s own reported results; the source does not describe an independent evaluation of the videos, code review or legal benchmark. It also does not specify a broader comparison set for the three-scene video measurements, so those figures should not be read as results across multiple projects.
The account does not report how often the new gates block builds, how many scenes they flag, or whether the revised videos perform better with viewers. It also does not provide enough detail to assess the legal score’s scoring method or why it remained zero despite the model answering all 10 tasks. The published runs are identified as public, but the source does not establish that the experiment has been independently reproduced.
Further Tests and Human Review
Meyer says the video checks are now part of the build process, where they can stop or flag scenes before rendering. The account does not give a schedule for further testing or say whether additional performance results will be published. The legal-model runs are available on lexassay.com, according to the November 9 update; readers can consult those runs, but the article reports no independent replication.
For now, the process Meyer describes keeps human approval alongside automated review and measurement. Whether the gates reliably improve later videos—and whether changes to the legal model can improve its score rather than only its task completion—remains to be shown.
Key Questions
What did Meyer build in one evening?
He says Claude built a hand-drawn video engine from scratch, based on a short list of visual traits and without using assets or layouts from the video that inspired the idea.
What was wrong with the first full video?
Meyer described it as slow and boring, with too much empty space, icons that did not support the narration and drawings that stopped moving when complete.
How did the revised video change?
For the first three scenes, Meyer reported frame coverage rising from 15% to 66%, still time falling from 29 seconds to zero, and camera moves increasing from three to 27. These are figures from his account, not a general industry comparison.
Did the legal-model experiment show that the model was better at law?
Not on the reported legal score. Meyer said the model answered all 10 published tasks, compared with four before the edit, but its legal score remained zero.
Were the results independently verified?
The source describes Meyer’s own build notes and experiments and does not report an independent evaluation. It says the legal-task runs are public on lexassay.com.
Source: ThorstenMeyerAI.com
