🔍 Read the full analysis: The Lessons Behind One Evening Of AI Video Building 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 he built a hand-drawn video engine in one evening, but its first full video was slow and visually weak. He describes how code review, human approval and measurable checks helped improve later videos, while a separate legal-model experiment showed why task completion alone does not prove quality.
Thorsten Meyer says he built a hand-drawn video engine in one evening, but its first full video the next morning was slow and visually boring. In an account published on ThorstenMeyerAI.com and updated November 9, he describes how code review, his own approval and measurements helped reshape the engine—and argues that a system finishing a task does not establish that its result is good.
Meyer says the initial idea came from a video style he admired. To avoid copying the original, he says he wrote down only general traits, including hand-drawn strokes, a continuous visual world and large numbers used as color accents. He says the team used no drawings, assets or layouts from that video, and built a new engine from scratch. Its colors came from his design system, and the handwriting font was openly licensed.
In Meyer’s account, Claude built the engine, while OpenAI’s Codex reviewed the code without making changes. Codex identified five issues in its first review, he says; those were fixed the same evening, with each fix paired with a test that failed on the earlier code. Meyer watched a roughly one-minute test video and approved it for publication. A full video produced the next morning did not meet his expectations: he cited excess empty space, icons that did not support the narration and finished drawings that remained still.
Meyer says the team measured frame coverage and the time images stayed still, then changed the engine and made checks part of the build. Across the first three scenes, he reports, filled-frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. He says the updated video, about the Nasdaq’s worst day in 19 months, went live a few hours after his critique. The original weak video remains online, unchanged, as an example of the first version’s shortcomings.
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 is a case study in the difference between producing a demonstration and creating a repeatable production process. Meyer’s central point is that a video can render successfully while still failing a human viewer. His response was to turn specific criticisms into measurable checks, rather than rely only on repeated visual tweaking.
That approach may help teams catch known defects earlier: according to Meyer, the engine now stops scenes that are too empty and flags objects that do not illustrate the narration. But those checks cover only qualities the team has chosen to measure. Frame coverage and camera movement do not, by themselves, establish that a video is accurate, clear or engaging. The account therefore offers both a workflow and a limit: metrics can enforce standards, but they do not replace editorial judgment.
Popular right now
Samsung 43-Inch M70H Series 4K Mini LED Smart TV (2026)
Philips 3300 Series Fully Automatic Espresso Machine with Milk Frother, Grinder, Hot and Iced Coffee Presets, SilentBrew, 15-Bar Pressure (E
Vitamix Touch & Go Advance 32-Ounce Bar Blender, White
As an affiliate, we earn on qualifying purchases.
A Separate Test of Model Quality
Meyer connects the video work to a separate experiment conducted the same week, asking whether directly editing a local model’s weights could improve its legal-task performance. He says Claude made the edit and that the model answered all 10 published Lexassay tasks, compared with four before. The updated article says the runs are public on lexassay.com.
That higher completion count did not mean the model passed the legal benchmark: Meyer says its legal score remained zero. He also reports that, in a separate review round, all 12 of his own tests passed while four of a reviewer’s six checks failed. These examples are his reported experiments, not independent evaluations included in the supplied account. Together, they frame the article’s recurring warning: a favorable or green test result can miss failures if the test does not measure the outcome that matters.
“A demo that runs is not a decision.”
— Thorsten Meyer
What the Report Does Not Establish
The account is a first-person report based on Meyer’s own build notes, incident notes and experiments. The supplied material does not include independent verification of the video measurements, the Codex findings, or the legal benchmark results. It also does not explain the benchmark’s scoring method or why the model’s legal score stayed at zero after it answered all 10 tasks.
The video figures apply to the first three scenes, not necessarily the full video or later projects. The account does not provide a formal audience evaluation, comparisons with other video tools, or a measure of how often the new gates catch defects in routine use. It also says an August guard produced about 1,300 warnings in a log nobody read and that an alert email did not go out; it gives no further details about the system or the changes made afterward.
Whether the Checks Hold Up
The account does not announce a scheduled release, follow-up test or independent audit. Meyer says the improved video is live and the runs for the 10 published Lexassay tasks are public. The practical next test is whether the video gates continue to flag weak scenes in later builds—and whether human review catches problems that the measured checks miss.
For the legal experiment, the published task results do not resolve the mismatch between completing all tasks and retaining a zero legal score. Further detail on scoring and evaluation would be needed to judge what the weight edit changed. Until then, Meyer’s examples support a narrower conclusion: completion, passing tests and production readiness are different claims, and each needs evidence suited to the outcome being judged.
Key Questions
What happened in the AI video project?
Meyer says he built a hand-drawn video engine in one evening. The first full video was unsatisfactory, so he measured visual issues, revised the engine and added checks to later builds.
What changed in the video after the revision?
For the first three scenes, Meyer reports that filled-frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. Those figures describe those scenes, not necessarily the entire video.
What roles did Claude and OpenAI’s Codex have?
According to Meyer, Claude built the engine and Codex reviewed the code. He says Codex reported five findings in the first review, which were fixed with tests designed to fail on the earlier code.
Did the legal model pass its benchmark?
No. Meyer says the model answered all 10 published tasks after a weight edit, up from four, but its legal score remained zero. The account does not explain the scoring method in detail.
Are the results independently verified?
The supplied account presents them as Meyer’s own build and experiment results. It points to public runs for the 10 Lexassay tasks, but does not provide independent verification of the video measurements or a separate evaluation of the reported findings.
Source: ThorstenMeyerAI.com
