🔍 Read the full analysis: What Does It Take To Bring An AI Idea To Production? One Evening Taught Me on ThorstenMeyerAI.com
Get hardware and tech essentials 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. He describes using a second AI for code review, human approval and measurable production gates to improve later videos; the account is a personal report, not an independent evaluation.
Thorsten Meyer says he built a hand-drawn AI video engine in one evening and put a short test video live that night, but the first full video produced the next morning was, by his account, slow and visually dull. In a November 6 personal report, Meyer describes how independent code review, human approval and measurable quality gates shaped the system before later videos went live.
Meyer says the concept came from a video style he admired, but he set a clean-room rule to avoid copying its drawings, assets or layouts. He says the team documented a short list of broad traits, including hand-drawn strokes and a continuous visual world, then used Claude to build a new engine. Its colours came from his own design system, and the handwriting font was openly licensed, according to his account.
He says OpenAI’s Codex reviewed the code but did not build the engine. Codex reported five findings in its first review round, Meyer writes, and each fix was paired with a test designed to fail on the old code. Later, Meyer watched a test video of about a minute and approved it; that version appeared in an article on local votes against data centers. The next morning’s full video did not meet his expectations: he identified excess white space, icons that did not support the narration and images that stopped moving after they were drawn.
Rather than rely only on subjective tweaks, Meyer says he measured how much of the frame contained content and how long scenes remained still. Across the first three scenes of a rebuilt video, frame coverage rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. Meyer says these measures became production gates: scenes that are too empty can be stopped before rendering, while objects that do not illustrate the narration trigger a warning.
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
How Review Became a Production Gate
The account focuses on a distinction that matters to teams deploying generative tools: a working demo is not the same as an approved production result. Meyer describes separate roles for code generation, adversarial review and human judgment, with tests intended to catch specific defects before they recur. The workflow offers a concrete example of how an AI-assisted project can be checked beyond whether it runs.
His reported measurements also show how qualitative feedback can be turned into operational checks. In this case, frame coverage and motion were tracked, then used as gates rather than informal preferences. Those figures describe the first three scenes of one rebuilt video; they do not establish a general performance benchmark or prove that the system works equally well across other subjects and formats.
Meyer also highlights limits in testing and monitoring. He says that in one separate review round all 12 of his own tests passed while four of a reviewer’s six checks failed. He also recounts an August incident in which a guard generated about 1,300 warnings in a log nobody read and an alert email did not go out. Together, the examples underline that tests and alerts only help when they detect relevant failures and reach someone able to act.
Products Worth Considering
LG UltraGear 32GX870A-B 32-inch 4K OLED Gaming Monitor
Logitech MX Anywhere 3S for Mac Compact Wireless Mouse, Space Gray
Apple MacBook Pro 14-inch (2021) – M1 Pro Chip, 10-Core CPU, 16-Core GPU, 16GB RAM, 1TB SSD, Space Gray
As an affiliate, we earn on qualifying purchases.
From Style Reference to Live Video
The article is a first-person account by Meyer, not a report of an independently audited product launch. He says the short test video went live the evening he built the engine, within an article about local opposition to data centers. The first full video in the style followed the next morning and remains online unchanged as an example of what he considered unsuccessful.
After critiquing that video, Meyer says the engine was adjusted to fill more of the image, move the camera with each sentence, keep completed drawings slightly animated and connect each object to something being said at that moment. A later video, on what Meyer describes as the Nasdaq’s worst day in 19 months, went live a few hours after his critique. The supplied account does not provide the publication date for that market event or independent verification of the comparison.
The same week, Meyer says he ran a separate experiment with a local model and a legal-answer benchmark after researchers demonstrated that one AI could repair another by editing its weights. Claude made the edit; the model completed 12 of 15 legal answers, compared with four before, while its legal score remained zero. Meyer presents this as a warning that completion counts alone can misrepresent performance, not as evidence that the model improved at legal reasoning.
“A demo that runs is not a decision.”
— Thorsten Meyer
What the Account Does Not Establish
The performance figures come from Meyer’s own build notes, incident notes and experiments, as described in his article. The source provides no independent audit, full benchmark methodology or comparative evaluation of the video engine. It is also unclear how the frame-coverage and motion checks were defined, what thresholds trigger a block, and whether the reported gains hold across a larger set of videos.
The account does not identify the legal benchmark, explain how its legal score is calculated, or provide the full results of the weight-edit experiment. It also does not give details about the five code-review findings or the reviewer’s six checks. These omissions limit what readers can infer from the examples; the reported results should not be treated as proof that AI-generated code or legal answers are reliable.
Further Tests Will Show Durability
Meyer says the video checks now act as gates in the production process, but his account does not set out a public release schedule, future milestones or plans for independent testing. The next evidence that would clarify the system’s reliability would include results across more videos, documented gate thresholds and details on how human approval and failed checks are recorded.
For the separate model experiment, the supplied article excerpt does not include the full story or say whether additional legal evaluations are planned. Until more information is available, the reported increase in completed answers should be read alongside the unchanged legal score and the limits Meyer describes.
Key Questions
What did Meyer build?
He says he built a hand-drawn AI video engine in one evening, using Claude to write the code and OpenAI’s Codex to review it.
Why was the first full video revised?
Meyer says it had too much white space, icons that did not support the narration and images that stayed still after drawing. He measured frame coverage and still time before changing the engine.
What changed in the first three rebuilt scenes?
According to Meyer, frame coverage rose from 15% to 66%, still time dropped from 29 seconds to zero, and camera moves rose from three to 27. These figures apply to the first three scenes, not a broad independent test.
Does the legal-model experiment show improved legal ability?
Not on the evidence provided. Meyer says the model completed 12 of 15 benchmark answers after a weight edit, up from four, but its legal score remained zero. The benchmark details are not included in the source material.
Source: ThorstenMeyerAI.com
