🔍 Read the full analysis: What Does It Take To Bring An AI Idea Into Production? 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 describes building an AI-generated video engine and publishing its work after code review, human approval and revisions guided by measured results. His examples also show why passing tests or completing more benchmark tasks does not, by itself, establish that an AI system is working well.
Thorsten Meyer says he took a hand-drawn AI video idea from an evening prototype to a published video within hours, but the release followed code review, a human approval step and revisions based on measured flaws. In a November 6, 2026, first-person report updated November 9, Meyer presents the project as a case study in why a working demo is not enough to establish that an AI system is ready for production.
Meyer says the concept began with a video style he admired. To avoid copying the original, he says the team wrote down a limited set of visual traits and built a new engine without using the source video’s drawings, assets, layouts or drawing code. Claude built the engine; OpenAI’s Codex reviewed it. Meyer describes Codex as a separate checker that reads code and reports problems rather than building the system. He says five findings from the first review were fixed that evening, each paired with a test designed to fail against the earlier code.
A roughly one-minute test video received Meyer’s approval and was published that evening. The first full-length video produced the next morning fell short of his expectations: he described it as slow and boring, citing empty space, icons that did not support the narration and images that stopped moving after being drawn. He says that version remains online as an example of what the team wanted to improve.
Rather than relying only on subjective tweaking, Meyer says the team measured frame coverage and the time images remained still. Across the first three scenes of a rebuilt video, he reports that the filled portion of the frame rose from 15% to 66%, still time fell from 29 seconds to zero, and camera moves increased from three to 27. Those measurements became production checks: scenes that are too empty are stopped before rendering, and the build flags objects that do not illustrate the narration. These figures describe three scenes in this project, not a general benchmark for AI video systems.
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
Why Production Needs More Than a Demo
The account makes a practical distinction between a system that can produce an output and one whose output is fit to publish. Meyer’s first full video rendered, but the team’s review found shortcomings that a successful run alone would not reveal. That matters to organizations deploying AI because a completion signal can be mistaken for quality unless teams define what acceptable output means and check it.
The reported workflow divides responsibility across different steps: one model builds, another reviews, automated tests check identified problems, and a human decides whether the result can go live. The video metrics then turn specific critiques into repeatable checks. This does not establish that the approach will work for every product, but it shows how a team can make particular quality expectations visible before release rather than relying on a favorable impression of a demo.
Meyer’s separate legal-model experiment reinforces the caution. He says a model answered all 10 published tasks in a benchmark after a weight-editing experiment, compared with four before, while its legal score remained zero. The report does not supply enough methodological detail in the supplied material to evaluate the experiment independently. Its stated lesson is narrower: answering more tasks is not equivalent to improving the capability the benchmark is meant to measure.
Products Worth Considering
Amazon Kindle Scribe (3rd Generation), 64GB, 11-inch Display, Graphite, with Premium Pen
DJI Mini 4 Pro Fly More Combo Drone with DJI RC 2
EcoFlow Delta 3 Max Portable Power Station with 220W Solar Panel, 2048Wh
As an affiliate, we earn on qualifying purchases.
From Style Idea to Measured Checks
Meyer’s account is a personal report based on his own build notes, incident notes and published experiments, rather than an independent evaluation of AI production practices. The timeline he gives is short: he had the video-style idea one evening, built and reviewed an engine that night, published a short test after approving it, and assessed the first full video the following morning. A revised video went live a few hours after that critique.
The project also involved a deliberate design boundary. Meyer says the team used broad visual characteristics—such as hand-drawn strokes, a continuous world rather than slides, and large numbers as accents—while creating its own assets and design choices. The account says the palette came from his design system and the handwriting font was openly licensed. These details explain how he says the team separated inspiration from copying; they are not an external legal assessment.
The article’s November 9 update says figures now refer to the 10 published Lexassay tasks and that the runs are public on lexassay.com. That update applies to the legal benchmark example, not to the three-scene video measurements. Meyer also describes an earlier monitoring failure from August: a guard recorded about 1,300 warnings in a log nobody read, while an alert email did not go out. The example highlights a gap between generating a warning and getting it in front of someone able to act.
““It was slow and boring.””
— Thorsten Meyer, describing the first full video
What the Report Does Not Establish
The account does not provide an independent review of the video engine, the code-review findings, or the benchmark experiment. The supplied material also does not specify the full scoring method behind the legal score of zero, the model’s baseline conditions, or enough detail to assess how the weight edit affected performance beyond the 10 published tasks. The task count is identified in the November 9 update, but the result should not be read as evidence of broader legal competence.
The video figures cover only the first three scenes of the rebuilt production, and the report does not offer a comparison against other video systems or a controlled evaluation of audience response. It also does not describe how the team verifies every automated check over time, or how often a reviewer’s findings are missed. Meyer’s examples document one creator’s workflow and incidents; they do not establish a universal production standard.
The August monitoring example leaves open what caused the alert email to fail and whether the logging and notification system has since been changed. More generally, the report does not say how the team handles disagreements between automated checks and human judgment, or what formal criteria govern approval beyond Meyer’s review.
The Next Test Is Reliable Oversight
Meyer says the frame-coverage and movement measurements now act as gates in the video build: a scene that fails specified checks is stopped before rendering, and an object that does not match the narration triggers a warning. The next practical test is whether those gates catch recurring defects consistently while still allowing useful creative variation. The account does not provide later results or a release schedule for further changes.
The legal experiment’s published task runs are available at lexassay.com, according to the November 9 update. Readers can consult those runs, although the source material does not set out a full independent replication or evaluation plan. For the monitoring issue, the report identifies the missed warning and email but does not say what follow-up work was completed. Further detail on alert delivery, review outcomes and benchmark methodology would clarify how the reported safeguards perform beyond these examples.
Key Questions
What happened in Meyer’s AI video project?
Meyer says he built a hand-drawn video engine, had its code reviewed, approved a short test video and then revised the first full video after finding it visually slow and sparse. He reports that a revised version was published a few hours after his critique.
What measurements did the team use?
For the first three scenes of the rebuilt video, Meyer reports that frame coverage increased from 15% to 66%, still time dropped from 29 seconds to zero, and camera moves rose from three to 27. He says those measures became checks in the production process.
Did the legal-model experiment prove the model improved?
No. Meyer says the model answered 10 published tasks rather than four after a weight-editing experiment, but its legal score remained zero. The task count alone does not show improved legal performance.
Who reviewed the video engine’s code?
Meyer says Claude built the engine and OpenAI’s Codex reviewed it. He describes Codex as a reviewer, not a builder, and says five findings were fixed with tests intended to fail against the earlier code.
Is this evidence that every AI project needs the same workflow?
No. The report documents one creator’s process and examples; it does not establish a universal standard. Its stated approach combines code review, tests, human approval and task-specific quality checks.
Source: ThorstenMeyerAI.com
