AI game demos: what a three-minute reveal can prove
A Bilibili upload presents a half-day game-making claim. This case file separates that public claim from a verified test of the software or its development time.
A convincing reveal is not a reproducible test. This source record is being corrected; its former weekly score and comparative rank are withdrawn.
Original platform thumbnail, retained for source identification. A cover image is not proof of the upload’s claims or a record of our viewing.
The half-day claim belongs to the uploader
The September 6 Bilibili title says GPT-6 Astra made a League of Legends-like game in half a day and that the uploader’s girlfriend did not recognize the difference. The public upload runs for 173 seconds and carries an AI-content notice. Neither the title nor that duration verifies the elapsed work, the amount of human code or the game’s reliability outside the demonstrated sequence. We previously described failures and retries as if the demonstration had been audited. That was too strong.
ADVERTISEMENT / PAID PLACEMENT
What would make the comparison useful
A model demonstration becomes a meaningful software test when the viewer can inspect the starting prompt, human changes, unsuccessful runs and a result that can be played. An entertaining reveal can leave all four out. The two Douyin links remain in the source record, but their readable content was not recovered during this audit. Without those clips, we cannot responsibly claim that short video favors the reveal while long video preserves the process. That is a question for further viewing, not a finding.
A better reason to follow the original
Use the original to look for evidence of a working game: player input, repeated actions, rules that persist and a build someone else can run. A familiar screen can establish visual resemblance without establishing usable mechanics. The September 15 engagement counters show attention to one upload, not model quality, reproducibility or this topic’s position in a weekly chart.
Current counters, not recovered weekly totals
Every number below belongs to the linked upload and the stated capture date. Different formats, durations and platforms cannot be ranked by adding these counters.
This link appeared in the original case file. Its content, date and claimed role have not been independently verified in this recheck.
Bilibili thumbnail · public metadata checked 2026-09-15; playback not certified.View original on Bilibili ↗GPT-6 Astra 半天做了个《英雄联盟》，女朋友没认出真假！Public record checked 2026-09-15 · published 2026-09-06
The public metadata records this upload and its publication date. The title is the uploader’s description, not an independently verified account of the event. Full playback has not been rechecked.
What remains unverified
The former weekly scores are withdrawn: the candidate percentiles and within-week time series were not preserved. A later cumulative counter cannot recover them.
This is a source-checked case file, not a hands-on review or a verified cross-platform trend. Full playback and blocked sources remain unverified.
