These 1,274 cases span just 47 days — and publication essentially stopped on August 12

By Alpha Lay · 2026-09-26

Every record in our case library keeps the original publication link. During a data-hygiene pass we noticed something: those links contain X (Twitter) post IDs, and those IDs encode an exact publication timestamp.

Once decoded, the time shape of this dataset turned out to be far steeper than we expected.

Where the data comes from

X post IDs are snowflakes: the high bits encode a millisecond timestamp. The formula is (id >> 22) + 1288834974657.

Of 1,274 records, 980 (76.9%) come from an x.com/<user>/status/<id> URL and therefore have a resolvable time. The other 294 point at GitHub repos, tutorial pages and other addresses with no timestamp.

So everything below is about those 980 — not all 1,274. The direction of that bias matters: GitHub and tutorial sources generally lag social media, so the true time distribution is probably later and flatter than what follows.

Finding 1: Every case falls inside 47.2 days

A corpus of 1,274 cases covering seven weeks. That tells you the public stock is very thin — this is not "two years of accumulation," it is a concentrated burst inside one short window.

Finding 2: The first 14 days are 70%, then it falls off a cliff

A 5.6x difference in daily rate.

The daily series is more telling:

From 24 to 1 in one day. Over the following thirty-plus days, no single day exceeded 25.

Finding 3: There are two peaks, not one

Looking at totals alone, it is easy to conclude "launched late July, faded in early August." But the curve has two peaks:

• First peak, 07-30 (137) — immediately after July 29. A textbook launch-day effect.
• Second peak, 08-06 (85) — a week later, at about 60% of the first.

The second peak is the more interesting one. It means attention did not simply decay: there was a distinct second wave roughly seven days in. That interval matches the cycle of "get access → work out how to use it → package the result as a case and publish it."

Finding 4: Cases continued after August 12, but they are not the same thing

The 297 cases after the cliff are not mere scattered leftovers — their composition differs systematically from the first 489. We take that apart in a separate piece (2D animation rises from 9.2% to 24.4%; median prompt length rises 36%).

One number here: "first/last-frame-to-video" cases number 37 in the later half and 10 in the earlier half.

What you can do with this

If you are sizing up "how many cases should I study": the public stock is seven weeks deep. If you want earlier reference material, it does not exist — you did not fail to find it.

If you pick topics or build tools: the second peak (08-06) is a reusable signal. Roughly a week after a model launches, a second wave of cases appears, usually of higher quality than day one (day one is mostly "I got it running"). Watch day seven.

If you are judging whether this space is cooling: the mid-August trough cannot simply be read as decline. Our collection window closed on 2026-09-19, and natural decay in public posting would pull the late numbers down anyway. This is the largest uncertainty in this piece.

Method and limitations

• The sample is 980, not 1,274: only records sourced from X post links. GitHub and tutorial sources carry no timestamp and typically publish later, so the real distribution is probably flatter.
• Post time is not generation time: we decode the publication moment. A creator may have finished the video days earlier; we cannot measure that lag.
• Right-truncation from the collection window: the index was generated 2026-09-19 and the last case is 09-14. Do not read mid-September lows as a trend.
• "Decline" is an inference: this piece reports posting counts only and does not attribute a cause. The August 12 cliff could be platform-side (recommendation, topic decay) or collection-side.
• All figures use 980 of the 1,274 index records.

Reproduce it

Every record keeps its original publication link, and records can be grouped by source site: https://cases.aishifu.shop/