What's new in AIOS Scribe
Latest
Each account can now start three new transcriptions a day instead of five, and only one video longer than 30 minutes in any rolling 24-hour window. The 120 minutes of new video a day still apply. A second long video within that window is refused, and shorter videos still work in the meantime. Long videos are where most of the cost goes, so this keeps the service free and available for everyone. Transcripts that already exist, and joining one already running, are still free and do not count.
For part of 15 November new transcriptions were refused: the service had used its monthly AI budget halfway through the month, because far more people are using it and long videos cost the most. The monthly budget is now twice as large, and each account can start up to 120 minutes of new video a day, on top of the five new transcriptions a day. A video that does not fit what is left today is refused with how long it is and how many minutes remain, so you can choose a shorter one now or come back later. Transcripts that already exist, and joining one already running, are still free and do not count.
A long video is now transcribed as many small jobs instead of one big one: each takes four one-minute parts and runs on its own, and the last to finish joins the transcript together. One job doing everything was being stopped partway by the platform, which lost the whole video. A part that is interrupted is retried once, and the video stops spending as soon as it can no longer succeed. Tested end to end on 15-minute, 58-minute and 104-minute videos, which finished in one to three minutes. The 15-minute limit is lifted.
From 6 to 18 October, 39 of the 51 transcription jobs for videos longer than about 15 minutes died partway and were later marked failed, while 36 of 37 jobs under 15 minutes completed. We could not see why, because the service kept no logs. We now refuse longer videos up front with a clear message, turned logging on, and found the cause: a platform limit on how many outside calls one job may make, which long videos exceed. The limit has been raised; the 15-minute cap stays until a long video has been tested end to end.
A stored transcript can no longer be silently replaced: each distinct output is its own version, one version is served to new requests, and every change is logged so it can be undone. A transcript that fails a serious automated check (broken or interleaved passages) is no longer offered to new requests; one that only raised warnings is still shown with a plain “flagged for review” note. Summaries are now tied to the exact transcript version they were written from. A held transcript can be regenerated by us under strict limits, and the new version replaces it only if its own review is better.
Transcribing now needs a free account, with five new transcriptions per account and per network address in 24 hours; cached transcripts and joining a running job are free. The summary became a tab next to the full text and timestamps, with a Markdown export. Transcripts can be translated with DeepL. You can opt in to one email when a report is ready or has failed. Automated quality evaluations of every transcript started being stored.
Some replies switched timestamp notation partway (1:07 arriving as 1.07), which made lines after the first minute show at the wrong time. These are now repaired, including for transcripts saved earlier. Videos are transcribed in overlapping one-minute parts, so a sentence that crosses a cut is no longer lost or cut in half.
AIOS Scribe moved to Cloudflare. A 49-minute video that had never completed was transcribed in 83 seconds by splitting it into parts. Third-party scripts were deferred until the page is usable, which cut blocking time on phones, and every guide and article got its own preview card.
The first guides were published, ads were removed from the transcription workflow, and copy that over-promised was rewritten. Database access was tightened so a browser cannot read stored transcripts directly.
A failed job now says why (refused content, too long, busy, timed out) instead of a generic message. Video length now comes from YouTube rather than the AI's estimate, which had been wrong for most videos. A monthly spending cap was added so the service cannot overspend. Asking the AI for timed segments only, not one long block of text, stopped it refusing several well-known talks.
Both now run in the application itself, with failures recorded, and privacy requests are stored first and emailed to us afterwards, so a slow email service never loses a request.