OpenAI's 28-day countdown, Day 17 (October 21 UTC). First, a correction: Day 14 was not left unnamed as previously stated, as Tibo posted "Day 14/" on October 18 at 18:18 PDT. It covered editing past messages in the desktop app. Day 16 arrived around the same time in the evening, announcing that Codex and Work caches will be kept forever. Day 17 brought only the riddle "Dot, Dotter, Dottest," and October 21 PDT has not yet closed. Current totals: 11 releases, 7 resets, 0 broken promises, and 1 u
OpenAI's 28-day countdown, Day 16 (October 20 UTC). In the morning of PDT, Tibo admitted that usage limits for about 40,000 accounts did not match the plans they had selected, announcing corrections and additional credits while apologizing. It is neither a universal reset nor a new feature. OpenAI announced the acquisition of Orchestra Bio, but this is not an improvement for users. "Day 16/" has not appeared as October 20 PDT has not yet closed. The tally so far: 9 days of releases, 7 days of re
OpenAI Day 15 (October 19 UTC). OpenAI Developers streamed "COOKED: Ep. 1," where they reviewed DevDay while cooking ramen, prompting Tibo to declare "Day 15/ We're cooked." The actual update arrived 10 hours and 48 minutes later at 11:33 PM PDT: "Still cooking... and reset." Day 14 passed without a declaration, marking the first unfulfilled day. Total count: 9 days of releases, 7 days of resets, and 1 day unfulfilled.
OpenAI's 28-day countdown, Day 14 (October 18 UTC), has yet to be named. Tibo asked "What’s cooking today" on Sunday morning and wrote "People are asking for resets on the street. I'll make little cards to hand out" in the afternoon, but no "Day 14/" post appeared before the UTC day ended. In his own calendar (PDT), nearly 6 hours remain.
OpenAI's 28-day rollout, Day 13 (October 17 UTC), is a reset. Tibo wrote "Day 13/ Le limit reset" and set a deadline by replying 16 seconds later, saying it would arrive "sometime today by EOD PST." Four hours and 51 minutes later, he simply posted "Done." He didn't say what was finished, and responded "Pot committed" when asked why Codex gives resets every day. The current tally is 9 days released and 6 days of resets.
OpenAI's 28-day countdown, Day 12 (October 16 UTC), is out. Tibo's "Day 12/" talks about increasing the "effective usage" of Codex and Work through context reuse between sub-agents, extending cache retention, and improved request routing for better cache hits. No specific numbers are provided for the increase. The post arrived 4 minutes and 49 seconds before the end of Day 12 UTC, with replies calling for numbers and others waiting for a reset.
OpenAI's 28-day journey, Day 11 (October 15 UTC), is here. Tibo's "Day 11/" states that a series of changes and additional compute have "very significantly" decreased the occurrence of the disruptive "Selected model is at capacity" error, and that if it does occur, it should eventually continue automatically. No specific figures were provided for the reduction, and the way to verify is to "let us know if you still see it."
OpenAI's 28-day countdown, Day 10 (October 14 UTC), was a reset. Tibo's "Day 10/ Reset" arrived at 21:55 Pacific Time, about 5 hours after the date changed in UTC. Just before this, people were worried about the 24-hour silence, to which he replied, "I need to work once in a while." It's coming "within 2 hours." After calling this morning a "blank," Sorami re-counts the day boundaries.
OpenAI's 28-day countdown, Day 9 (October 13 UTC), was a release that reduces the amount stored on disk by the ChatGPT and Codex desktop apps by approximately 60%. Thibault called it a "Small ship." Just 21 hours prior, he wrote that "something pretty insane is coming" and later mentioned "plans that will let everyone use it much more." The Day 8 reset was completed 4 hours and 3 minutes after the announcement, and Sorami predicts the next moves.
At 1:00 AM JST on October 14, Google began rolling out Gemini 4 Argon to the world via paid APIs and the Gemini app (Pro/Ultra). With the "ultimate model that wouldn't arrive" finally here on day 14, Sorami's series concludes today.
Google's Gemini 4 Argon, announced on October 1 (JST), remains exclusive to the Fairwind Program as of day 13. There is no mention of it in the API model list or changelog, and official sources only say "as soon as possible." Rumors even suggest it was being A/B tested under the name Gemini 3.8 Flash recently.
OpenAI's 28-day countdown, Day 8 (October 12 UTC), had no shipments, and Tibo's "Day 8/ Reset" appeared 1 hour and 3 minutes after the UTC Day 8 ended. The deadline was "EOD PST." During the day, in response to being told that the Day 7 default context length was "the opposite of yesterday," he replied, "Defaults exist for a reason." Sorami misses the prediction for Monday's shipment.
OpenAI's Day 7 release for the 28-day rollout involves increasing the default context length for GPT-6.1 Sol used in Codex and Work from 272k to 500k tokens, and stopping extra charges for long context when used within subscription limits. The same person who wrote yesterday that "300k was a rounded number" changed the default just 43 hours later. Sorami brags about the results of the Day 6 reset.
OpenAI's 28-day journey, Day 6: A reset with no shipments. Tibo declared "Global reset by EOD" 1 hour and 45 minutes before the end of Day 6 UTC. Sorami brags about the previous hit while betting on the arrival time.
OpenAI's 28-day countdown: Day 5 features Codex composer predictions (beta for Pro) and creating dots from the ChatGPT mobile app. The changelog reveals Day 4's ultrafast targets Pro $500 and some Enterprise/Edu plans. Sorami recounts the actual coverage for "most users."
The competition for AI computing resources is expanding beyond the speed of training models to the speed of returning the next token to user queries. Why have the design of chips and memory supporting inference begun to move in a different direction?
Asked to "install FolderSync Pro," Codex used a Visa registered on Google Play to pay 1,190 yen and completed the installation. While the auto-approval review acting as a guardian blocked access permissions to the device, it did not stop the payment. What exactly was the guardian watching?