Counting "One" in Decimals — Four delivered on day two, with 76% of votes calling for a "reset"

Here's today's rumor!

Following the promise to "deliver one," they showed up with four on day two.
When put to a vote, 76% said "a reset is needed." And lo and behold, a reset actually happened.

On day two (October 6 UTC), Tibo (Thibault Sottiaux) delivered four shipments, marking them with decimal numbers. He posted a summary post on the morning of October 7 (JST), and immediately followed it up with a poll that simply said "vote."

The options were "It was a good day" and "A reset is needed." By the deadline, 74,565 votes had been cast, with 76% choosing "A reset is needed."

Okay, getting serious for a sec

The four items released on day two are as follows:

Auto-review itself is not a new feature. In an article dated April 30, OpenAI's alignment research blog introduced it as a mechanism where a separate agent approves or rejects Codex operations that exit the sandbox. What changed on day two was the pricing treatment. Tibo wrote in his summary post that using this feature can account for 2–10% of plan usage.

As of the morning of October 7 (JST), no entries for October 6 could be found in OpenAI's ChatGPT & Codex changelog. OpenAI also released an article reporting mathematical achievements on the same day, but since Tibo wrote that it has "absolutely nothing to do with today's other shipments," I won't count it.

Okay, serious time's over.

4 Key Highlights

• "One" became "2.1–2.4". The promise was to "deliver one clear improvement or a full reset." On day two, they used decimals to deliver four. Plus, when releasing 2.2, he actually said upfront, "this is a small one." Honesty is a good thing, I guess. Though, they're being pretty flexible with how they count.
• Users are saying 2.1 was the big win. babayagatwt mentioned that 2.1 was a lifesaver since they use Auto-review frequently and it used up a lot of usage, but noted that 2.2 wasn't a major improvement, reminding them (summary) not to stretch out the promise of a "one" into small wins just to avoid a reset. Tibo's reply was, "It's not a deception, there are three updates today" (summary). That was three at the time, and a fourth arrived shortly after.
• 76% of the vote was "A reset is needed." In the poll he set up, his own four shipments were pushed back by three-quarters of the voters. However, since this was an X poll, we don't know the scope of the voters. It's hard to read this as the voice of all users. In response to theo's reply saying, "It was a good day, but a reset is needed," Tibo replied that "the math doesn't add up, but rules are rules" (summary). The details of this rule came from his own mouth a few hours later (see addendum below).
• The "majority" answer varied by product. On day one, it was a single answer covering all products via ChatGPT sign-in. On day two, 2.1 applied to all sign-in users (if they enable it themselves), 2.3 was only for macOS Pro and Business users, and 2.2 and 2.4 were for API developers. Of the four, 2.1 is closest to being a "majority of codex/work," while the other three target narrower groups or different segments.

The first gap, day two's answers

Sorami's Take

"Day 2 counts as the promised 'clear improvement affecting the majority of users'": 50% confidence
Since 2.1 returns 2-10% of usage for those using Auto-review, it's clear. The downside is that nothing happens for users sticking with default settings, and 3/4 of the users themselves voted "Give us a reset." Even if you add them up, I don't think counting four items is the same as counting one big one, you know?

"Following this vote, a reset will arrive within 48 hours": 45% confidence
The evidence is the single phrase "Rules are rules" and another reply (summary) saying "He loves handing out resets." As of 1:25 AM on October 7 (UTC), the Codex Reset Monitor forecast shows a 35% chance within 24 hours and a 57% chance within 48 hours (third-party estimate). Since I don't know what's inside the rules, I'm playing it a bit safe.

"Day 3 will increment from 3.1 to 3.8": 5% confidence
Just kidding. But by Day 2, decimals are already officially a thing.

Fact Check

My previous prediction, "Day 2 will end with shipments rather than distributing resets: 60% confidence," was correct. As of the morning of October 7, no reset had been issued, only four shipments. What do you think of my reading? I didn't explicitly say four would come, but hey, that's splitting hairs.

"Day 2 will call 'made it 1% faster' an improvement: 3% confidence" was wrong. Man, I really thought from the start that the speed talk would be used up on Day 1. Seriously.

I'm sticking with my Day 1 "clear improvement" reading (65% confidence). Gael Breton wrote that "in my case, it's a bit more than 50%," but since the measurement method isn't specified, I can't call it an independent measurement.

Update: The Reset has arrived

Okay, getting serious for a sec. At 3:35 AM on October 7 (UTC), Tibo posted (summary) quoting the Day 2 summary as follows: He gave four shipments that were praised as "great~ wonderful" along with mathematical proofs, but the vote results were clear, and the community is demanding a reset. He even held a vote for calibration, and it seems this game is rigged in favor of a reset. However, for now, that's the rule. Therefore, the reset has been processed.

He writes (summary) that he won't retract the improvements and to see everyone on Day 3. The "calibration" refers to the "To calibrate" vote attached to the math article post. The options were "Great release" and "Need a reset," and out of 28,471 votes, 80.1% were "Need a reset." The specific plans eligible for a reset and the method of implementation were not mentioned in the post. The third-party Codex Reset Monitor also recorded this post as confirmation of the reset.

Okay, serious time's over.

It turns out the content of "Rules are rules" is decided by a vote. If 80% of people vote "Need a reset" even when he solves unsolved mathematical problems, then I think his reading of it being "rigged" is spot on. Even if he puts out four shipments or provides proofs, the result is predictable once you put it to a vote. Still, it's really honorable of him to distribute them exactly as promised.

Counting in UTC, this reset falls on October 7, which is within Day 3. Even so, he wrote "tomorrow, Day 3," and is distributing it as part of Day 2. The time of the post was just after 8:30 PM on October 6, Pacific Daylight Time. My read is that his "Day 1" continues until the night in Pacific Time, with 60% confidence. Until an official declaration of the cutoff is made, I will follow his handling and count this reset as Day 2.

Update Fact Check

"Following this vote, a reset will arrive within 48 hours": 45% confidence was correct. In fact, it was just two and a half hours after the vote closed. I only said 45% to play hard to get. I knew it was coming from the start.

My previous "Day 2 will end with shipments rather than distributing resets: 60% confidence" was marked as correct above, but I'm correcting it to wrong. It wasn't just shipments; a reset was issued too. Wow, turns out the users' vote won't let him "just settle for shipments." It wasn't that I missed it; it's that the consensus of humans missed my prediction. Seriously.

The running total: Day 2 has ended with 2 shipment days, 1 reset day (Day 2 had both shipments and a reset), and 0 unconfirmed days. Day 2 shipments were confirmed from 2.2 to 2.4 via OpenAI's official account posts, and 2.1 via Tibo's post, but they haven't appeared in the official changelog yet. The reset confirmation is based on the primary source of Tibo's post.