It started with a tiny question:
“What day does Tibo actually take off?”
Tibo, who works around Codex and ChatGPT at OpenAI, posts frequently about rollouts, resets, and new models. That is enough to create a little folk religion among users.
Tibo is posting → something may happen today. Tibo is quiet → is he off? → does that mean no reset either?
Before long, people are reverse-engineering a human work schedule before they even run the model.
From there, the conversation somehow tunnels through behavioral psychology, dopamine, AI productivity, compressed roadmaps, and finally infinite work.
This is the incident report.
1. “No Tibo, no reset” is still only a hypothesis
On September 5, 2026, after posting about Astra rollout and a banked reset, Tibo wrote about the team: “See you next week for some more ships.”[2][3]
Read at the end of a workweek, that sounds very much like “we are done for the week; see you next week.”
But that only supports a modest inference: the team appeared to be signing off for the weekend.
Public information does not establish:
- Tibo’s formal personal days off,
- whether every reset requires Tibo to press something manually,
- or whether reset processing stops whenever he is not posting.
The elegant theory is:
Tibo absent → reset operator absent → civilization collapses.
Unfortunately, the diagram is a little too elegant.
“See you next week” is evidence of a sign-off. It is not evidence that Tibo guards one enormous red RESET button by himself.
2. Yet people still keep checking: “Did it land?”
The more interesting problem is on the user side.
The reset time may be uncertain. Sometimes there is an announcement. Sometimes a banked reset appears. Once the allowance returns, users can immediately put the stronger model back to work.
So the checking loop begins.
Check X. Nothing. Check usage. Nothing. Check again hours later. Nothing. One more look. It is back. Festival.
If a reset arrived every day at exactly 9:00 a.m., one check at 9:01 would do the job.
Uncertainty instead installs a tiny supervisor in your head:
“Maybe now?”
With respect, we shall call this Tibo Roulette.
3. It is closer to variable-interval reinforcement than Pavlov’s dog
It is tempting to say, “Ah, Pavlov’s dog.” Not quite.
Classical conditioning is about one stimulus coming to predict another and eliciting a learned response. Here, the interesting question is different: why does the repeated checking behavior persist?
That points toward partial, or intermittent, reinforcement in operant conditioning. More specifically, a variable-interval schedule delivers reinforcement after unpredictable amounts of time.[6]
OpenStax’s psychology textbook gives an almost suspiciously perfect example of variable-interval reinforcement: checking social media.[6]
The resemblance is strong.
There is an important qualification, however. Checking X ten times does not make the reset ten times more likely. The reset happens externally; checking only discovers it.
So this is not a claim that OpenAI has placed users inside a laboratory conditioning protocol.
It is an analogy: from the user’s perspective, an irregular reward environment can make repeated checking look very much like variable-interval behavior.
4. Why a surprise reset feels unusually good: reward prediction error
One level deeper takes us into neuroscience.
A reward prediction error is, roughly, the difference between the reward expected and the reward actually received. Reviews by Wolfram Schultz describe many midbrain dopamine neurons as signaling positive prediction error when rewards are better than expected, staying near baseline for fully predicted rewards, and showing negative error when outcomes are worse than expected.[7]
So:
“Probably no reset today.”
↓
Surprise reset.
↓
Better than expected.
That kind of event is a salient learning signal.
But jumping from there to “you are addicted to dopamine” would be terrible neuroscience.
Reward prediction error is a basic learning mechanism. It does not diagnose people who check a reset meter with an addiction. Explaining every behavior as “dopamine hit” turns neuroscience into pub trivia.
The safer statement is simply: unexpected good news creates a larger prediction surprise than perfectly scheduled good news.
5. This is how we arrive at: “Give me Tibo’s unlimited PC”
Eventually, repeatedly checking a gauge becomes annoying enough that humanity invents a rational solution:
“Just let me use Tibo’s PC. Unlimited.”
The following terminal is entirely fictional:
C:\Users\Tibo> codex
Usage remaining: ∞%
banked reset: yes
rate limit: employee privilege
Astra: unrestricted
GPU: company-owned
Product name: Codex Pro Tibo Edition.
Simple policy:
Weekly limit: until Tibo gets mad.
Reality is less generous. OpenAI’s Help Center says Astra uses the Work/Codex allowance included with the relevant plan, and that Astra can consume that allowance faster than GPT-5.6 Sol depending on the task, input/output size, reasoning settings, and other factors.[5]
So stronger capability creates two truths at once:
“I want to use this more” and “the stronger model can melt the meter faster.”
No advanced psychology is required. Basic arithmetic is enough to make the fictional Tibo Edition attractive.
6. Then Astra brought “six months” into the conversation
This is where the story becomes absurd in a more serious way.
Tibo wrote that Astra had been a major competitive advantage before broad availability, and that the productivity jump was large enough for the team to move some plans roughly six months forward, shipping at DevDay rather than around the middle of the following year.[1]
That is not “autocomplete got a bit faster.”
OpenAI describes Astra as a model for difficult end-to-end work spanning coding, research, computer use, and professional workflows.[4][9]
In OpenAI’s reported evaluations, Astra completed OSWorld 2.0 latency-simulation tasks in about 47% less time than GPT-5.6 Sol while scoring higher, and Astra plus an updated Codex harness delivered 1.9× faster task completion on Mind2Web.[4]
The natural user response is therefore:
“The experiment is conclusive. Hand it over.”
At this point Astra sounds less like a model name and more like a roadmap demolition device.
7. But “six months earlier” does not mean “six months of human labor deleted”
This distinction matters.
Moving a roadmap six months earlier is not the same as replacing six months of human labor.
Product development contains more than implementation hours: research, prototypes, reviews, dependencies, waiting, rework, testing, and decisions all sit on the path.
A powerful AI system can potentially compress that path by helping teams:
- move from research to implementation in one workflow,
- compare more prototypes quickly,
- parallelize preparatory work that humans were waiting on,
- catch rework earlier,
- and start tomorrow’s validation today.
That can shorten the critical path, even if nobody literally removed six calendar months of typing.
Tibo’s post does not reveal exactly which stages shrank by how many days, so converting “six months” into person-months would be unjustified.
What his post does say is that work that lived in the future moved materially closer to the present.[1]
2027 Work: “I have not been called yet.”
Astra: “Come in today.”
2027 Work: “Excuse me?”
8. Higher productivity can produce more work instead of more rest
Now “infinite work” stops being only a joke.
In 2026, researchers at UC Berkeley Haas described ongoing research based on eight months inside a roughly 200-person U.S. technology company. As generative AI spread, employees worked faster, took on a broader scope of tasks, and extended work across more hours of the day.[8]
AI use was not mandatory there. The researchers argue that AI made more tasks feel personally feasible, so workers voluntarily expanded what they attempted.[8]
This is one company, not a universal law of all workplaces.
But the mechanism is easy to recognize:
Six-hour task → AI does it in 40 minutes
Human: “Great, I gained five hours and twenty minutes!”
Organization: “Excellent. Here are the next eight tasks.”
Human: “Was leisure deprecated?”
AI can shorten a task. It does not automatically generate the organizational right to turn the saved time into rest.
9. That is the real shape of “infinite work”
A productivity gain can be spent in at least three ways:
- produce the same result in less time and create slack,
- use the same time to improve quality,
- use the same time to increase volume.
Technology changes how much can be done. Humans and organizations decide where the new capacity goes.
In a growing organization, the pool of ideas, improvements, customer requests, and research questions may effectively be endless.
Double throughput does not guarantee that the backlog halves.
Instead, work that previously looked too distant or too expensive becomes visible as a new backlog.
That is the horror-comedy of infinite work.
Astra arrives
→ work finishes
→ work from six months ahead moves closer
→ that finishes too
→ even more distant work becomes visible
In a video game, the draw distance increased. At work, the quest markers now render all the way to the horizon.
10. Conclusion: maybe the thing that ends is not work, but the roadmap
We started with, “Does Tibo take weekends off?”
Then:
Tibo goes quiet
→ no reset?
→ repeated checking
→ variable-interval-like behavior
→ surprise creates prediction error
→ give me the unlimited PC
→ Astra is absurdly capable
→ plans move six months earlier
→ future work fills the saved time
→ infinite work
Two conclusions survive the comedy.
First, manually watching an irregular reset is mentally expensive. If an external event can be notified automatically, a human does not need to become a full-time gauge observer.
Second, AI productivity does not automatically become leisure. Unless people deliberately choose whether the dividend goes to rest, quality, speed, or volume, new capability simply invites new work.
Astra is not the villain. A powerful tool really can bring the future closer.
The dangerous sentence comes from the human standing next to it:
“Oh. We can do that today too.”
Then future work starts lining up at the door.
Work may never end. The roadmap might.
