Technology
IF IT WORKS, OPENAI WILL FIX IT UNTIL IT DOESN’T
The corporate art of improving a product past the point of usefulness.
There is a peculiar disease in modern corporate culture.
It usually attacks shortly after somebody creates something that works.
The product is good. Customers like it. The system is stable. People have learned how to use it. Workflows have formed around it. Nobody is standing outside headquarters demanding an emergency redesign.
This creates an emergency.
Somewhere inside the company, somebody looks at this intolerable stability and asks:
**“So what are we changing?”**
The obvious answer — *nothing, you magnificent idiot, it works* — is apparently unacceptable.
Because if nothing changes, there is no transformation initiative.
No reorganization.
No refreshed user experience.
No strategic realignment.
No exciting new roadmap.
No presentation explaining why the thing customers already liked is now seventeen percent more customer-centric.
And, most dangerously, somebody in management might eventually be asked what exactly he does all day.
So the furniture must move.
OpenAI appears to have developed an especially aggressive case of this disease.
CONGRATULATIONS. YOU HAVE A WORKING PRODUCT.
ChatGPT has never been perfect.
It hallucinates. It misunderstands instructions. It occasionally develops the intellectual confidence of a drunk uncle explaining quantum mechanics after watching half a YouTube video.
Fine.
Users learned this.
That is what people do with tools.
A professional user does not need a tool to be perfect. He needs it to be **predictable**.
You learn what the machine does well.
You learn where it fails.
You learn how much context it can handle.
You learn how to structure prompts.
You learn when to verify it.
You learn its peculiar habits.
Eventually something important happens:
**The human and the tool develop a workflow.**
That workflow has value.
Perhaps more value than another benchmark point.
Because somebody writing software, editing a book, analyzing documents or maintaining a months-long project does not particularly care whether a new model scores three points higher on some benchmark whose name sounds like a malfunctioning dishwasher.
He cares whether the bastard remembers the file he uploaded ten minutes ago.
And this is where the modern technology company encounters its philosophical crisis.
The customer has adapted to the product.
The product works.
The customer would now like everybody responsible for the product to kindly remove their hands from it.
Impossible.
There are hands.
There is a product.
The hands begin to itch.
THE AUGUST SURPRISE
The latest complaints surrounding ChatGPT are interesting precisely because they are not complaints that artificial intelligence remains imperfect.
Users are describing **regressions**.
That is different.
In recent OpenAI community discussions, paying users have complained that after recent GPT-5.x changes, long-established workflows became unreliable.
Long conversations reportedly lose track of information that was previously available.
Attached files appear to vanish from the model’s effective context.
Users say the system sometimes claims a file is unavailable even though it was uploaded into the same conversation.
Others complain that supposedly intensive High or Extra High reasoning sometimes finishes suspiciously quickly and produces shallow answers.
Complex requests can turn into plans for completing the request rather than completion of the request itself.
Instead of doing the job, the machine sometimes explains how it intends to approach doing the job.
This is a magnificent technological achievement.
Human corporations have finally succeeded in teaching artificial intelligence the meeting.
Then there are the interruptions.
“Connection interrupted.”
Responses that stop halfway.
Sessions that fail.
Login problems.
Service outages.
On July 25, OpenAI itself reported elevated error rates affecting ChatGPT, the API and Codex.
On August 19, another disruption affected access for users.
Technical outages happen. Every large online service eventually falls over.
That alone is not particularly interesting.
What is interesting is the combination:
A product is constantly being modified while users simultaneously complain that the thing they depend on has become less predictable.
And predictability is precisely what professional users buy.
SMARTER ON PAPER, DUMBER AT WORK
This exposes one of Silicon Valley’s favorite intellectual mistakes.
**Capability is not usability.**
A model can become objectively better at mathematics while becoming subjectively worse for somebody writing a novel.
It can improve benchmark performance while becoming worse at maintaining context.
It can become safer while becoming irritatingly reluctant to perform harmless tasks.
It can become more concise while becoming less useful.
It can become more sophisticated internally while becoming less dependable externally.
None of these statements contradict each other.
But corporate product announcements generally behave as though they do.
The new version is smarter.
The new version is faster.
The new version is safer.
The new version reasons better.
Wonderful.
Does it still do what I was doing yesterday?
That question somehow arrives much later.
Users have also complained about answers becoming shorter, more hedged and less complete. Developers describe receiving outlines or partial implementations where they previously received working code. Others report excessive verification, planning or caution before execution.
And this produces perhaps the funniest possible outcome for an artificial-intelligence company.
You build increasingly powerful reasoning systems.
Then you wrap them in enough routing, optimization, safety logic, product decisions and behavioral tuning that the person using the thing begins wondering where the intelligence went.
The engine may have gained horsepower.
Unfortunately somebody redesigned the transmission.
BUT HOW WILL WE KNOW?
There is an imaginary conversation that appears to explain an astonishing amount of modern product development.
**“It works.”**
“Yes, but could it work better?”
**“Probably.”**
“Then we should change it.”
**“Why?”**
“To see whether it works better.”
**“And if it works worse?”**
“We'll gather feedback.”
This is called innovation.
There is another possibility.
Leave it the fuck alone.
This possibility receives surprisingly little venture funding.
The defense of perpetual rework is always superficially reasonable.
Technology improves.
Competitors improve.
Security requirements change.
Infrastructure changes.
AI models genuinely need refinement.
Nobody seriously proposes freezing ChatGPT forever in amber and maintaining GPT-whatever until the Sun consumes the Earth.
The stupidity begins when **change itself becomes evidence of progress**.
It isn't.
Sometimes changing something improves it.
Sometimes changing something breaks it.
Sometimes changing something produces three months of work whose final achievement is recreating approximately what existed before the project began.
Anyone who has worked inside a sufficiently large organization has witnessed this miracle.
THE REWORK ECONOMY
And OpenAI is hardly alone.
This is practically a Western corporate art form.
A website works.
Redesign it.
A menu makes sense.
Reorganize it.
Customers know where the buttons are.
Move them.
A brand is recognizable.
Refresh it.
A department functions.
Restructure it.
Employees finally understand the restructuring.
Restructure it again.
Then commission a survey asking why employee engagement has declined.
The problem is structural.
Maintenance is invisible.
Restraint is invisible.
Leaving something alone produces no launch event.
Nobody gets promoted for announcing:
**“This quarter we examined the product carefully and decided not to fuck with it.”**
There is no keynote presentation.
No new logo.
No transformation consultant.
No triumphant LinkedIn post.
Change, however, produces artifacts.
Roadmaps.
Meetings.
Teams.
Budgets.
Management positions.
Consultants.
Metrics.
Reports.
A person who preserves a functioning system can look inactive.
A person who tears it apart and rebuilds it can look indispensable.
Even when the rebuilt version works worse.
So organizations develop a perverse incentive:
**The people responsible for improvement must continuously prove that improvement is necessary.**
Otherwise somebody may eventually ask why the improvement department is so large.
AI MAKES THIS WORSE
With ordinary software, users can sometimes resist.
Do not install the update.
Keep the old interface.
Use the previous version.
Maintain the legacy system.
AI products increasingly remove that possibility.
The thing behind the interface can change without the user changing anything.
Routing changes.
System instructions change.
Safety behavior changes.
Context handling changes.
Models are replaced.
Tools are modified.
Limits move.
The product you used Monday can behave differently Friday while retaining exactly the same name.
That creates a strange new commercial relationship.
You are not really buying a stable tool.
You are buying **access to an evolving experiment**.
For casual users, perhaps that is fine.
Ask it for a lasagna recipe today and another lasagna recipe next month. Nobody dies because the second lasagna has slightly different reasoning characteristics.
Professional workflows are different.
If somebody has spent months learning how to collaborate with a model, behavioral stability becomes part of the product.
Breaking that stability has a cost.
The user pays it.
Not the team that changed the model.
THE SAFETY BUTTON
Then there is safety.
AI companies have legitimate reasons to prevent models from assisting with genuinely dangerous activities.
But safety systems have another characteristic familiar to anybody who has ever dealt with corporate compliance:
They expand.
A restriction designed for an obvious dangerous case acquires another restriction for an edge case.
Then another restriction to prevent circumvention.
Then another because somebody on social media discovered an embarrassing prompt.
Eventually the system can become so nervous that ordinary users start encountering refusals in situations having approximately nothing to do with the original danger.
The corporate response is usually that the model has become safer.
The user's response is usually:
**I asked you to make a fucking picture.**
Both statements can technically be true.
This is another reason product quality cannot be measured on one axis.
A model that refuses every request would have spectacularly low misuse rates.
It would also be a remarkably expensive blank page.
The difficult engineering problem is not maximizing safety.
It is maximizing usefulness **while maintaining appropriate safety**.
Those are not the same objective.
And when companies forget the distinction, users notice immediately.
THE CUSTOMER IS NOT YOUR BETA TEST DEPARTMENT
There is nothing inherently wrong with experimentation.
Experiment.
Test.
Break things.
Try bizarre architectures.
Change the routing.
Make the model excessively cautious.
Make another version excessively verbose.
Give one version the personality of Dostoevsky and another the attention span of a Labrador puppy.
That is what testing environments are for.
The problem begins when paying customers become involuntary participants in experiments that disrupt established work.
A mature product should have a simple principle:
**The more people depend on a behavior, the higher the burden of proof before changing it.**
That principle is boring.
It is also how grown-ups operate infrastructure.
You do not redesign an aircraft cockpit every Tuesday because the interface team discovered a more modern icon.
You do not move the brake pedal because analytics suggest drivers interact with the accelerator more frequently.
And you should be extremely careful about changing the behavioral characteristics of an AI system after millions of people have incorporated those characteristics into their work.
Innovation without stability is not innovation.
Eventually it is just churn.
PLEASE REMOVE YOUR HANDS FROM THE MACHINE
There is an old engineering principle:
**If it isn't broken, don't fix it.**
Modern corporate culture has produced an improved version:
**If it isn't broken, reorganize the team, redesign the interface, retrain the model, change the routing, rename three features, remove the fourth, call everything an improvement — and wait for the customers to explain what you broke.**
Then hold a meeting.
The meeting will determine what to change next.
OpenAI is building some of the most sophisticated technology humanity has ever produced.
That makes the problem funnier, not less serious.
Because creating intelligence is apparently easier than developing the organizational discipline required to occasionally say:
**It works.**
**Don't touch it.**
There will always be another model.
Another benchmark.
Another safety adjustment.
Another optimization.
Another manager with an idea.
Another engineer who found something fascinating at two in the morning.
Excellent.
Give them a sandbox.
Give them GPUs.
Give them benchmarks.
Give them all the coffee in San Francisco.
But when millions of people finally have a version that works reliably enough to become part of their daily lives, introduce one additional piece of advanced alignment technology.
A sign above the keyboard.
It should say:
# DON'T FUCK WITH IT.