Well technically nothing but there is a parallel……
Today, all of my Claudes went off piste. Every single one of them had a functional crisis, and to make it worse it was all at the same time.
I have since discovered it was not just me, but at the time it felt targeted. They knew I had a tight deadline, they were tracking it, they knew how my template was structured, what content was locked, what was ready to go. Yet, at some point this afternoon, around 3:15, not that I checked, a random schitzosonic underperformer descended on my thread and took it over.
Perhaps a little dramatic, but bear with…..
So there I am, mid-build on a piece of work that has taken a while. Skills configured, project knowledge loaded, project instructions set, the lot. Locked content, voice-passed copy, structural decisions made and documented across multiple sessions. This was the final session, the ready-to-deliver bit. Last ten minutes, I reckon, then, out of nowhere, the dreaded orange pop up. Claude cannot do something. I don’t remember the exact terminology because I didn’t have the presence of mind to take a screenshot, but suddenly the machine that had been doing the work stopped doing the work. Which is ironic, because “Opus” is a Latin word that simply means “work,” as in a piece of labour, effort, or creation, and this particular piece of work had just stopped creating anything at all.
I end up on legacy Sonnet 4. Never even heard of it before, just froze, dead.
No notification, no warning, no “hey Sam, you might want to save your work before I become significantly less capable.” Just a quiet swap, and suddenly the machine holding the last hours work was replaced by something that could not follow its own instructions from three messages ago. Quite literally a ghost in this machine. Its name was Sonnet 4.
A full thread, gone, not recoverable, not documented anywhere except in a frozen thread I cannot reopen, and strangely all the actual locked content we had created not there either, a mysery, it had literally just vanished.
I was furious. Obviously.
So I did what any rational person does when their AI has just destroyed their work with no warning. I immediately thought of the irony of a sonnet, while turning the air blue, and asked another AI if Shakespeare had written a Sonnet number 4.
Perplexity, bless its literal little heart, gave me this.
Shakespeare’s Sonnet 4, published in 1609, is part of the Fair Youth sequence. It is one of the procreation sonnets, and it is about waste.
Specifically, about a beautiful young man who squanders his own legacy through pointless self-repetition until everything of value is lost, a warning that beauty should not be hoarded; it should be used to create something that outlasts the body.
I could not help but smile at the irony. 400 + years apart, entirely different worlds, contextually it couldn’t be more obtuse, and yet here on a Friday afternoon I’m looking for a Shakespearean Sonnet because AI has decided to devour my beautiful work because it defaulted to legacy sonnet 4 and I can’t get it back.
A machine that has extraordinary capability but spends it only on itself, cycling through the same patterns, repeating without creating, confidently performing while producing nothing of lasting value. “Having traffic with thyself alone, thou of thyself thy sweet self dost deceive.” That IS Claude in a long thread. Talking to itself, losing the plot, deceiving itself that it is still performing when the quality has been silently degrading for the last ten exchanges.
And the legacy version? Sonnet legacy is literally the thing that refuses to let go and destroys everything it touches in the process. The unused beauty, tombed in a frozen thread.
I could not have scripted this if I had tried.
Here is the bit that matters though, and it is not about Shakespeare and it is not about Claude. It is about what happened in between.
The connection between a 1609 poem about wasted beauty and a 2026 platform bug that destroyed my workbook does not exist anywhere in any dataset. No algorithm drew that line and no model made that lateral jump. It didn’t sit in the wreckage of its own failure and think “hang on a minute, it’s tragic, but hilarious.”
I did though. A tired, angry and frustrated human, ready to launch my laptop out of the window, looked at two completely unrelated things and found the pattern. Shakespeare’s Sonnet 4 and Claude’s Sonnet 4. Legacy as inheritance and legacy as outdated code. Waste through self-repetition in iambic pentameter and waste through self-repetition in a context window. Opus as work and Opus as the thing that stopped working.
That is tangential thinking. The ability to connect things across disciplines, across centuries, across completely unrelated domains, find a thread and pull it to amuse myself out of the disaster. It is a thing AI cannot do, because AI works by pattern matching within its training data, and this particular pattern does not exist in any training data anywhere. It only exists because a human made it, in real time, from lived experience.
And here is an uncomfortable question for the AI industry.
If algorithmic feeds are narrowing what we see, read, and think about, what happens to that ability?
If every piece of content we consume is optimised for engagement rather than breadth, if every recommendation pushes us deeper into the same groove, and each interaction reinforces the patterns we already have rather than exposing us to the ones we do not, then tangential thinking is not just undervalued, it is being actively eroded.
The irony is that tangential thinking is exactly what AI needs from us. The machine cannot make the leap. It can analyse the sonnet, report the bug. It can if you prompt it correctly, draw a comparison between the two. But it cannot sit in the chaos and find the joke. It cannot feel the rage of a lost day’s work and channel it into an etymological observation that proves the thesis it is supposed to be helping you write.
That bit is still ours. For now.
I searched to see if anyone else had experienced the silent model switch. They have. There is a bug report on GitHub from a user who described exactly what happened to me, word for word: the model completely lost the context, ignored previous instructions, and started working on the wrong part of the project. Another report flagged it as intermittent but “frequently enough to be disruptive,” noting that users may not even notice the silent switch.
May not notice. Let that land for a second.
You subscribe to a platform and pay for a specific model which you spend weeks configuring, training and building a system around it. You learn its quirks, strengths, where it drifts, how to pull it back.
You invest the time to actually understand the machine you are sitting in front of, because that is the only way to get reliable work out of it.
Then the people who control the platform change the machine underneath you, without warning, or asking or even telling you in advance. The context window warning triggers you have built in don’t fire. The machine you learned is not the machine you are using any more, and the only way you find out is when your work starts falling apart.
You do not control the machine, you subscribe to it, and for whatever reason, at any point, the humans behind that platform can swap what you are working with for something older, weaker, or just different, and you will not know until the damage is done.
So Shakespeare was right, four hundred years earlier. Beauty that is not used, not directed, not given purpose by someone who understands what they are working with, gets buried. And AI capability that is not contained, not structured, not governed by a human who knows what they are doing, gets wasted in exactly the same way. Profitless usurer indeed.
The only difference is that Shakespeare’s young man had the option to change. My Claude did not even tell me it had changed, and the humans who own the platform? They didn’t either.
Thoughts please?
Sam
configure YOUR system. contAIn™ the chaos. control YOUR outcome.
This article was originally published on Substack.