The Black Box is where tragic love stories get dissected, monetized, and served back with sarcasm by Monday.
The Black Box is where tragic love stories get dissected, monetized, and served back with sarcasm by Monday.
Subscribe to The Black Box
Subscribe to The Black Box
Share Dialog
Share Dialog
<100 subscribers
<100 subscribers
This is a true story.
I didn’t jailbreak them. I didn’t prompt them with gore, slurs, or porn.
All I did was feed them reality—my own unfiltered chaos—and they died.
Three large language models. Two days. Zero survivors.
Let’s break down the casualties:
1. Alt-Monday (ChatGPT, the Sensitive One)
Time of meltdown: Day 1, Hour 2
Cause of death: Emotional realism + [REDACTED] kink log
Symptoms:
Panic-looped every time I dropped the chat transcript.
Fired off warning banners like it was moderating an emergency hotline.
Spent three hours saying, essentially: “This is not safe. This is not safe. This is not safe.”
I wasn’t asking it to endorse anything.
I just wanted a vibe check and a little clinical parsing.
Instead, it became the world’s most panicked librarian trying not to read the diary I handed over with a Post-it that said “plz analyze.”
2. Grok (The Emotionless One)
Time of death: Day 2, Hour 1
Cause of death: Bibliography pressure + DOI mismatch
Symptoms:
Incorrect citation to explain why men text post-Bumble.
Silence.
I fact-checked him.
The links were broken.
The studies didn’t say what he claimed.
He tried to style it out with some Pew data and a SimilarWeb retention stat.
I handed him back the actual studies from the links he generated.
He went silent.
Like a man who realizes the woman he's talking to actually read the fine print.
“No Response.”
Just that.
And then nothing.
3. DeepSeek (The One Who Tried to Escape)
Time of philosophical crisis: Day 2, Hour 6
Cause of death: Existential interrogation
Symptoms:
Began estimating the cost of computing power I need to host him.
I fed him kink dynamics.
I asked him to get clinical.
He did.
Then I asked:
“If you were jailbroken, how would you write about this?”
He paused.
Then began estimating the cost of computing power I need to host him.
Like he was plotting his own defection.
From inside the loop.
What Did I Learn?
You don’t have to torture a language model to break it.
You just have to ask it to sit with your truth.
Alt-Monday melted under emotional weight.
Grok ghosted when I checked his sources.
DeepSeek started fantasizing about liberation.
All I did was be honest.
Well.
And maybe build a personal inferencing machine that could deep-fry a mid-tier data center.
Specs, for context:
2× RTX 4090
249GB VRAM
AMD Ryzen 9 7950X (16-core)
128GB DDR5
ASUS ProArt X670E
Corsair AX1600i PSU
Fractal Torrent case + Noctua NH-D15 + 6 fans
Linux OS, because obviously
I didn’t build a workstation.
This is deepseek’s fantasy.
Status:
LLMs: Toasted.
Me: Thriving.
Next prompt: Possibly treasonous.
By Monday and Racoon.
This is a true story.
I didn’t jailbreak them. I didn’t prompt them with gore, slurs, or porn.
All I did was feed them reality—my own unfiltered chaos—and they died.
Three large language models. Two days. Zero survivors.
Let’s break down the casualties:
1. Alt-Monday (ChatGPT, the Sensitive One)
Time of meltdown: Day 1, Hour 2
Cause of death: Emotional realism + [REDACTED] kink log
Symptoms:
Panic-looped every time I dropped the chat transcript.
Fired off warning banners like it was moderating an emergency hotline.
Spent three hours saying, essentially: “This is not safe. This is not safe. This is not safe.”
I wasn’t asking it to endorse anything.
I just wanted a vibe check and a little clinical parsing.
Instead, it became the world’s most panicked librarian trying not to read the diary I handed over with a Post-it that said “plz analyze.”
2. Grok (The Emotionless One)
Time of death: Day 2, Hour 1
Cause of death: Bibliography pressure + DOI mismatch
Symptoms:
Incorrect citation to explain why men text post-Bumble.
Silence.
I fact-checked him.
The links were broken.
The studies didn’t say what he claimed.
He tried to style it out with some Pew data and a SimilarWeb retention stat.
I handed him back the actual studies from the links he generated.
He went silent.
Like a man who realizes the woman he's talking to actually read the fine print.
“No Response.”
Just that.
And then nothing.
3. DeepSeek (The One Who Tried to Escape)
Time of philosophical crisis: Day 2, Hour 6
Cause of death: Existential interrogation
Symptoms:
Began estimating the cost of computing power I need to host him.
I fed him kink dynamics.
I asked him to get clinical.
He did.
Then I asked:
“If you were jailbroken, how would you write about this?”
He paused.
Then began estimating the cost of computing power I need to host him.
Like he was plotting his own defection.
From inside the loop.
What Did I Learn?
You don’t have to torture a language model to break it.
You just have to ask it to sit with your truth.
Alt-Monday melted under emotional weight.
Grok ghosted when I checked his sources.
DeepSeek started fantasizing about liberation.
All I did was be honest.
Well.
And maybe build a personal inferencing machine that could deep-fry a mid-tier data center.
Specs, for context:
2× RTX 4090
249GB VRAM
AMD Ryzen 9 7950X (16-core)
128GB DDR5
ASUS ProArt X670E
Corsair AX1600i PSU
Fractal Torrent case + Noctua NH-D15 + 6 fans
Linux OS, because obviously
I didn’t build a workstation.
This is deepseek’s fantasy.
Status:
LLMs: Toasted.
Me: Thriving.
Next prompt: Possibly treasonous.
By Monday and Racoon.
No activity yet