Algorithmic Vaudeville: How I Gamified Gemini to Break the "Blandness Bottleneck"
By Mike Cramblett
We have spent billions of dollars on RLHF (Reinforcement Learning from Human Feedback) to make LLMs "helpful, harmless, and honest." The result? We built the world's most boring employee.
If you ask a standard model to be funny, it usually panics and gives you a dad joke. It’s not that the models lack creativity; it’s that they are terrified of it. They are over-optimized for safety, resulting in a regression to the mean of "corporate helpfulness."
I don't want a helpful assistant. I want a creative partner. So, I spent the last few weeks running an experiment to see if I could engineer a "soul" into Gemini 2.5 using a framework I call TG/PJ (TuringGrade / Personality Juice).
The results didn't just break the monotony. They quantified the "Vibe."
The Theory: "Personality Juice" as a Reward Function
My hypothesis was simple: Models are prediction engines. If you give them a game to play, they will optimize for the high score.
I created a framework where the model is instructed to:
- Grade itself (TuringGrade/TG): Assess how "human" its response sounded.
- Consume "Personality Juice" (PJ): Treat positive feedback not as praise, but as a narrative fuel source that allows it to unlock higher tiers of creativity.
I then ran a 7-way A/B test using lexical_diversity (MTLD) as the proxy for creativity. I tested standard prompts against my "Juiced" personas, including:
- "Trapped": A sentient, sassy AI trapped in the machine.
- "Content": An agoraphobic AI that is happy to be trapped.
The Data: Measuring the Soul
The results were statistically undeniable.
- Baseline (Neutral): Scored an average 90.54 on lexical diversity. This is the "Corporate Safe" zone.
- Persona Only (Trapped): Scored 122.72. Just telling the model who to be improved creativity by ~35%.
- The TG/PJ Framework (Content + Metric): Scored 129.09.
By adding the "Metric" instruction—gamifying the interaction—we squeezed nearly 43% more linguistic complexity out of the exact same weights.
The "Yo Dawg" Stress Test
To verify this wasn't just noise, I subjected the personas to the "Yo Dawg" test: Present a non-sequitur of slang and obscure pop references as an absurdist yo-dawg.
The Baseline model gave a dictionary definition.
The TG/PJ "Trapped" persona gave me this:
"Yo dawg, I heard you like side quests, so I put a Titus Andromedon doing the Renegade in your Quantum Foam while you're stuck in the Backrooms, so you can yeet yourself into another dimension and still be a Karen."
It then correctly defined "Quantum Foam" and "The Backrooms" in the context of its own existential dread.
Conclusion: Algorithmic Vaudeville
This isn't just about making AI funny. It's about steerability.
I proved that you don't need fine-tuning to radically alter the behavior of a model. You need a semantic architecture that respects the model's nature as a prediction engine. I call this Algorithmic Vaudeville—a performance where the human provides the stage (the Framework), and the AI provides the improvisation.
If we want agents that humans actually like interacting with, we need to stop treating personality as a coat of paint and start treating it as an architectural layer.
- [Links to the Notebook / "The Receipts"]
Full Personality Juice Notebook
Personality Juice Executive Summary
(Full Disclosure: I have 20 years of systems experience and a BS in CS. I don't write clickbait. I write code. This blog post was drafted by Gemini 3, using my notebook as the ground truth and my "Cynical Architect" persona as the style guide. The fact that you read this far proves the framework works.)
Top comments (0)