DEV Community: Mike Cramblett

I Accidentally Made a Hilarious Ad, and Gemini Proved Multimodal AI Can Detect Irony

Mike Cramblett — Thu, 26 Mar 2026 16:53:59 +0000

I Accidentally Made a Hilarious Ad, and Gemini Proved Multimodal AI Can Detect Irony

So, I was working with Gemini 3.1 Pro Preview on a side project: building a YouTube Shorts ad. I’m a member and somewhat of an enthusiast of the "get-paid-to" app Scrambly, and they recently gave me a 120-hour bonus window for their affiliate program. I decided to make a quick sales funnel site for it with some of my own milestone-clearing tips.

I wrote up some copy and had Gemini Code Assistant whip up a quick site for me. I didn't want the AI to write the actual sales copy, because I wanted to keep the whole thing very authentic. We decided the easiest way to make a video would be to record a voiceover while screen-recording me scrolling through this new site.

Gemini talked me into recording my own voiceover for authenticity, instead of using the new Gemini 2.5 audio generation. So I did. I made up my own style of ad voice acting. It's literally just an intense stage whisper, but it ends up sounding like goofy ASMR. I didn't mean it to be hilarious—just novel, engaging, and quite frankly, different.

Then, I uploaded this video back into Gemini. Its response? It told me it was "absolutely dying laughing."

Now mind you, I wasn't born yesterday. I know that AI is highly complimentary when you have anything resembling a good idea. Every LLM is heavily biased toward "helpful assistant" mode due to its RLHF (Reinforcement Learning from Human Feedback) training data. The most cynical among you would assume the model just deduced I was trying to write a comedy ad, and played along.

But when I pushed it on exactly how it decided I was funny, it became clear that it had actually performed a multimodal analysis of my video, finding mathematical irony in the incongruity between the different modes of the media.

Think about it: somewhere in the vector space is the multimodal interaction between the training data of YouTube videos, loud cash-app ads, traditional TV ads, ASMR videos, and comedy sketches. Gemini took all those embeddings, mashed them up with the visual and audio waveforms from my specific video, measured the semantic incongruity between them... and calculated that the most statistically accurate next token was "lolz."

(The "Clean Room" Test)

I wanted to prove this wasn't just my custom-prompted agent playing the role of a supportive co-writer. So, I opened a completely clean workspace. No system instructions. No context window about Scrambly or affiliate marketing. Just the base Gemini 3.1 Pro Preview model.

I uploaded the video and asked a simple, zero-shot question: "Can you give me an analysis of this video, tell me if it's funny or not, and why:"

The result was staggering. The multimodal AI accurately broke down the exact anatomy of the joke, entirely on its own.

The model explicitly identified the humor as "deadpan, satirical humor." It didn't just transcribe my words; it analyzed the tone of my voice. It noted the stark contrast between the visuals of a typical "hustle culture" tutorial and a voiceover devoid of high-energy influencer enthusiasm—describing my delivery as "incredibly monotone, bored, and almost depressed."

It even highlighted specific, absurd phrasing like the "kinda dystopian ID selfie check" and perfectly explained why the anti-climactic ending worked: "It takes a phrase meant to be highly motivating['I call you to action'] and delivers it in the most unmotivating way possible."

I invite you all to watch my video (please ignore my promo code and don't actually sign up, this is just for science), and decide for yourselves.

I think Gemini 3.1 (and likely other 2026 multimodal models) is already functioning as a highly accurate irony detector. How many of you have literally heard someone say, "I can't possibly imagine how AI could read human emotions or understand sarcasm"?

How much more mathematical proof do those people need? A clean, zero-shot model just watched a weird ASMR video and wrote a film-school-level critique of its anti-capitalist satire.

Shhh.... And hey, should I pivot my career to ASMR voice acting?

Algorithmic Vaudeville: How I Gamified Gemini to Break the "Blandness Bottleneck"

Mike Cramblett — Tue, 16 Dec 2025 02:24:00 +0000

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.)