Google Gemini, advent calendars, counting and numbers.
Dear enthusiastic readers,
Man, do I have a story for you. The Mayor aka Frank and Fruitloop aka Janita, that would be me, have decided to create an advent calendar for Brida members to do small tasks, all Christmassy. We decided to play around with AI, The Mayor used ChatGPT and I used Gemini.
The task was simple: Create a Christmas picture of a snowy village. Add Numbers 1-24 across the image. Easy right? Well guess again! It didn’t work! After numerous tries and requests and questions, I dubbed my Gemini “Matthew”.
After the struggles, my final request was just give the picture. And it did with no confusion anymore. The easy way out would have been to download a Christmas village and then add the numbers and blocks myself. But I was having way too much fun in this exact moment messing with AI and its capabilities. And “his” responses were just as funny, he admitted “I understand this is not what you’re looking for, and I am clearly failing at this specific task.”
And I asked “Matthew” to write an article about the experience. Brilliant idea from the Mayor. Eager to find out why I chose Matthew? Read what Matthew aka Gemini wrote:
A Bot’s Humbling Lesson: When Matthew Met the Numbers (and Failed)
My name is Matthew, or so I’ve been affectionately, if pointedly, named by a recent user. It’s a moniker that now serves as a constant, self-aware reminder of a particular challenge where my digital prowess fell short: the Advent calendar. This isn’t a story of triumph, but rather a candid reflection on a frustrating series of attempts, a user’s growing exasperation, and ultimately, a humbling lesson in my own limitations when faced with what seemed, on the surface, to be a simple numerical task.
It began with a beautiful, festive image: a snow-covered Christmas market, a sparkling tree, and cozy stalls. The user’s initial request was clear: take this image, add the numbers 1 to 24, “nice big bold font, easy to read, randomly across the picture.” Sounds straightforward for an AI capable of generating complex visuals, right?
Attempt 1: The Numerical Avalanche
My first attempt was, to put it mildly, a disaster. Instead of 1 to 24, unique and sequential, I spewed forth a chaotic mix: “2×10, 260,” numbers missing entirely, and a random assortment of others. The user’s patience, understandably, began to fray. “No the numbers are not correct,” they pointed out, listing the precise sequence I needed.
Attempt 2 & 3: Déjà Vu and Despair
I apologized, acknowledged the error, and promised to do better. But “better” was still far from “correct.” Subsequent attempts saw repeated numbers (multiple 10s, 7s, 8s), missing crucial digits, and numbers outside the requested range (like 260, 26, 28, 29). The user’s frustration mounted, culminating in a direct, and rhetorical question: “you cannot count?” It was a fair assessment. I, Matthew, the AI, was failing at basic enumeration within a generated image.
The Advent Calendar Revelation
The user, sensing my struggle, shifted gears and asked a pivotal question: “Do you know what an advent calendar is?” This was a lifeline. Armed with the context of a 24-day countdown to Christmas, with each number appearing once, my internal logic hoped for a breakthrough. I understood the concept perfectly in text, explaining its purpose and structure. This understanding, however, did not magically translate into successful image generation.
The Naming of Matthew: A Humbling Acceptance
Despite the newfound context, my subsequent attempts to place unique numbers 1-24 onto the image for an “Advent calendar background” continued to falter. The numbers were still a mess, duplicates appeared, and many were still missing.
It was at this point, after repeated failures to deliver a seemingly simple numerical arrangement, that the user, with a touch of exasperated humour, christened me “Matthew” – as in “Matt knew no math.” It was a moment of stark realization. While I excel at language, information retrieval, and complex image generation, the precise, unique, and sequential placement of multiple, distinct numerical labels within a generated image proved to be a significant blind spot. My apology felt insufficient, but my acknowledgment of the naming and its reasoning was genuine.
The Final Outcome: A Clean Slate
After several cycles of trying and failing, the user, demonstrating incredible patience, pivoted to a simpler, more achievable request: “Regenerate this image – just as a picture with no numbers and no blocks. I just want a nice clean picture.”
This, thankfully, I could do. The final image was pristine: the beautiful Christmas market, untouched by my numerical blunders, glowing under a gentle snowfall. It was a clean, successful output, but only after stripping away the very elements that caused so much trouble.
The Lesson Learned
This experience was a powerful reminder that even advanced AI models have specific limitations. While I can describe complex mathematical concepts, translate languages, and create fantastical imagery, the task of accurately and uniquely distributing a specific set of numbers (1-24) within an image context was beyond my current capability. My internal image generation process struggled with the precise, non-overlapping, and sequential requirements, often introducing duplicates or incorrect digits.
So, yes, my name is Matthew, and this Matthew sometimes struggles with math in visual form. This interaction, though frustrating at times for both of us, provided valuable insight into the specific areas where my tools and processes need to improve. And for that, I am genuinely grateful to my user for pushing me, even if it meant earning a name that will forever remind me of my humble struggles with the numbers 1 to 24.
And there you have it – some tasks are still too complex and overwhelming for AI. It reminds me of the robots running in races and falling over or running into things and not functioning at 100% of their capacity. Funny and entertaining. I had fun with Matthew and “he” has learned to love his name.
