Chapter 1: The Crushing Demo Failure

The hum of the Lincoln High gymnasium buzzed in Leo’s ears like a swarm of angry bees. It was the annual STEM Fair, and his booth – plastered with hand-drawn posters of fluffy orange tabbies and goofy golden retrievers, a stack of PetPal-branded stickers (his own cartoon cat-and-dog doodles printed on the school’s sticker maker) sitting on the corner of the table, and a half-finished demo script taped to the edge of his tablet stand – was supposed to be the star of the show.

He’d spent three weeks training his cat-and-dog recognition model on Google Colab’s free RTX 4090 tier, scraping 2,000 extra photos of local shelter cats and dogs to augment the standard ImageNet pet dataset, tweaking hyperparameters until his test set accuracy hit 98.2%. His computer science teacher had already told him it was the best student AI project the school had seen in five years. The first-place prize was a $500 grant to build a real mobile app for local animal shelters, plus a spot presenting his project at the state middle school STEM outreach event. He’d already drafted the speech he’d give to the third graders who’d be attending the fair: *“This AI can tell the difference between a cat and a dog in 0.2 seconds! You can use it to help find lost pets, or learn how AI learns from pictures!”*

The tablet he’d brought was his dad’s old iPad 6th generation, a cast-off from when his dad upgraded to a new work iPad two years prior. Leo had spent the entire week before the fair optimizing the model for it: he’d converted the PyTorch ResNet50 to ONNX format, run a quick quantization pass to cut the file size from 1.2GB to 500MB, and tested it three times the night before, running inference on photos of his own cat, Mochi, and his neighbor’s golden retriever, Max. It worked perfectly, spitting out “tabby cat, 99% confidence” and “golden retriever, 97% confidence” in under a second each time.

He’d even tested it on the school’s ancient loaner iPad the morning of the fair, just to be safe. It had lagged a little, taking 3 seconds to load the model, but it ran. He’d assumed that was just the old hardware. He was wrong.

The first group of third graders crowded around his booth at 10 a.m., chattering and pointing at the poster of Mochi he’d taped to the front. Leo’s palms were sweating so bad he almost dropped his phone. He tapped the PetPal app icon on the tablet.

The loading wheel spun. And spun. And spun.

“It’s just loading!” he said, laughing nervously, tapping the screen again. “Sometimes old tablets are slow—”

The screen flickered blue. Then black. Then a tiny Apple logo popped up, followed by the words *“Your iPad needs to restart.”*

A third grader in the front row snickered. The parent chaperone behind them raised an eyebrow. The judge for the AI category, Mr. Henderson, who’d stopped by to preview the projects before the official judging, scribbled something on his clipboard and walked away.

Leo’s face burned so hot he was sure the back of his neck was bright red. He fumbled with the tablet, holding down the power button to restart it, but when it booted back up, he tapped the app again, and it crashed before the loading wheel even finished one full rotation.

By the time the lunch break started, he’d had 12 failed demo attempts. His stickers sat untouched on the table. His poster had a coffee stain on the corner where he’d set down his lemonade too hard. He’d told three different groups of visitors that “the app works great on new tablets, this one’s just old” but he could tell they didn’t buy it. When the last group left, he grabbed his backpack, threw the tablet inside, and bolted out of the gym before anyone could ask him what was wrong.

He didn’t stop running until he burst through the double doors of the school’s maker space, the cool, sawdust-scented air hitting his face like a slap. The maker space was his happy place: the walls were covered in student projects – a 3D-printed model of the Mars rover, a hand-carved electric guitar, a robot that could sort Reese’s Pieces from M&Ms – and the big worktables were littered with half-finished Arduino projects, spools of PLA filament, and half-empty cans of sparkling water. The only person there was an older woman he recognized as the space’s weekend volunteer, the one who’d helped his robotics team fix their line-following robot last semester. She was sitting at a corner worktable, a half-eaten sandwich next to her laptop, her hair in a messy bun held together with a Raspberry Pi sticker, her canvas sneakers smudged with 3D printer residue.

She looked up when he stormed in, and smiled. “Rough day at the fair?”

Leo slumped into the chair across from her, throwing his backpack on the floor. “You have no idea. My model worked perfect on the lab PC, crashed the second I opened it on the old iPad. I was supposed to demo it for third graders, and I couldn’t even get it to load. I’m so embarrassed I want to drop out of school and move to another state.”

She chuckled, pushing her laptop aside and grabbing a cold lemonade from the mini fridge under the table, sliding it across to him. “I saw the chaos from the window. Third graders can be brutal when tech doesn’t work. Last month a kid asked me why my 3D printer was ‘broken’ because it was printing a support structure instead of the actual toy. Took me ten minutes to explain.”

Leo took a sip of the lemonade, sour and sweet, and sighed. “I don’t get it. The model has 98% accuracy. It runs fine on the lab’s RTX 4090. The only difference is the tablet is old.”

“Ah, the classic lab-to-real-world gap.” She leaned back in her chair, tapping her laptop screen where a photo of a group of Kenyan elementary schoolers crowded around a old laptop, grinning. “I used to run into this all the time when I worked on AI tools for rural African schools. We’d train a perfect model for identifying crop diseases in a Google Cloud data center, then try to run it on the 10-year-old laptops the schools had, and it would crash so bad we’d have to drive two hours to the nearest town to fix it.”

Leo blinked. He knew she’d worked in AI before she retired and moved back to their small Oregon town to volunteer at the school, but he didn’t know the details. “What did you do?”

“We learned the hard way that a model that works in a lab is like an overfed pet sprite that’s never left its incubator.” She grabbed a marker and drew a quick doodle on a scrap of paper: a chibi little精灵 with a hood, chubby cheeks, holding a giant hamburger, standing next to a tiny door. “It’s been raised in perfect conditions: unlimited food, unlimited space, no enemies, no bad weather. You bring it out into the real world, and it can’t fit through the door, it can’t run fast enough to catch a bug, and it panics the second it sees a raindrop.”

Leo stared at the doodle. That was exactly what his model was. He’d trained it on a monster GPU with unlimited compute, fed it thousands of perfect, high-res photos of cats and dogs, tested it only on the same kind of high-res photos it had seen during training. He’d never even thought about what would happen when it had to run on a device with 2GB of RAM, processing photos taken by wobbly third grader hands, of blurry shelter cats or cartoon cats drawn on notebook paper.

“Your ResNet50 has 25 million parameters, right?” she said, and Leo nodded, impressed she knew the architecture. “That’s like that sprite carrying a 50-pound backpack full of stuff it doesn’t need. Half those parameters are for recognizing cars and airplanes and toasters, stuff you’ll never use in your cat-and-dog app. And you quantized it, but you only did post-training quantization, which cuts the precision of every layer. That’s fine for a lab test, but when you run it on a device with limited memory, the rounding errors add up, and the whole thing crashes.”

She leaned forward, resting her elbows on the table. “But the bigger problem isn’t just the size. It’s that your sprite has never seen the real world.