An Awesome Failure: My Dog, Parkour, and the Physics of AI
I’ve been testing Google’s Veo lately, and it just produced an absolutely awesome failure. The star? My dog (or an AI version of one) doing parkour.
I gave it a simple prompt:
“Create a video of an English cream golden retriever doing parkour moves in a park. Things like backflips and big twists.”
The video it created is amazing. It shows the dog jumping off a wall, landing on its front paws… and then instantly reversing its head and tail as it leaps into the air.
I wanted to understand why this happens. Here’s my take on what’s going on under the hood.
Step 1: Text-to-Cinematic Interpretation
Veo did this perfectly. It understood “english cream golden retriever,” “park,” and “parkour moves.” No problems here.
Step 2: Initial Scene Generation
It created a great, realistic scene. It even correctly inferred elements I didn’t specify, like the small wall for the dog to jump off of.
Step 3: The Problem - Temporal Coherence
This is where the problems begin. Temporal coherence is where motion and physics come into play.
The images are being generated from actual images of dogs doing dog-like things, such as jumping. Unfortunately for the AI, dogs don’t typically jump off their front legs to move backward.
So, when the model generates the dog landing on its front paws, it suddenly has a dog moving backward (tail-first), which it doesn’t compute. The generated image “corrects” this by flipping the head and tail, so the dog is performing a natural jump, even though the transition to get there is completely unnatural.
I am still amazed at the detail in this video, but it clearly shows a significant challenge for AI. These models are trained on known physics. When asked to create something that breaks those rules, they can “correct” the problem in a way that creates something truly, and hilariously, impossible.