Torras.ai

Recently, I stumbled upon an intriguing puzzle on Twitter (yes, I'm old school and still call it Twitter). Thanks to a post by World of Engineering.

I found myself scratching my head over a spatial intelligence challenge. It took some effort, but I finally cracked it!

This puzzle presented a perfect scenario to test GPT-4o and its ability to process images and handle spatial tasks. Surprisingly, despite several attempts, GPT-4o couldn’t manage to get the right answer, not even close!

Curious, I decided to run a little experiment. I asked a bunch of friends and family to try their hand at the same puzzle. Unlike GPT-4o, they all managed to solve it—though, like me, most needed pen and paper to work through the problem.

The method I used involved giving letters to all unknown distances:

And then combining all unknown edge lengths to create a system of equations, where "P" is the perimeter we are looking for. The equations look like this:

1. P = a + b + c + 12 + 8 + d + b + e + b + a + f
2. a + b = 10
3. b + c = 8
4. d + e + f = 12

You can actually solve these equations for P, even when the rest of variables would remain unknown.

Interestingly, while GPT-4o floundered with the spatial aspect of the puzzle, it nailed the equations when presented in text form. In fact, it solved them perfectly—a clear step up from its predecessor, GPT-3.5.

If you're curious about the solution, you can actually ask GPT-4 directly by visiting their site and entering the equations or click below:

Solve (maybe) the Equations with GPT-4

Additionally, Cesar Lizarraga offered a brilliantly intuitive solution that simplifies everything, no paper needed!

This experience highlights a critical point: despite advancements in AI, humans still hold the upper hand in spatial reasoning and intuitive problem-solving. It's a reminder that while AI can perform specific tasks with astonishing accuracy, it seems then General Intelligence are still largely human domains, so we need will to wait for AGI a bit more!.