That's a terrible analogy because a human can trivially count the beans. If GPT has any hope of being a decent image model, then counting most definitely is one of the tasks that it needs to be capable of doing correctly. It's astounding that the benchmarks and tests have ignored such basic image tasks. Even a five year old can count them correctly.
GPT is primarily an LLM which predicts the next words in a response. To give it the capability to count the beans you will need a specialized GANN or image recognition software that can locate and categorize beans specifically then create a count. (Duct taping the torque wrench to the hammer)
Then, when you throw a bunch of skittles in there you will need to train the GANN to ignore the skittles. (Duct taping a screwdriver to the torque wrench.)
This is all an effort to approximate a general intelligence like a 5 year old which has had thousands of tools duct taped to their brains over a hundred million years of evolution.
