ChatGPT passes the U.S. medical licensing examination.

The biggest difference between machine learning and human is that human uses fuzzy logic. The machine uses precise logic for its actions. When a machine operates, it uses precisely described action chains. The problem is that no pre-programmed action chain has a match for a real-life situation. 

On the streets are some surprises. Like suddenly coming cars are everyday situation. Nobody can predict all possible situations that robots can face in real life. And that makes the system hard to program. 

Precise logic means that programmers must describe every single thing that requires notice or reactions to the machine. The machine cannot operate. If something is missing from its code. 

When the learning system will face that kind of situation it says "unknown situation". And then ask for orders from the human operators. 

The reason. Why ChatGPT passed the U.S. medical examination, but the AI cannot recognize U.S. marines that closed the base because the type of information that AI must handle in the examination is limited. 

In the examination, the AI uses a precise and limited type of information. The AI must just handle letters with precise ASCII values. AI is best in that kinds of missions. 

And when marines closed the target. The AI must handle uncertain types of information. The difference between human and machine learning is that the coders must precisely describe the missions to the AI. 

Sometimes that the person takes a box or some unexpected item like red paper. And cover the head by using that item is enough to cheat the AI. The reason for that is simple. Nobody ever described a person whose head is covered with paper to AI. And that means the AI goes into a trap. 

In the test, marines used trees and boxes. And the AI didn't recognize them. Nobody even describes the case. The two men walk to the target area and keep the box over their heads. Nobody describes. The situation for the AI is that the box with four feet are closing it. 

The AI is the support system. Not the main operator. But the systems advancing. And that kind of test is needed for the R&D work of that system. The same thing that made it easy to cheat the AI is making it hard to make robots that can go shopping for people. 

There are so many variables that making perfect descriptions for every possible situation is hard. Programmers are hard to describe all possible situations, and blocks to the computer. 

But why the AI that mission is to launch an alarm if some unauthorized person is in house works? Sometimes regular alarm systems make alarms. When the burglary alarm launches that system sends information to the guard, who checks the situation from the screen. 

The fact is that the AI uses precise logic. If the alarm system is on, the AI simply knows that nobody should move in some area. And then if something moves in some room, it sends information to the human operators. The AI can detect cases like possible fights in public transportation. And then make an alarm to human operators who make decisions on how to act in that case.

https://www.unilad.com/news/us-marines-defeat-pentagon-ai-robot-cardboard-boxes-175338-20230121

https://scitechdaily.com/the-rise-of-artificial-intelligence-chatgpts-stunning-results-on-the-us-medical-licensing-exam/

https://astronomyandtechnology.blogspot.com/