Here's an example of it failing to understand some very simple Python code that any human with basic programming knowledge would not mess up:
------
Me: What is going to be the output of this code when `x = 0` and when `x = 900`:
```python
x = 0
for i in range(100):
if x > 50:
x +=1
else:
x += 5
if x == 90:
print("X is not 90")
else:
pass
```
GPT-4: When x = 0, the code will execute as follows:
The initial value of x is 0.
The for loop iterates 100 times.
During the first 11 iterations, since x is not greater than 50, it will increment by 5 each time.
After the 11th iteration, the value of x will be 55, which is greater than 50.
For the remaining 89 iterations, x will increment by 1 each time.
The final value of x after 100 iterations will be 55 + 89 = 144.
During these iterations, the value of x never becomes 90, so the print statement is never executed.
When x = 900, the code will execute as follows:
The initial value of x is 900.
The for loop iterates 100 times.
Since x is already greater than 50, it will increment by 1 each time.
The final value of x after 100 iterations will be 1000.
During these iterations, the value of x never becomes 90, so the print statement is never executed.
In both cases, the output of the code will be empty, as the print statement is never executed.
--------
That is not correct, when x = 0 it will print "X is not 90". I basically create some simple code that is very unlikely to exist in its training data because it doesn't do anything useful and it cannot execute the code. A large amount of the reason that these models seem so impressive is that they were trained on almost everything that exists on the internet, so probably any question you ask it has already been asked somewhere and any useful code you present to it or ask it to generate is included in its training data, or at least something close enough.
So, when I go out of my way to create simple tasks or questions that are highly unlikely to exist in its training data, yet a human could easily handle, it often fails.