... and yet a 4 year old can beat the world's most advanced supercomputer in any creative task whatsoever, even something as simple as "draw your house for me."
Humans have feelings (affect valence) which tells our organic neural networks what information to keep and what to discard. Feelings shows us how important data is. What shows computers what data is important? In the current paradigm: humans.
Andrew Ng (Stanford, Coursera) proposed the rule of thumb that AI is good at substituting humans at 1 second tasks (tasks that take humans about 1 second). You can chain several 1 second tasks into more complicated tasks, like driving: 1) visually identifying an obstacle, 2) judging relative speed, 3) deciding the easiest way to avoid a collision. Nobody has ever suggested that abstract reasoning, which is required to draw any correlation to human IQ, can be broken down into 1 second tasks.
A computer cannot imagine, feel, explore, or even learn (no, it cannot). You can give a neural net a vast amount of examples and it will do a good job "predicting" unseen examples. But the neural net is constant, the problem definition is constant. I believe we agree that learning is the ability to create internal models of reality and adapt to your environment. Computers fail at every verb and noun in that sentence.
To conclude, there is no path that takes us from the present paradigm into a future where a singularity is reached and AI "takes over".