Alex,

Thanks for that example.  It shows the importance of the unconscious computation that is performed in the human cerebellum, whose perceptions and actions are totally unconscious.   I urge everyone to click on the link in your note.

There is an important reason why the human drone experts lost in the competition with the fully automated drone:  the humans used a combination of high-speed cerebellar computation (as the unmanned drone does) with the much slower (and conscious) decision making in the cerebral cortex.   Those conscious decisions slowed their performance.

Compare that with the high-speed performance by the gymnastic champion Simone Biles.  She devoted years of conscious effort to train her cerebellum to perform the various motions automatically.  Before each competition, she perfects the training for each routine she performs.  In a performance that has multiple routines, she uses her cerebral cortex to check the positions and timing for each routine.  Then she launches a pretrained routine that is totally under the control of the unconscious cerebellum.

All of us use the cerebellum for routine processing in walking, eating, driving a car, or typing on a keyboard.   Mathematicians take advantage of that high-speed processing in the most complex kinds of math.   But writing a proof uses the slower conscious processing in the cerebral cortex to check whether the high-speed calculations are correct.

Note that the processes in the cerebellum are precise for what they do.  The errors can occur when the decisions for running them (made by the cerebral cortex) are not correct.

Note that none of these processes, either by the cerebrum or by the cerebellum, could be performed by the LLMs.  The Large Language Models might respond to a verbal command to execute a routine by the cerebellum.  But all their operations are probabilistic, and they're based on vague and often ambiguous natural language.   They can't  do the precise checking and testing that guarantee accuracy.

LLMs are useful.   But they're just one more tool in the huge toolkit of AI technology.  They do a limited range of operations very well, but they can't do the whole job.

John