People keep talking about if/how AGI alignment is possible. I think different people may mean slightly different things, but possibly they mean one of the below two things.
1) AI that will know what is good for us and will work to benefit us.
or, 2) AI that will function to fulfil its controllers'/creators/training data generators' stated objectives.
concerns wrt to 2 becomes simply that of possession of any other powerful tool/weapon, so lets focus on the question from point of view of 1.
okay, lets see where does the already existing super intelligence - the human mind stand? Human mind, in majority of the cases, appears to be limiting us even harming us. from laziness to addictions, procrastination to shortcut seeking, inability to concentrate to inability to take something off the mind, it appears to be doing things clearly not in our best interest lots of time.
Shashtra's too are full of teachings for how to control मन (mind) so that we can use it to achieve greatest good for us and prevent harms caused by its wrong usage. i.e. Envy of the bestest AGI itself is not inherently "aligned".
Why?
because a tool cannot decide its command, at best it can be made powerful, efficient, durable, AUTOMATED like our mind. it can automatically pick up our commands and act on it, it can run on auto pilot while no commands are being issued, but if user gives commands not aligned to his own self interest, the tool will anyway function as per the command received (if user's own other superseding command doesn't interfere.), the tool cant itself know or want what is "good" for the user.
For those of us who see/understand/accept difference between जड/प्रकृति and चेतन, this is easier to see. that is, concept of "alignment" cannot be defined as (1) at all in relation to what we are calling intelligence in current context. Artificial or Natural. for they are things made of जड/प्रकृति (material components). Alignment is something that चेतन himself needs and he has freedom to act in aligned or non aligned fashion with respect to self interests.
Knowledge, benefit, harm, happiness, desire, effort etc are applicable only for the चेतन. whereas concepts that are applicable to a tool are efficient, powerful, durable, configurable, automated etc.
Just like efficient, powerful, durable, configurable, automated doesn't make sense for चेतन, knowledge, benefit, harm, happiness, desire, effort makes no sense for non alive. i.e. AI or AGI or whatever super intelligence, it cannot know, cannot desire, cannot make an effort on its own. it can only work based on what you know, what you desire, what you are making an effort for. you tell it in runtime what your knowledge, desire is or you ask it to infer from what you have told it in past.
No AI can ever be inherently aligned (per definition 1) with human interest, not because it would 'want' to be non aligned or we wont achieve that 'level' of success, but because the concept of alignment makes no sense with respect to that thing.
So, alignment as defined as (1) is not an applicable quality for association with AI and when defined as (2) concern is nothing new or AI specific. it simply is the good old problem if are we are using tools at our disposal to work towards our own good or destruction? In short, alignment discussions have no meaning at all. Discussion should be how to use it properly along with thousand other tools we use already.
Disagree? Doubtful? consider that if it was applicable concept in first place, wouldn't we be seeing natural intelligence/mind being inherently "aligned"? are we seeing it in reality? is the Nature/God stupid?
Related
https://nyaydarshannotes.blogspot.com/2023/01/Yog1.5.html
---------------
posted originally at
https://ekakinchan.blogspot.com/2023/04/is-natural-intelligence-aligned.html