In its most basic form: the continuous formation/recall of associations and disassociations between stimuli.
That is the atom of thinking. But even that rules out ANNs pretrained using gradient descent since they can’t learn in response to stimuli.
The next simple step up is the ability to overwrite those learned associations which probably is solved by completing the first task but it’s important to note.
The following step up is being able to process those associations/memories themselves as stimuli which can be associated forming new connections/memories that can be recalled and associated and so on ad infinitum.
That rules out finite non-recursive neural nets, and that still puts us no where near AGI. It’s basically just an ml taxonomic tree builder
The next necessary piece for actual high level thinking is curiosity and action. The system needs to be able to take a concept and a relation (abstract idea of a particular kind of association between two concepts) which currently don’t form a known relationship, and attempt to find the missing concept which would complete that relationship. (Most likely this involves either finding relationship chains between the two which in all other known cases are equivalent to a known relation OR following some procedure to experience the relationship directly as stimuli.)
Lastly, in order to experience the idea of thinking, we first need to have the ability to form temporal associations, but also have working memory of recent past thoughts, and finally the ability to compare those with previous recent memories or the most current stored thoughts. Finding patterns in our thinking helps us be better at thinking. (This is a prerequisite for the experience of free will and consciousness)
That’s what higher level thinking entails.
Of course tuning the system’s response to recognized associations, how memories are stored and recalled, how frequently the system experiences curiosity, etc. will all impact the overall intelligence of the model, but considering most current “AI” don’t even check off the first box, that’s all a bit out of scope.
In its most basic form: the continuous formation/recall of associations and disassociations between stimuli.
That is the atom of thinking. But even that rules out ANNs pretrained using gradient descent since they can’t learn in response to stimuli.
The next simple step up is the ability to overwrite those learned associations which probably is solved by completing the first task but it’s important to note.
The following step up is being able to process those associations/memories themselves as stimuli which can be associated forming new connections/memories that can be recalled and associated and so on ad infinitum.
That rules out finite non-recursive neural nets, and that still puts us no where near AGI. It’s basically just an ml taxonomic tree builder
The next necessary piece for actual high level thinking is curiosity and action. The system needs to be able to take a concept and a relation (abstract idea of a particular kind of association between two concepts) which currently don’t form a known relationship, and attempt to find the missing concept which would complete that relationship. (Most likely this involves either finding relationship chains between the two which in all other known cases are equivalent to a known relation OR following some procedure to experience the relationship directly as stimuli.)
Lastly, in order to experience the idea of thinking, we first need to have the ability to form temporal associations, but also have working memory of recent past thoughts, and finally the ability to compare those with previous recent memories or the most current stored thoughts. Finding patterns in our thinking helps us be better at thinking. (This is a prerequisite for the experience of free will and consciousness)
That’s what higher level thinking entails.
Of course tuning the system’s response to recognized associations, how memories are stored and recalled, how frequently the system experiences curiosity, etc. will all impact the overall intelligence of the model, but considering most current “AI” don’t even check off the first box, that’s all a bit out of scope.