One of the most compelling issues regarding AI would be in behavior and relationships.
What are some of the methods to address this? For example, friendship, or laughing at joke? The concept of humor?
Artificial Intelligence Stack Exchange is a question and answer site for people interested in conceptual questions about life and challenges in a world where "cognitive" functions can be mimicked in purely digital environment. It only takes a minute to sign up.Sign up to join this community
Relationships and normal social behavior require a human to possess a reasonable "theory of mind", a skill in understanding and modeling the thought processes that happen in the minds of others, and making reasonably accurate predictions on how particular actions will be understood by others.
In general, this might be treated as any other machine learning/prediction task - while this skill is quite complex and too hard for our current systems, there doesn't seem to be any obvious qualitative barrier that needs to be breached. Notably, there's no reason to believe that a mind needs to be able to experience a certain "feeling" in order to model that other minds can have it - an AI agent could form a causal model of how friendship works in human relationships and use that to exhibit behavior that's consistent with friendship in all aspects and/or facilitate friendship behavior towards it by particular humans, if it fits the agent's goals. Whether you'd consider that "real friendship" is pretty much just a matter of how you define the word, with some parallels to the "p-zombie" discussion.
In general, what you are describing implies a hierarchical sequence model, in which mannerisms adapt to the regime or paradigm in effect. Expressive modalities are how we recognize the operative context from the behaviors of other agents. For an artificial agent, avoiding un-canniness would involve clustering the factors underlying the classification of discourse contexts, tagging them with corresponding factor models for manner in a way which closely aligned to the classifications which human agents make. This involves developing an adequate latent representation for both spaces, which is likely to involve quite a lot of training data, or some clever transfer learning in conjunction with one-shot techniques (generally, taking samples as modes).
When the context becomes one of friendship, for example, the style factors of expression should extend to factors of trust, collaboration, disclosure, sympathy. In order to implement these, layered abstractions in the representation of behaviors will need to be crafted, presumably by training. In order to align these learned categories to human categories, which I suppose to be essential to emotional fluency, exploiting the structure of corpus semantics - including distributional characteristics of the linguistic labels and inferences - when estimating loss gradients seems a natural strategy, which exploits cultural learning.
These comments are necessarily speculative, as no such systems are current, to my knowledge. At least, not in any significant degree of maturity or application. Certainly other approaches are conceivable. I am just extrapolating plausible strategies, in the context of current technologies, strategies for implementing the kinds of behaviors you describe in a useful way.
More, specifically, I am considering the question as a manifold-learning problem, and hence requiring a representation model in which the inputs and outputs vary over state-spaces, with the behavior being learned as a mapping between those spaces. Each of those components (the input space representation and its natural topology, the output space representation likewise, the mapping between them, and the sequence model hierarchy which extrapolates from the manifold to a control process emsemble) has its own peculiar challenges, and the challenges to creating an adequate implementation are a confound of those. Still, it seems more a question of time and money than of notional feasibility.
Methodology for a social enabled AI.
Regarding social interactions, I believe that trying to build a copy of human behaviour based on technologies we understand, might not be effective .
Instead, I would start from the roots of how human grows and learn, or better, by what is each human trying to solve from childhood to the end of their life. In that way we will be able to build artificial beings, able to exploit the best of what technology might provide.
It seems to be based on emotions.
As an example, humour is the kind of social trick a gifted social artificial life could learn and master, first, like a child, then like all of us, in order to empathize with others when this common emotions arise, before to analyse the conditions in witch it arises so that to be able to reproduce it, then eventually if the need become a priority, by theorizing its mechanics, possibly mastering it and using it like a tool.
(I didn't studied the details, but humour is often generated by instigating surprise to others, it requires modelling how others generally think, leading them to be surprised.)
Mastering social interactions require modeling
Humour, friendship as examples of social interaction should be naturally discovered by IA like by Human, and eventually used in order to fit once goal. On the other hand this require intelligence to be sensible to surprise in a positive way, leading either to laugh, interest, curiosity or fear depending specific conditions. Mastering that process requires to model both the world (how you think) and how others do model it (how they think).
Modeling is greatly accelerated by interactions and communication
Additionaly, Human may trigger others laugh without building a conscious theory , this help us understand intelligence and build an artificial one: Language is here used to transmit initially abstract concepts built by intelligence. It helps us accelerate our modeling of the world, that would remain very basic if only based on our own experience.
Most important skill for Modeling is observation, logic comes second, interactions comes third
Unconscious resolution of problems seems to be a characteristic of live-beeings, but also of machine-learning techniques like deep learning that is criticised for that. Because we would like it to be able to explain us the concepts they discovered and used to solve the problems we gave it. This is because we would like that tools to build abstract and possibly original models, before to give it tha ability
A causal and statistical approach backed by hierarchical goals including the hunger for discovery
For this to happen in a complex world made both of physics and though, a causal statistical approach seems required, backed by, to give the direction to take: a hierarchical list of goal two of wich
should manage to find their best possible place behind that to survive, allowed or helped by the conscious to be a finished beeing.
For discovery goal to reach its best level, biological related mechanics involving reproduction, variability, complexity and adaptation to environment, could be involved, but this seems to go beyond the question.
Still some work remains to be done in order to give our reasoning tool the ability to comunicate, that is to master a language, and to use it given the knowledge of what concepts and semantic domains, interlocutors do already master, to then build new concept in a transmissible form like words, images, video or just body-movements, face expression or so.