Engineering Kindness: Building A Machine With Compassionate Intelligence (2015)

I wasn't expecting compassionate intelligence via predicate calculus based expert systems. That's very 1980s, as is the whole paper. There's no mention of machine learning. There's a lot here about "common sense", which remains an unsolved problem in AI.

Thinking about common sense without getting into philosophical tangles is difficult. It's helpful to consider common sense as the ability to predict the consequences of actions. If you can predict consequences, (probably probabilistically) you have the basis for planning. And you can avoid getting into trouble, which is what common sense is mostly for.

This is known technology. In modern control theory, you set up a predictor, have it watch the "plant", and self-tune until it can predict what the plant will do. This is called adaptive control, and it's a kind of machine learning, with similar math but different terminology. Once you have a good predictor, you can invert the predictor function and solve for the inputs which will yield the desired outputs. This works on systems which are reasonably continuous. Many industrial processes are controlled by such systems.

Prediction is a prerequisite for compassionate intelligence. It's not having internal emotional state that's crucial. It's having a predictive model of the other party's emotional state.

Of course, once you have a tuned model of somebody, you now know what buttons to push to make them do things, in either the "buy" or "vote" spaces. Whether something is compassionate is a design decision.

Interesting proposal. I definitely think that implementing metacognitive facilities, as well as internal representations of the perceived emotional states of the actors around the AI, seem like good steps in the right direction. Predicate logic will never work, though.

One other aspect I found a bit odd:

> As an agent designer, it gives one confidence in choosing “insight meditation” as a philosophy of mind because the mental processes have been documented to create compassion in even the most hardened criminals (Ariel and Menahemi, 1997).

This strikes me as a bit of an absurd criteria for choosing what is, essentially, a software algorithm. This system is so fundamentally different from a human being, that I don't think any such parallels are warranted.

It's kinda funny, but I kinda don't want to make fun of it, so I'm left speechless. Maybe others too? Hence no comments on the content in a hour?

I do think that the Turing test is still the greatest test for strong AI, machines are still very far from passing it, and I feel that understanding human emotion should be a part of the general strategy of "understanding human things" on a level sufficient to pass a very long Turing test session:

* From a technology standpoint it's not clear that what we'd call human emotion requires a different approach than other "human things."

* From an application standpoint, a machine much dumber than an adult (as in totally can't pass the Turing test) but with developed emotional facilities is kinda scary, perhaps like an alien child ("human-like feelings" with undeveloped reasoning, and the reasoning is undeveloped in ways we're not familiar with - at least I'm conjecturing that this machine will not be able to pass "the human children Turing test" any more than the adult version.) It's not clear what good you can do with such a machine.

All of the above from a serious angle and without dwelling on the joy of sexprs like (Feels (Feels (Feels (Object)))).

Option (a) create a 'motherly' machine which is guaranteed compassionate towards children and small furry mammals,

Option (b) create a Buddha-like machine which is guaranteed benevolent towards everyone and everything in the Universe.

Neither of these work! Wrt (a) the flip side of maternal empathy is a ferocious attitude towards potential predators who threaten her cubs. How does the programmer guarantee no human will be placed in this category, mistakenly or otherwise? Besides, before one attains motherhood one must pass the 'terrible twos' which many psychologists regard as the peak stage of physical violence.

Wrt (b) an AI with no desires has no incentive to interact with the world. So its mind will not develop.

Is there a third option?

Oh, this was killed because of the redirect through twitter, http://www.emotionalmachines.org/papers/engineeringkindnessw...