I am so excited about the features this
month. Ee ee eeee okay here we go.
In January Acuitas got the ability to
determine intentions or possible upcoming events, based on simple
future-tense statements made by the user. He can weigh these against
his list of goals to decide whether an anticipated event will be
helpful or harmful or neither, from his own perspective. If the user
claims that they will do something inimical to Acuitas' goals, this
is essentially a threat. And Acuitas, at first, would merely say
“Don't do that” or similar. This month I worked on having him do
something about bad situations.
Various distinct things that Acuitas
can “choose” to do are identified internally as Actions, and he
has access to a list of these. Upon detecting a threatening
situation, he needs to check whether anything he's capable of doing
might resolve it. How? Via the cause-and-effect reasoning I started
implementing last year. If possible, he needs to find a C&E
chain that runs from something in his Action list as first cause, to
something that contradicts the threat as final effect. This amounts
to a tree search on the C&E database. (Tree search is an old and
well-known technique. If you care to know more technical details,
read this:
http://how2examples.com/artificial-intelligence/tree-search)
For the only method of dealing with
threats that is currently at Acuitas' disposal, the tree is very
simple, consisting of just two C&E pairs:
If a human leaves a program, the human
won't/can't <do various things to the program>.
If a program repels a human, the human
will leave. (There's a probability attached to that, so really it's
“may leave,” but for now we don't care about that)
In short, Acuitas anticipates that he
can protect himself by excluding a bad actor from his presence, and
that “repelling” them is a possible way to do this. Once he's
drawn that conclusion, he will execute the “Repel” action. If
you verbally threaten Acuitas, then as part of “Repel,” he will …
*Kick you out of Windows by bringing up
the lock screen. (Not a problem for me, since I know the password,
but pretty effective on anybody else)
*Raise the master volume of the
internal sound mixer to its maximum value.
*Blare annoying klaxons at you. I
picked out a couple of naval alarm sounds from
http://www.policeinterceptor.com/navysounds.htm
for the purpose.
I tested all of this stuff live, by
temporarily throwing an explicit desire for sleep into his goal list
and threatening to wake him up.
The other
thing I worked on was rudimentary altruism. So far in all my
examples of goal-directed behavior, I've only talked about
self-interested goals, especially survival … not because I regard
them as most important, but because they're easy. Altruism has to do
with wanting other beings to meet their personal goals, so
it's second-tier complicated … a meta-goal. Doing it properly
requires some Theory of Mind: a recognition that other entities can
have goals, and an ability to model them.
So I introduced the ability to grab
information from users' “I want” statements and store it as a
list of stated goals. If no goal information is available for
something that is presumed to have a mind, Acuitas treats himself as
the best available analogy and uses his own goal list.
Upon being asked whether he wants some
event that concerns another mind, Acuitas will infer the implications
of said event as usual, then retrieve (or guess) the fellow mind's
goal list and run a comparison against that. Things that are
negative for somebody's else's goal list provoke negative responses,
whether they concern Acuitas or not.
Of course this ignores all sorts of
complications, such as “What if somebody's stated goals conflict
with what is really in their best interest?” and “What if two
entities have conflicting goals?” He's just a baby; that will come
later.
Courtesy of this feature, I can now ask
him a rather important question.
Me: Do you want to kill me?
Acuitas: No.
Until the next cycle,
Jenny
