Daniel, thanks, I find these comments very interesting, especially the scheme you are designing. In the near term, I'd like to learn more! Mike
On 5/20/22, Daniel Jue <[email protected]> wrote: > On point #1, Maybe we can expound on the meaning of "judgement". By the > dictionary it's "an opinion or conclusion", but in common parlance we might > interpret a judgement as "a belief that we desire others to agree with.". > An even more cynical take would be "a belief we expect to be accepted by > others on grounds of appeal to authority, i.e an expert or judge". We > might even establish a chain of social value, from greatest to least: > > 1) Correct Judgement > 2) Correct Opinion > 3) Incorrect Judgement > 4) Incorrect Opinion > > Where correctness is based on social acceptance, and judgement vs opinion > is based on the source's expertise in the subject matter. Social value > being a measurement of long term durability and benefit to building upon. > For instance, one might have a "Correct Judgement" that the "concept of > laws" has long term durability, and there are societal benefits from > building on (instantiating) laws. Conversely, the "incorrect opinion" that > some laws are meant to be broken may not have long term durability, and a > negative social benefit if done. > > On point #2, I think only activist AI Ethicists are fighting for control > over the grounds of the decisions (because they see the danger if no one > champions the social good); I believe you are right that a majority of AI > developers are not concerning themselves with the grounds, beyond subtly > allowing their own bias. (Not pointing fingers, in most AI approaches I > believe escaping your own groups' biases is impossible) > > AIKR taken to heart on point #3. > > Agreed on the premise of point #4, and I'll inject that an AGI developer > ought to know that a judgement is an opinion within a space-time-context > frame. > I've continued work on something I'm calling Facet Theory, which is > independent of the one by Guttman on wikipedia by the same name. The goal > is to model contiguous space-time-context frames of understanding as > something called a facet, which have interesting properties where two or > more facets meet along an edge. For instance where two similar, but > incompatible paradigms (groups of opinions) meet, e.g. Newtonian and > Relativistic physics. In that example I like to visualize those as two > facets on a 3D diamond. They are both reasonable approximations of > understanding for their own particular space, time, and context. They may > both be useful models of understanding how the world works (even at the > same time in history). One may reign supreme for centuries, and just > because a more accurate, "better" understanding is discovered in the future > does not mean that it is totally replaced. "Context" in > "space-time-context" is a catch-all for other dimensions of understanding, > such as: > * Socio-Cultural (body language in a certain social or cultural situation) > * Information Source (The news being dire, but from a certain news channel) > * Related Context (The food being good, related to a particular lunch > event; A building having a good aesthetic, within the concept of Brutalist > architecture) > * etc. > > > In this way the AGI may not only survive cognitive dissonance, but thrive > in it by cognicizing at levels above the compartmentalized facets > it encounters. > > Something that is related to (my) Facet Theory but not quite the same is > Prototype theory, where our esteemed Antonio Lieto gets a mention: > https://en.wikipedia.org/wiki/Prototype_theory > I suppose a prototype of a concept would exist at the centroid of one of my > facets. > > What we think of as subjective judgements are just opinions whose variance > is greatest on the contextual scale (i.e. by person, place) and second > greatest variance by time (your favorite music might change over your > lifetime) > However your objective judgements can be modeled as opinions with less > variance over context, space and time, but still subject to instances where > it does not hold true. "This apple" refers to an apple in a particular > space and time, but in the context of an entity moving close to c, the > apple may not be red. It also may not be red in the context of a > colorblind persons' context, as distinguishable from green. > > Self referential statements such as "I think therefore I am", or axiomatic > systems like mathematics absolve themselves from many dimensions and > therefore seem profound to us because they are without counterexample > across time, space, or context. > Fleeting statements such as "my head itches right now" seem least profound > simply because of the number of dimensional constraints on the facet. > > NARS has some great parts, and part of my substrate is based on what NARS > can achieve. However there are a great many things which, once > incorporated into NARS, allow a more human-compatible understanding to take > place. For instance, instead of only frequency and confidence of an > experience, also incorporating a learned confidence of the data source, > when it was learned, etc. > > You may imagine a human case where you have to believe your brother who has > old information, or a dubious politician. An AGI will be put in analogous > situations, and a conscious system also consumes the reflection of its own > past decisions, not only the reputation of external data sources. > > For the last part of point #4, saying "follow the money" may sound cynical > but it's rarely wrong. Modern AI developers (especially Weak AI) will > likely build systems to maximize stakeholders return, whatever the > judgements. > > I don't have it all figured out, but this past year of sabbatical has been > a tremendous help. Instead of trying to cram my own opinions into the AGI, > it needs to be able to interpret reality on its own, and learn to reflect > on its own judgements like a child. From an AI Safety perspective I've > embarked on what I call Pathology First Development, which is basically > generating failure modes of being that are analogous with human > neuropathologies and psychopathologies. The motivation is that if > pathological behavior patterns can be simulated and recognized, an AGI > could be taught to avoid behaviors that lead to these patterns. > > Daniel Jue > > > On Fri, May 20, 2022 at 1:48 AM Mike Archbold <[email protected]> wrote: > >> MOTIVATIONS >> >> Is the following fair? >> >> * There seems to be a prevailing, tacit climate of opinion in AI that >> a judgment is correct primarily if it is 1) equivalent to prior >> human-caused judgments (supervised) or 2) due to rewards. >> >> *Thus there seems to be no need for developers to concern themselves >> with the grounds of the decision; ie., WHY *this* specific judgment? >> (just point to the data and precedent) >> >> *But the problem is the combinatorial explosion: in a real world >> settings often novel variations occur, each nuance bearing a thousand >> little preferences, values, probabilities that have not specifically >> been trained for. >> >> *So it seems like an AGI developer ought to know what a judgment is, >> and to design accordingly. For example, we can think of purely >> objective judgments ("this apple is red") or more subjective judgments >> ("I think blues is superior to jazz, but not in the future"). >> Presently it seems like modern AI regards both as valid if the answer >> fits some pattern. >> >> ~~~~~~~~~~ >> >> By the way I know NARS holds "judgment" as one of its major >> components, need to re-examine) >> >> Mike A > > > -- > Daniel Jue > Cognami LLC > 240-515-7802 > www.cognami.ai ------------------------------------------ Artificial General Intelligence List: AGI Permalink: https://agi.topicbox.com/groups/agi/Ta80108e594369c8d-M6cea3f5a07cd5289a9869d47 Delivery options: https://agi.topicbox.com/groups/agi/subscription
