Now, we see the lens for what it truly is –
A pow­er­ful but flawed entity grop­ing its way through an inter­nal environment.

We probed the inner work­ings of the generic mind and famil­iar­ized our­selves with lenses in our pre­vi­ous post. We also learned from the mere exis­tence of issues like the prob­lem of for­eign con­text that lenses can be less than per­fect. Now it’s time to take this issue to the next level and call our atten­tion to greater flaws lurk­ing just beneath the sur­face. We will begin by casu­ally observ­ing a lens in its nat­ural envi­ron­ment, and will even­tu­ally end by tor­tur­ing the poor spec­i­men with a bat­tery of thought experiments.

The Internal Environment

What is the best word to describe an abstract space gov­erned by a set of rules, capa­ble of host­ing enti­ties that inter­act with one another accord­ing to these rules? A group? A field? A sys­tem? I tend to think of these as envi­ron­ments, no mat­ter how abstract its rules or enti­ties may be. Just as the Earth is an envi­ron­ment occu­pied by count­less liv­ing organ­isms, the Internet can be thought of as an envi­ron­ment for data and pro­grams, and per­haps the law can be seen as an envi­ron­ment filled with doc­u­ments and his­toric cases. In much the same way, each generic sen­tient being con­tains a unique inter­nal envi­ron­ment host­ing a com­plex inter­ac­tion between mem­o­ries, emo­tions, and beliefs. Every obser­va­tion, action and moment spent in deep thought will affect the rela­tion­ship between the enti­ties in this inter­nal envi­ron­ment, even if the effects can be quite sub­tle. Conversely, the inter­nal envi­ron­ment will strongly affect the judg­ments a generic will make and the actions he will take. In fact, the inter­nal envi­ron­ment is prob­a­bly just as impor­tant to a generic as the exter­nal envi­ron­ment that he lives in.

It’s pos­si­ble to cre­ate envi­ron­ments that only con­tain mind­less rule-abid­ing enti­ties, but wouldn’t it be more inter­est­ing if the enti­ties had a greater aware­ness of their envi­ron­ment? Generics are more than just pas­sive objects in their exter­nal envi­ron­ment; they also use their senses and deduc­tive skills to bet­ter under­stand that envi­ron­ment. They can make pre­dic­tions about the future, and make informed deci­sions based on these pre­dic­tions. Of course, a generic only can only learn about their envi­ron­ment through the small win­dow that his senses pro­vide, and his model of the envi­ron­ment can only be built based on what he knows¹.

A direct par­al­lel can be made with lenses. The lens occu­pies a cen­tral posi­tion in a generic’s inter­nal envi­ron­ment, being involved directly or indi­rectly with almost every­thing there. It too can respond to changes in its envi­ron­ment using, among other things, a set of inter­nal senses. Just like the exter­nal senses, the inter­nal senses only pro­vide a small win­dow into the com­plex activ­ity of the inter­nal envi­ron­ment. Naturally, the lens can only make casual esti­mates and eval­u­a­tions based on what it knows from its senses. It’s impor­tant to note that since all gener­ics have a unique inter­nal envi­ron­ment, the inter­nal senses are com­pletely pri­vate. A generic can­not sense any­thing about the inter­nal envi­ron­ment of another generic, and even if he could he wouldn’t be able to make any sense² out of the read­ings with respect to his own inter­nal environment.

Why Evaluation

My pre­vi­ous post paid plenty of atten­tion to the causal esti­mate, with a hard focus on its role in causal pre­dic­tion and the prob­lem of for­eign con­text. It prob­a­bly goes with­out say­ing that prior and pos­te­rior infor­ma­tion are also quite impor­tant, but what about the eval­u­a­tions? Even if we wave our hands and accept that eval­u­a­tions are needed to give emo­tional con­tent to the expe­ri­ence foun­da­tion, we still haven’t assigned any inher­ent mean­ing to the eval­u­a­tions. What stops the lens from cre­at­ing com­pletely ran­dom evaluations?

Just like how the lens uses causal pre­dic­tion to exer­cise the accu­racy of its causal esti­mates, the lens uses value pre­dic­tion to put its eval­u­a­tions to the test. On the sur­face, value pre­dic­tion works in the exact same way as causal pre­dic­tion – the lens just uses an eval­u­a­tion to make pre­dic­tions rather than a causal esti­mate. However, this might sound like non­sense. Evaluations don’t cap­ture cause-and-effect rela­tion­ships in the phys­i­cal world, so how can they be used to make any kind of prediction?

If you think about it, causal esti­mates relate the phys­i­cal world to itself (even if the phys­i­cal world includes inter­ac­tions with other gener­ics), whereas eval­u­a­tions tie exter­nal events and actions to inter­nal beliefs and emo­tions. In a sense, eval­u­a­tions can rep­re­sent cause-and-effect rela­tion­ships between the exter­nal and inter­nal envi­ron­ments. A generic can use value pre­dic­tion to pre­dict their future actions from their emo­tions or pre­dict how an event will affect his beliefs about the world. In any case, infor­ma­tion about the inter­nal envi­ron­ment is col­lected using the inter­nal senses, and can be made avail­able as either prior or pos­te­rior information.

Because the inter­nal senses can­not be shared, value pre­dic­tions are dis­tinctly self-cen­tered. This doesn’t mean other gener­ics can’t be involved though. When some generic A gives his coworker generic B an impor­tant task, A may pre­dict that B will work hard and even­tu­ally fin­ish the task based solely on a feel­ing of trust, even if he had no fac­tual evi­dence to back this par­tic­u­lar pre­dic­tion³. Some generic C might be con­vinced that a group of author­ity fig­ures will soon make a deci­sion that he strongly dis­agrees with, sim­ply because the group demon­strated very dif­fer­ent moral beliefs on unre­lated issues in the past. Generics some­times assume that the inter­nal envi­ron­ments of other gener­ics are sim­i­lar to their own, and per­form value pre­dic­tion on behalf of another generic. This is risky and almost always mis­taken though. If generic A pre­dicts that generic B will spend the whole day on the task because A him­self felt very moti­vated at the time, then he is using value pre­dic­tion and is quite likely to be wrong. This is not to be con­fused with the case where A has evi­dence to believe B too wanted the task done quickly, where he is using casual pre­dic­tion and might not be wrong after all.

Other Uses of the Lens

There are some other tasks that the lens can do, though these capa­bil­i­ties are used less often than event inter­pre­ta­tion, causal pre­dic­tion or value pre­dic­tion. For one, it can per­form causal retro­d­ic­tion and value retro­d­ic­tion. Retrodiction is the oppo­site of pre­dic­tion, where cause and effect are stud­ied in oppo­site direc­tion. Just as some effects can be deduced once one knows the causes, some causes can be deduced once one knows the effects. This can be used to learn more about an event once a generic learns more about its effect on the future, or to recon­struct older mem­o­ries based on more recent memories.

The lens can also per­form the more exotic task of event con­struc­tion, where both prior and pos­te­rior infor­ma­tion are cre­ated out of thin air to cor­re­spond to a given causal esti­mate and eval­u­a­tion. This can be use­ful to cre­ate exper­i­ments that ver­ify the accu­racy of either the causal esti­mate or the eval­u­a­tion. I sup­pose it can also be used to cre­ate cau­tion­ary tales to illus­trate cer­tain beliefs one has about the world, or fic­tional sto­ries built around cer­tain emo­tional states one had at the time.

Overgeneralization

The lens can indeed do many tasks, but it doesn’t always do them well. It def­i­nitely makes mis­takes, and not just because the exter­nal and inter­nal senses can’t present a per­fect view of the exter­nal and inter­nal envi­ron­ments. The lens even­tu­ally has to assume that the causal esti­mates and eval­u­a­tions learned from the past will gen­er­al­ize to future sce­nar­ios, and some­times it may be overly zeal­ous in its gen­er­al­iza­tion. This leads us to one of the most notable quirks of the lens – its ten­dency to over­gen­er­al­ize⁴.

There can be sev­eral kinds of over­gen­er­al­iza­tion. In uncon­di­tional gen­er­al­iza­tion, the lens tends to pro­duce the same out­puts regard­less of the inputs it receives. For exam­ple, sup­pose generic A’s lens con­stantly pro­duces eval­u­a­tions related to dread or anx­i­ety dur­ing event inter­pre­ta­tion, no mat­ter what events occur. Maybe A expe­ri­enced many events in the past that really made him feel dread­ful and anx­ious – so many that his lens learned to inter­pret all future events in a sim­i­lar way. More sub­tly with pin­ning, the lens tends to pro­duce a fixed out­put given a spe­cific value for one of its inputs, regard­less of the other inputs. A nice exam­ple here would be a lens per­form­ing causal pre­dic­tion which always assumes that good inten­tions (causal esti­mate) lead to good out­comes (eval­u­a­tion + pos­te­rior infor­ma­tion), regard­less of the sit­u­a­tion (prior infor­ma­tion). In this exam­ple, we say that the expec­ta­tion of a good out­come is pinned to the assump­tion of good intention.

In all cases of over­gen­er­al­iza­tion, the lens learns a false pat­tern that doesn’t hold in gen­eral. What hap­pens when a generic encoun­ters evi­dence against his gen­er­al­iza­tion? Let’s sup­pose that the lens in the first exam­ple learns from its inter­nal senses that it feels excited after some event, which does not mesh well with the eval­u­a­tion of the event as dread­ful. On a lucky day, the lens will loosen itself and pre­dict less dread­ful­ness for the future. But it also has plenty of ways to avoid the truth. It can try to find strange rea­sons to inter­pret the event as dread­ful any­way, or it could treat this spe­cific event as a rare excep­tion to the rule and con­tinue to over­gen­er­al­ize. In the worst case, the gen­er­al­iza­tion sur­vives even after being chal­lenged many times. What if two gen­er­al­iza­tions dis­agree on a sin­gle event? Let’s pit the first gen­er­al­iza­tion in the pre­vi­ous para­graph against the sec­ond. If generic A learns that a well-mean­ing friend wants to help him on a task⁵, will he feel con­fi­dent that the task will be done well, or will he feel anx­ious and dread­ful? This exam­ple of dis­so­nance is some­what like apply­ing an unstop­pable force to an immov­able object. If the two gen­er­al­iza­tions clash against one another often enough, one of them could dis­ap­pear indef­i­nitely. An even more bizarre result would be a mod­i­fi­ca­tion to both gen­er­al­iza­tions which empha­sizes the com­mon­al­i­ties between the two. Perhaps the unstop­pable force and the immov­able object both sur­vive the encounter with noth­ing more than a scratch, and generic A gets close to sat­is­fy­ing both gen­er­al­iza­tions by feel­ing grate­ful, but guilty and undeserving.

It’s prob­a­bly near impos­si­ble to com­pletely avoid over­gen­er­al­iza­tion, but it does seem like there are ways to mit­i­gate its effects. One could try to develop a cul­ture of open-mind­ed­ness that encour­ages gener­ics to drop their gen­er­al­iza­tions in the face of coun­terex­am­ples. Failing that, it is also pos­si­ble to use dis­so­nance to one’s advan­tage. One can set up sce­nar­ios that force a generic to rec­on­cile two con­tra­dic­tory gen­er­al­iza­tions, which can either out­right elim­i­nate one of the gen­er­al­iza­tions, or con­vert both into forms that are eas­ier to deal with. Then again, over­gen­er­al­iza­tion might not even be as bad of a prob­lem as it seems; a more con­tro­ver­sial point of view might even con­sider gen­er­al­iza­tions to be char­ac­ter­is­tic per­son­al­ity traits of a generic sen­tient being, with the asser­tion that forc­ing all gener­ics to believe only in fac­tu­ally grounded state­ments robs them of their indi­vid­u­al­ity. I par­tially agree with this sen­ti­ment – in my eyes, most cases of over­gen­er­al­iza­tion are just inter­est­ing quirks of a generic’s lens, not mal­adies that must be cured against all costs.

1. Can one even describe an exter­nal envi­ron­ment with­out ulti­mately ref­er­enc­ing sen­sory per­cep­tion? Do the senses give qual­i­ta­tive expe­ri­ences in a con­scious mind, or are they more sim­i­lar to sen­sor read­ings in a mech­a­nis­tic sys­tem? Quite sur­pris­ingly, these ques­tions are actu­ally not all that rel­e­vant to this field; every­thing in this post applies no mat­ter what answer you have to either question.
2. Pun not intended.
3. Why the feel­ing of trust then? It could have been earned through com­pletely dif­fer­ent means, such as by hav­ing friendly con­ver­sa­tions in the past. It’s even pos­si­ble that A hap­pened to be in a good mood, allow­ing him to trust every­one at the moment.
4. It should be noted that what counts as over­gen­er­al­iza­tion and what doesn’t is a mat­ter of opin­ion. Still, there are many cases where it’s clear that a generic gen­er­al­izes too far even after find­ing strong evi­dence against their beliefs. The oppo­site prob­lem – under­gen­er­al­iza­tion – feels less like a mis­take and more like some kind of sub-opti­mal use of knowledge.
5. Assume that beyond all rea­son­able doubt, A knows that this friend has good intentions.