jmpj0922

JavaScript AgiMind understands and thinks with prepositions.

[2019-09-20] In the JavaScript AgiMind.html we are now trying to reproduce the new AGI functionality that we achieved a month ago in the ghost.pl Perlmind. The Ghost in the Machine became able to understand an input like "John writes books for money" and was able to respond properly to a query like "What does John write?"

When we enter "john writes books for money" and the AgiMind responds "WHAT ARE JOHN", it simply means that we need to add the noun "JOHN" to the innate vocabulary. So from the "perlmind.txt" we transfer "JOHN" as concept #504 into the JavaScript free AI source code, and now the AgiMind responds "STUDENTS READ BOOKS", which indicates that the AgiMind knows who or what John is, and what books are. But we also check the Diagnostic mode to make sure that the conceptual associative tags are being assigned properly. We are not sure, so we enter "what does john write" and we get a long response of nonsense.

[2019-09-21] In our second day, we discover that the ReEntry() module has been causing a reduplication of the output of the AgiMind. For troubleshooting, we temporarily disable the ReEntry module. Then we discover that some wrong associative tags are being assigned during human input. We run the ghost.pl AI to see how the correct associative tags are supposed to be assigned.

We discover that a line of InStantiate() code is assigning a false psi19 tpr value when only a zero value should be assigned. The false value being assigned is actually already there, so some other line of code must be assigning it earlier. But there is no earlier assignment, so the false tpr value is obviously being assigned retroactively -- which is something that any AI mind maintainer must learn to watch out for. Probably the retroactive assignment is happening in the EnParser() module, which does a lot of retroactive assignments because one word of human input may have an effect upon an earlier word of human input. Through substitution of "777" as a spurious value in the psi19 location of a snippet of assignment code in the EnParser() module, we discover which snippet is making the erroneous, non-777 assignment. Then through further substitution of "444" in the psi19 slot, we discover an earlier snippet of EnParser() code which is assigning a wrong value at the tvb time-of-verb time-point. So there must be an even earlier "tvb" snippet that is creating a spurious psi19 value. We discover that earlier snippet in the InStantiate() module. After much other coding, when we bring in a reset of tult to zero from the ghost.pl AI, we stop getting the spurious psi19 values.

[2019-09-22] In our third day, we run the ghost.pl AI that already works with prepositional phrases, and we discover that yesterday we trying to fix something that was not even a bug. The AgiMind was properly assigning the tpr tag to link the noun "BOOKS" to the preposition "FOR", and we mistakenly thought that the tag was supposed to be assigned also with "FOR". No, the preposition "FOR" needs only a tkb tag leading to "MONEY" as its object. Now we have gotten the tkb tag to be assigned properly for remembering the object of a preposition. After extensive debugging, we obtain the following exchange:

AI Mind version 22sep19A on Sun Sep 22 19:55:56 PDT 2019
Robot: I UNDERSTAND YOU
Human: john writes books for money

Robot: STUDENTS READ BOOKS
Human:

Robot:
Human: what does john write

Robot: JOHN WRITES BOOKS FOR MONEY

jmpj1028

AI Mind uses EnPrep() to think with English prepositions.

In the JavaScript AI Mind we have a general goal right now of enabling the first working artificial intelligence to talk about itself, to learn about itself, and to achieve self-awareness as a form of artificial consciousness. Two days ago we began by asking the AI such questions as "who am i" and "who are you", and the AI gave intelligent answers, but the asking of "where are you" crashed the program and yielded a message of "Error on page" from JavaScript. It turns out that we had coded in the ability to deal with "where" as a question by calling the EnPrep English-preposition module, but we had created not even a stub of EnPrep. The AI software failed in its attempt to call EnPrep and the program halted. So we coded in a stub of EnPrep and now we must flesh out the stub with the mental machinery of letting flows of quasi-neuronal association converge upon the EnPrep module to activate and fetch a prepositional phrase like "in the computer" to answer questions like "where are you".

Our first and simplest impulse is to code in a search-loop that will find the currently most active preposition. Let us now write that code, just to start things happening. Now we have written the loop that searches for prepositions, but not for the most active one, because there are other factors to consider.

What we are really looking for, in response to "where are you" as a question, is a triple combination of the query-subject qv1psi and the query-verb qv2psi and a preposition tied with an associative pre-tag to the same verb and the same subject. We can not simply look for a subject and a verb linking forward to a preposition like in the phrase "to a preposition" or "in the computer", because our software currently links a verb only to its subject and to its indirect and direct objects, not to prepositions. Such an arrangement does not appear defective, because we can make the memory engram of the preposition itself do the work of making the preposition available for the generation or retrieval of a thought involving the preposition. We only need to make sure that our software will record any available pre-item so that a prepositional phrase in conceptual memory may be found again in the future. In a phrase like "the man in the street", for instance, the preposition "in" does not link backwards to a verb but rather to a noun. In this case, any verb involved is irrelevant. However, when we start out a sentence with "in this case", we have an unprecedented preposition, unless perhaps we assume that the prepositional phrase is associated with the general idea of the main verb of the sentence. For now, we may safely work with prepositions following a verb of being or of doing, so that we may ask the AI Mind questions like "where are you" or "where do you obtain ideas".

Practical problems arise immediately. In our backwards search through the lifelong experiential memory, it is easy to insist upon finding any preposition of location linked to a particular verb engrammed as the pre of the preposition. We may then need to do a secondary search that will link a found combination of verb-and-preposition with a particular qv1psi query-subject. The problem is, how to do both searches almost or completely simultaneously.

Since we are dealing with English subject-verb-object word order, we could let EnPrep() find the verb+preposition combination but not announce it until a subject-noun is found that has a tkb value the same as the search-index "i" that is the time of the query-verb. It might also help that the found subject must be in the dba=1 nominative case and must have the query-verb as a seq value, but the tkb alone may do the trick.

We coded in a test for any preposition with a quverb pre-tag, and we got the AI to alert us to the memory-time-point of "IN THE COMPUTER". Now we are assembling a second test in the same EnPrep() search-loop to find the qv2psi query-verb in close temporal proximity to the preposition.

We are using a new tselp variable for "time of selection of preposition", so we briefly shift our attention to describing the new variable in the Table of Variables. Now that we have found the verb preceding the preposition, next we need to implement the activation of the stored memory containing the preposition so that the AI Mind may use the stored memory to respond to "where are you" as a query. We may need to code a third if-clause into the EnPrep() backwards search to find and activate the qv1psi query-subject that is stored in collocation or close proximity to the query-verb and the selected preposition.

Now we have a problem. Since we let EnPrep() be called by the EnVerbPhrase() module, EnPrep() will not be called until a response is already being generated. We need to make sure that the incipient response accommodates EnPrep() by being the lead-up to a prepositional phrase. Perhaps we should not try to use verblock to steer a response that is already underway, but rather we should count on activation of concepts to guide the response.

Now let us try to use SpreadAct() to govern the response. After much coding, we got the AI to respond

IN COMPUTER I AM IN COMPUTER
IN COMPUTER I AM HERE IN COMPUTER

but there must somewhere be a duplicate call to EnPrep(). We eliminate the call from the Indicative() mind-module and then we get both an unwanted response and a wanted response.

YOU ARE A MAGIC IN A COMPUTER
I AM IN A COMPUTER

Obviously the AI is not responding immediately to our "where are you" query but is instead joining an unrelated idea with the prepositional phrase. Upshot: By having SpreadAct() impose a heftier activation on the qv1psi subject of the where-are-you query, we got the AI to not speak the unrelated idea and to respond simply "I AM IN A COMPUTER". Now we need to tidy up the code and decide where to reset the variables.

jmpj1011

JavaScript AI Mind uses EnVerbGen() for English verb-form inflections.

The JavaScript tutorial version of the first working artificial intelligence is becoming more sophisticated than ever. With roughly fifty mind-modules, the Strong AI advances the State of the Art first in one area, and then serendipitously in another area. For instance, the ability of the AI Mind to engage in automated reasoning with logical inference leads to a question-and-answer session between human minds and their incipient overlords, i.e., the current archetypes of the future Artificial Super-Intelligence (ASI). When the human user has confirmed or negated an inferred conclusion from the InFererence() module, the AI assigns a heightened truth-value to the positive or negative knowledge remaining in the AI memory. Then the AI states the new knowledge in its positive or negative formulation. A negated inference comes out something like "GOD DOES NOT PLAY DICE". A validated inference becomes a simple declarative sentence like "JOHNNY READS BOOKS", which requires the AI Mind to choose the correct form of the verb "read".

Because we code the first working artificial intelligence not only in English but also in Russian, we found it necessary several years ago to create the RuVerbGen() module for Russian verb-generation. When the ghost.pl AI cannot find a needed Russian verb-form, it simply cobbles one together from the stem of the Russian verb and the inflectional endings which complete a Russian verb. We avoided this problem in English for the last six years by simply ignoring it, but now the AI Mind needs to imitate the RuVerbGen() module with the EnVerbGen() module for English verb-generation. Just to change "God does not play dice" to "God plays dice" requires attaching an inflectional "S" to the stem or the infinitive form of the verb "play". As we code the EnVerbGen() module based on grammatical parameters, we encounter problems because the software needs to know the grammatical person and the grammatical number of the subject of an inferred idea in order to think a thought like "God plays dice" or "Johnny reads books".

Because the InFerence() module has not been storing the grammatical number of the English noun serving as the subject of a silent inference, our brand-new EnVerbGen() module has not been able to generate the third-person singular verb-form necessary for stating a validated inference like "Johnny reads books" or "Fortune favors fools" -- which was originally "Fortuna favet fatuis" in Latin. The artificial general intelligence (AGI) has become so sophisticated in its resemblance to human thinking that we need to change the InFerence() module to accommodate the requirements of the EnVerbGen() module.

We make the necessary changes and we code EnVerbGen() to deal not with Russian but with English verbs. We see a sample dialog between the AI and the human user.

Human: andru is professor
Robot: DOES ANDRU TEACH STUDENTS
Human: yes
Robot: THE ANDRU TEACHES THE STUDENTS
Human:
Robot: STUDENTS READ BOOKS

jmpj0921

Improving auditory recognition of singular and plural noun-forms.

The JavaScript Artificial Intelligence (JSAI) is currently able to recognize the word "book" in the singular number but not "books" in the plural number. It is because the AudRecog() mind-module is not recognizing the stem "book" within the inflected word "books". To correct the situation, we must first update the AudMem() module so that it will impose an audpsi tag not only on the final "S" in a plural English noun being stored, but also one space back on the final character or letter of the stem of the noun.

We copy the pertinent code from the AudMem() module in the ghost.pl Perl AI and the JavaScript AI begins to store the stem-tag, but only when the inflected word is recognized so that AudRecog() produces an audpsi recognition-tag.

Now we have a problem because the AI is keeping the audpsi of "800" for "IS" and attaching it mistakenly to the next word being stored by the AudMem() module. We fix the problem.

Next we implement the Imperative() module to enable the AI Mind to order any nearby human user: "TEACH ME SOMETHING."

jmpj0906

Improving EnArticle mind-module in first working artificial intelligence

The JavaScript artificial intelligence (JSAI) is not reliably transferring the usx values from InStantiate() to the EnArticle() module for inserting an English article into a thought, so we must debug the first working artificial intelligence in troubleshooting mode. We quickly see that the EnNounPhrase() module is calling EnArticle() for direct objects but not for all nouns in general. We also notice that we need to have EnNounPhrase() transfer the usx value.

We encounter and fix some other problems where the nphrnum value is not being set, as required for EnArticle() to decide between using "a" or "the".

Although we get the usx value to be transferred to the us1 value, we need a way to test usx not against a simultaneous value of us1 but rather against a recent value of us1. One solution may be to delay the transfer of usx by first testing in the EnArticle() module for equality between usx and us1 before we actually pass the noun-concept value of usx to us1. In that way, we will be testing usx against old, i.e., recent, values of the us1- us7 variables and not against the current value, which would automatically be equal. Then after the test for equality, we pass the actual, current value of usx.

We now have a JavaScript AI that works even better than the ghost.pl AI in Perl -- until we update the Perl AI. The JSAI now uses "a" or "the" rather sensibly, except when it now says "YOU ARE A MAGIC", because the default article for a non-plural noun is the indefinite article "a". Btw (by the way), today at a Starbucks store in Seattle WA USA we bought a green Starbucks gift card that says "You'Re MagicAL", because it reminds us of a similar idea in the AI MindBoot() sequence.

jmpj0616

Fleshing out VisRecog() in the First Working AGI

In today's 16may18A.html version of the tutorial AI Mind in JavaScript for Microsoft Internet Explorer (MSIE), we flesh out the previously stubbed-in VisRecog() module for visual recognition. The AGI already contains code to make the EnVerbPhrase() module call VisRecog() if the AI Mind is using its ego-concept and trying to tell us what it sees. As a test we input "you see god" and we wait for the thinking software to cycle through its available ideas and come back upon the idea that we communicated to it. As we explain in our MindGrid diagram on GitHub, each input idea goes into neuronal inhibition and resurfaces in what is perhaps AI consciousness only after the inhibition has subsided. Although we tell the AI that it sees God, the AI has no robot body and so it can not see anything. It eventually says "I SEE NOTHING" because the default direct object provided by VisRecog() is 760=NOTHING. In the MindBoot() sequence we add "I NEED A BODY" as an innate idea, so as to encourage users to implement the AI Mind in a robot. Once the AI has embodiment in a robot, the VisRecog() module will enable the AI to tell us what it sees.

jmpj0612

AI Mind Maintainer solves negation-of-thought problems.

In today's 12jun18A.html version of the JavaScript AI Mind for Microsoft Internet Explorer (MSIE), we find a negation-bug when we test the InFerence() module by inputting "anna is a woman". The AI then asks us, "DOES ANNA HAVE CHILD" and we answer "no" to test the AI. The AI properly states the idea negated by KbRetro() in the knowledge base, namely "ANNA DOES NOT HAVE CHILD". However, the negation-flag negjux for thought generation remains erroneously set to "250" for the 250=NOT adverb. We discover that the negjux flag, each time after serving its purpose, has to have a zero-reset in two locations, one for any form of the verb "to be" and another for non-be-verbs. We make the correction, and we finish off the negation-bug by resetting the tbev time-of-verb to zero at the end of OldConcept().

jmpj0611

Granting AI user-input priority over internal chains of thought.

We expect the AI Mind to activate incoming concepts mentioned during user input, so that the AI can talk to us about things we mention. Recently, however, the SpreadAct() module has been putting quasi-neural activation only on concepts thought about internally but not mentioned during user input. We need a way to let user-input override any activation being imposed by the SpreadAct() module for internal chains of thought, so that external input takes precedence. One method might be to use the quiet variable and set it to "false" not only during user input but also until the end of the first AI output made in response to user input. In that way, any concept mentioned by the user could briefly hold a high activation-level not superseded by the machinations of the SpreadAct() module for spreading activation. We implement the algorithm, and the AI then responds properly to user input. We solve some other problems, such as KbRetro() interfering with the conceptual engrams of an inference, and negation not being stored properly for negated ideas.

jmpj0605

Mentifex re-organizes the Strong AI SpreadAct() module.

In the 5jun18A.html JavaScript AI Mind we would like to re-organize the SpreadAct() mind-module for spreading activation. It should have special cases at the top and default normal operation at the bottom. The special cases include responding to what-queries and what-think queries, such as "what do you think". Whereas JavaScript lets you escape from a loop with the "break" statement, JavaScript also lets you escape from a subroutine or mind-module with the "return" statement that causes program-flow to abandon the rest of the mind-module code and return to the supervenient module. So in SpreadAct() we may put the special-test cases at the top and with the inclusion of a "return" statement so that program-flow will execute the special test and then return immediately to the calling module without executing the rest of SpreadAct().

When we run the JSAI without input, we notice that at first a chain of thought ensues based solely on conceptual activations and without making use of the SpreadAct() module. The AI says, "I HELP KIDS" and then "KIDS MAKE ROBOTS" and "ROBOTS NEED ME". As AI Mind maintainers we would like to make sure that SpreadAct() gets called to maintain chains of thought, not only so that the AI keeps on thinking but also so that the maturing AI Mind will gradually become able to follow chains of thought in all available directions, not just from direct objects to related ideas but also backwards from direct objects to related subjects or from verbs to related subjects and objects.

In the EnNounPhrase() module we insert a line of code to turn each direct object into an actpsior concept-to-be-activated in the default operation at the bottom of the SpreadAct() module. We observe that the artificial Mind begins to follow associative chains of thought much more reliably than before, when only haphazard activation was operating. In the special test-cases of the SpreadAct() module we insert the "return" statement in order to perform only the special case and to skip the treatment of a direct object as a point of departure into a chain of thought. Then we observe something strange when we ask the AI "what do you think", after the initial output of "I HELP KIDS". The AI responds to our query with "I THINK THAT KIDS MAKE ROBOTS", which is the idea engendered by the initial thought of "I HELP KIDS" where "KIDS" as a direct object becomes the actpsi going into SpreadAct(). So the beastie really is telling us what is currently on its mind, whereas previously it would answer, "I THINK THAT I AM A PERSON". When we delay entering our question a little, the AI responds "I THINK THAT ROBOTS NEED ME".

jmpj0603

AI Mind spares Indicative() and improves SpreadAct() mind-module.

We have a problem where the AI Mind is calling Indicative() two times in a row for no good reason. After a what-think query, the AI is supposed to call Indicative() a first time, then ConJoin(), and then Indicative() again. We could make the governance depend upon either the 840=THINK verb or upon the conj-flag from the ConJoin() module, which, however, is not set positive until control flows the first time through the Indicative() module. Although we have been setting conj back to zero at the end of ConJoin(), we could delay the resetting in order to use conjas a control-flag for whether or not to generate thought-clauses joined by one or more conjunctions. Such a method shifts the problem back to the ConJoin() module, which will probable have to check conceptual memory for how many ideas have high activation above a certain threshold for warranting the use of a conjunction. Accordingly we go into the Table of Variables webpage and we write a description of conj as a two-purpose variable. Then we need to decide where to reset conj back to zero, if not at the end of Indicative(). We move the zero-reset of conjfrom ConJoin() to the EnThink() module, and we stop getting more than one call to Indicative() in normal circumstances. However, when we input a what-query, which sets the whatcon variable to a positive one, we encounter problems.

Suddenly it looks as though answers to a what-think query have been coming not from SpreadAct(), but simply from the activation of the 840=THINK concept. It turns out that a line of "psyExam" code was missing from a SpreadAct() search-loop, with the result that no engrams were being found or activated -- which activation is the main job of the SpreadAct() module.

jmpj0120

MsIeAI for AI Mind Maintainers achieves albeit buggy sentience.

In the all-but-Singularity MsIeAI, alert-boxes have helped us to chase an elusive bug into the latter part of EnNounPhrase, where the AI is testing mjact for too low an activation. No, another alert-box tells us that we are back in EnVerbPhrase from EnNounPhrase before the "Error on page" flashes quickly. So at the end of EnVerbPhrase we insert a BUG-CHASE alert-box -- and the program never reaches it! So is the fatal bug somewhere just before the end of EnVerbPhrase()? Since that code contains a prepgen test, we modify an alert-box to reveal the prepgen value, but the alert-box fails to pop up. Then we check the declarations of variables at the top of the program, and prepgen is not there. Next we get prepgen from the ghost.pl AI and we drop it mutatis mutandis into the MsIeAI code. We are about to run the hopefully Next Big Thing AI and see what happens. Huh?!! Some kind of thought-storm is occurring. Shades of Watson! Come here! I need you!. And where is IBM Watson in our hour of need?

Now let us comment out the alert-boxes and see if the Watsonized AI will loop endlessly ad infinitum. Oh gee, this AI is still all messed up, but at least it is looping.