Strong AI

Purpose: Discussion of Strong AI Minds thinking in English, German or Russian. Privacy policy: Third parties advertising here may place and read cookies on your browser; and may use web beacons to collect information as a result of ads displayed here.

Sunday, April 30, 2017

pmpj0430

Improving the storage of the number-flag for nouns.

Today in ghost199.pl we will try to make the AI error-free even before we go back to adding in the functionality already present in some of our obsolete AI Minds. For instance, we have not yet coded the negation of verbs into our Perlmind source code. Consequently, if you tell the AI something like "You are not a boy", it fails to attach a negative juxtaposition $jux flag to the verb during comprehension of the input sentence. A few cycles of thought later, the AI may then assert "I AM A BOY" because it has been informed of the negated proposition without the ability to process the negation.

We debug the AI by letting ghost199.pl think on its own without human input. Eventually the Perlmind erroneously says "I AM ROBOTS", which is grammatically incorrect because of the plural noun. We intuit immediately that the AI is retrieving the most recent engram of the concept #571 "ROBOT" without insisting on a singular number. We inspect other recent thoughts of the AI and we see that it thinks "KIDS MAKE ROBOTS" but it stores the word "ROBOTS" as a singular noun. We must look and see if the InStantiate() mind-module has a proper $num flag for storing "ROBOTS" correctly as a plural noun. We see that the OldConcept() module looks up the stored num(ber) of a found engram and tentatively assigns the same value to the $num flag, but there really needs to be an override if a different value is needed.

In the otherwise obsolete but still rather advanced 24jul14A.F version of MindForth, some AudInput code checks for an "S" at the end of an input noun as a reason to assign plural number to the noun. Let us try to implement the same test in the Perl AI. First we test for the presence of an 83=S, but we must also make sure that the "S" is the final character of a noun. First in OldConcept() we comment out the line of code that was transferring the found num(ber) of a noun to be the same number for a new instance of the noun, regardless of the presence or absence of a terminating "S". Then we notice that "ROBOTS" stops being stored as singular, and becomes plural. We create a variable $endpho to hold onto each previous character in AudInput() to test if a word ends in 83=S. Thus we are able to store a plural number if a noun ends in "S".

Sunday, April 23, 2017

pmpj0423

Stubbing in MindMeld() and stopping derailment of thought.

We function now as an AI Mind Maintainer debugging the Perlmind free AI source code. In the ghost198.pl AI we first stub in the audacious MindMeld() module to nudge AI practitioners into devising a way for two AI Minds to share their dreams. Then we deal with some problems pointed out on Usenet by persons who have downloaded the Perlmind and evaluated its functionality.

We run ghost198.pl with "dogs are mammals" as input and we press the Escape-key to halt the AI after its first response, "I HELP KIDS". We notice immediately three problems with how the word "DOGS" is stored in the @psy and @ear memory arrays. For some reason, "DOGS" is being assigned new-concept #3002, even though the Tutorial display of diagnostic messages indicates that the AI is preparing to assign new-concept #3001 to the first new concept. We check the MindBoot() sequence to make sure that "DOG" is not already a known concept in the AI; it is not. Now let us inspect the source code to see where the new-concept number $nxt is incremented from 3001 to 3002. We see that the end of MindBoot() clearly assigns the number 3001 as the value of the $nxt variable. Now let us search for the $nxt++ increment. It is happening towards the end of the NewConcept() module. We immediately wonder if $nxt is being incremented before AudMem() stores the concept-number. We insert into AudMem() a diagnostic message to let us know the $nxt value before storage. The first diagnostic message does not tell us enough, so we insert a second diagnostic into the AudMem() module. It also does not help us.

In the AudInput() module we use some diagnostic messages to learn that the "S" in "DOGS" is first being stored with the correct $nxt value of "3001" and then a second time with the incorrect value of "3002". Perhaps we should increment $nxt not in NewConcept() but in AudInput(). We move the $nxt++ increment from NewConcept() into AudInput(), and we stop getting the wrong values of the $nxt variable.

A second problem is that the concept of "DOGS" is being stored with a zero instead of "2" for "plural" in the $num slot of the @psy conceptual flag-panel. The most recent incarnation of the InStantiate() module does not seem to address the $num value sufficiently, so let us inspect recent or older MindForth code. We discover that the obsolete 24jul14A.F version of MindForth uses some complex tricks to assign the num(ber) of a concept being stored, so we will put aside this problem to deal with more serious issues.

The third and presumably more serious problem is that the input word "DOGS" is being stored with the $nxt concept number "3001" only on the "S" phoneme and not on the "G" at the end of the word-stem "DOG". Let us leave that problem also aside for a while, because entering "dogs are mammals" repeatedly is running into more serious problems. FOr instance, all three words of the input are being stored erroneously with the same $rv recall-vector, which can cause the wrong auditory memories to be retrieved. Let us see if the previous ghost197.pl does the same error. Yes, and so does the ghost196.pl AI. However, we should not find it difficult to correct the $rv problem. We fix the problem by resetting $rv to zero at the end of the InStantiate() module. Now the Perlmind no longer goes off the rails of thought, and so we upload it to the Web.

Wednesday, April 12, 2017

pmpj0412

Ghost Perl Strong AI cycles through Normal; Transcript; Tutorial; Diagnostic Mode

It is time now in ghost197.pl to show a clean human-computer interface (HCI) and to stop displaying masses of diagnostic messages. Accordingly in the AudInput module we change the user-prompt to say "Tab cycles mode; Esc(ape) quits AI born [date-of-birth]". We insert if-clauses to declare which user input mode is in effect: Normal; Transcript; Tutorial; or Diagnostic. Near the start of ghost197.pl we set the $fyi to a default starting value of unity ("1") so that the human user or Mind-maintainer may press the Tab-key to cycle among user input modes. In AudInput() we insert code to increment $fyi by one point with each press of the Tab-key and to cycle back to unity ("1") for Normal Mode if the user in Diagnostic Mode presses Tab again.

In the MainLoop module we change a line of code to test for $fyi being above a value of two ("2") and, if so, to display the contents of the @psy conceptual array and of the @ear auditory memory array. Thus the user in #3 Tutorial Mode or in #4 Diagnostic Mode will see the storage of current input and current output in the memory arrays. We consider the display of conceptual memory data in Tutorial Mode to be an extremely powerful tool for teaching how the artificial general intelligence (AGI) works. After any input, the user may see immediately how the input goes into memory and how the values in the flag-panel of each row of the @psy array represent the associative tags from concept to concept and from engram to engram.

Next we start commenting out or deleting the display of various diagnostic messages. Over time and over multiple releases of the Ghost AI source code, any AI coder may decide which messages to display in both Tutorial and Diagnostic Modes, or in only one of them. Although we comment out a message involving Russian input, we do not delete the diagnostic message because we may need it when we turn back on Russian as an input language. Russian has become much more important in our Ghost Perl AI because we need Russian or German to demonstrate Machine Translation by Artificial Intelligence. When we have commented out most of the diagnostic messages, we need to put back in some code to show what the user is entering.

Tuesday, April 11, 2017

pmpj0411

Stubbing in the MetEmPsychosis module.

[2017-04-10] Today in ghost195.pl we stub in MetEmPsychosis() as an area for Perl code that will enable an AI Perlmind to either move itself across the Web or replicate itself across the Web. We foresee the advent of a kind of "AiBnb" or community of Web domains that invite and encourage AI Minds to take up temporary or long-term residence, with local embodiment in a robot and with opportunities for local employment as a specialized AI speaking the local language and familiar with the local history and customs.

[2017-04-10] In the AudInput() module today we insert the Cyrillic characters of the Russian alphabet for each line of code that converts lower case to upper case and sets the $hlc variable to "ru" as the human-language-code for Russian. We have not yet turned the Russian language back on again, but we will need it to test out our ideas for Machine Translation by Artificial Intelligence.

Coding VisRecog to say by default: I SEE NOTHING.

[2017-04-11] Today in ghost196.pl we would like to port in from MindForth the code that causes any statement of what the AI is seeing to default to the direct object "NOTHING," so that Perl coders and roboticists may work on integrating computer vision with the AI Mind. We make it clear that the visual recognition (VisRecog) system needs only to supply the English or Russian name of what it is seeing, and the AI will fill the slot for direct objects while generating a sentence about what the AI sees. The VisRecog mechanism does not need to be coded in Perl or in Forth. It only needs to communicate to the Perlmind a noun that names what the AI is seeing. When the generated statement passes through reentry back into the Mind, even a new noun will be assigned a concept-number and will enter into the knowledge-base (KB) of the AI.

First we declare the subject-verb-object variables $svo1, $svo2, $svo3, and $svo4 to hold a value that identifies a concept playing the role of subject, or verb, or indirect object, or direct object in a typical sentence being generated by the AI. If there is no direct object filling the slot for the object of the verb "SEE", then the VisRecog() module must try to fill the empty slot. Until a Perl expert fleshes out the VisRecog() code, the word "NOTHING" must remain the default object of the verb "SEE" when the ego-concept of "I" is the subject of the verb. We ran the AI and we typed in "you see kids." After a spate of outputs, the AI said, "I SEE KIDS," but we would really prefer for the AI to say, "I SEE NOTHING" as a default.

After coding a primitive VisRecog() module, next we go into the part of the EnVerbPhrase() module where it is looking for a direct object. We set conditions so that if the subject is "I" and the verb is "SEE", VisRecog() is called to say "NOTHING" as a direct object, and EnVerbPhrase() stops short of saying any other direct object by doing a "return" to the calling module. We now have a Perlmind that invites the integration of a seeing camera with the AI software.