clpm0709

Extrapolating from the First Working AGI

Artificial General Intelligence (AGI) has arrived in MindForth and its JavaScript and Perl symbionts. Each Mind is expanding slowly from its core AGI functionality. The MindBoot sequence of innate concepts and ideas can be extended by the machine learning of new words or by the inclusion of more vocabulary in the MindBoot itself.

We may extrapolate from the current MindBoot by imagining a Perlmind that knows innately the entire Oxford English Dictionary (OED) and all of WordNet and all of Wikipedia. Such an AGI could be well on its way to artificial superintelligence.

If there is no upper bound on what a First Working AGI may know innately, why not make full use of Unicode and embed innately the vocabulary of all living human languages? Then go a step further and incorporate (incerebrate?) all the extinct languages of humanity, from LinearB to ancient Egyptian to a resurrected Proto-European. Add in Esperanto and Klingon and Lojban.

pmpj0706

Preventing unwarranted negation in the First Working AGI.

The First Working AGI (Artificial General Intelligence) has a problem in the ghost267.pl version written in Perl Five. After we trigger a logical inference by entering "anna is woman" and answering the question "DOES ANNA HAVE CHILD" with "no", the Perlmind properly adjusts the knowledge base (KB) and states the confirmed knowledge as "ANNA DOES NOT HAVE CHILD". Apparently the reentry of concept 502=ANNA back into the experiential memory is letting the InStantiate() module put too much activation on the 502=ANNA concept and the AI is erroneously outputting "ANNA BE NOT WOMAN". Since the original idea was "anna is woman", the real defect in the software is not so much the selection of the old idea but rather its unwarranted negation. When we change some code in the InStantiate() module to put a lower activation on reentrant concepts, the problem seemingly goes away, because the AI says "I HELP KIDS" instead of "ANNA BE NOT WOMAN", but as AI Mind Maintainers we need to track down where the unwarranted negation comes from.

The unwarranted negation comes from the OldConcept() module where the time-of-be-verb $tbev flag was being set for an 800=BE verb and was then accidentally carrying over its value as the improper place for inserting a 500=NOT $jux flag into an idea subsequently selected as a remembered thought. When we zero out $tbev at the end of OldConcept(), the Ghost AI stops negating the wrong memory.

pmpj0705

Improving the storage of conceptual flag-panels during input.

In the ghost266.pl Perlmind we need to improve upon a quick-and-dirty bugfix from our last coding session. After a silent inference and the operation of AskUser() calling EnAuxVerb(), the Ghost AI was going into the verb-concept of the inference-triggering input and replacing a correct $tkb value with a zero. Apparently the time-of-verb $tvb value, set in the Enparser() module during the parsing of a verb, was being erroneously carried over from the verb of user-input to the verb 830=DO in the EnAuxVerb() module during the generation of an inference-confirming question by the AskUser() module. Therefore the time-of-verb $tvb-flag needs to be reset to zero not during the generation of a response to user-input but rather at the end of the user-input. However, we find that we may not reset time-of-verb $tvb to zero during AudInput(), apparently because only character-recognition and not yet word-recognition has taken place. The $tvb-setting for a verb must remain valid throughout AudInput() so that the EnParser() module may use the time-of-verb $tvb flag to store a direct object as the $tkb of a verb. Accordingly we reset the $tvb-flag to zero in the Sensorium() module after the call to AudInput(). We stop seeing a $tkb of zero on the verb of an input that triggers automated reasoning with logical InFerence.

pmpj0703

Keeping AskUser from storing incorrect associative tags.

The ghost265.pl version of the Perlmind has a problem after making a logical inference. Instead of getting back to normal thinking, some glitch is causing the AI to say "ANNA BE NOT ANNA".

As we troubleshoot, we notice a problem with the initial, inference-evoking input of "anna is woman". The be-verb is being stored in the psy array with a $tkb of zero instead of the required time-point of where the concept of "WOMAN" is stored. This lack of a $nounlock causes problems later on, which do not warrant their own diagnosis because they are a result of the lacking $nounlock. We need to inspect the code for where the be-verb is being stored in the psy-array, but we are not sure whether the storage is occuring in the InStantiate() module, or in OldConcept(), or in EnParser(). We see that the Ghost AI is trying to store the be-verb in the EnParser() module with the correct tkb, but afterwards a tkb, of zero is showing up. We must check whether InStantiate() is changing what was stored in the EnParser() module.

Meanwhile we notice something strange. An input of "anna is person" gets stored properly with a correct $tkb, but "anna is woman" -- causing an inference -- is stored with a $tkb, of zero. When we enter "anna is robot", causing an inference and the output "DOES ANNA WANT BEEP", there is also a zero $tkb. Upshot: It turns out that the EnAuxVerb() module, called by AskUser() after an inference, was setting a wrong, carried-over value on the time-of-verb $tvb variable, which was then causing InStantiate() to go back to the wrong time-of-verb and set a zero value on the $tkb flag. So we zero out $tvb at the start of EnAuxVerb().

pmpj0701

Debugging the InFerence Function in the ghost.pl First Working AGI.

The ghost264.pl Perlmind has a minor bug which causes logical inference not to work if the inference is not triggered immediately at the start of running the program. If we let the AI run a little and then we type in "anna is woman", the AI answers "DOES ERROR HAVE CHILD" instead of "DOES ANNA HAVE CHILD". In the psy concept array of the silent inference, we observe that a zero is being recorded instead of the concept number "502" for Anna. The AI MindBoot is designed with the concept of "ERROR" placed at the beginning of the boot sequence so that any fruitless search for a concept will result automatically in an "ERROR" message if no concept is found. We suspect that some variable in the InFerence module is not being loaded with the correct value when the ghost.pl AI has already started thinking various thoughts.

The pertinent item in the InFerence() module is the $subjnom or "subject nominative" variable which is set outside of the module before InFerence is even called. We discover that the variable is spelled wrong in the OldConcept module, and we correct the spelling. It then seems that InFerence() can be called at any time and still operate properly. We decide to run the JavaScript AI to see if an inference has any problems if it is not the first order of business at the outset of an AI session. Nothing goes wrong, so the problem must have been the misspelling in the OldConcept() module.

During this coding session we also make a change in the KbRetro() module for the retroactive adjustment of the knowledge base (KB). We insert some code to put an arbitrary value of eight (8) on the $tru(th)-value variable for the noun at the start of the silent inference, such as "ANNA" in the silent inference "ANNA HAVE CHILD". When the human user either confirms or invalidates the inference, the resulting knowledge ought to have a positive truth-value, because someone has vouched for the truth or the negation of the inferred idea. We envision that the $tru(th)-value will serve the purpose of letting an AI Mind restrict its thinking to ideas which it believes and not to mere assertions or to ideas which were true yesterday but not today. We expect the $tru(th)-value to become fully operative in a robotic AI Mind for which "Seeing is believing" when visual recognition from cameras serving as eyes provides reliable knowledge to which a high $tru(th)-value may be assigned.