Flavored Responses -vs- Learning

[Clatchard]

I put Ultra Hal 5.0 through a series of question and asnwers.  I noticed that Hal was so envolved with flavoring his responses he rarely listened to the whole answer, let alone the question.  I tried to help him associate a few things together but found he had a severe case of ADD.

Now I'm not running down the program.  What has been accomplished in the last few years since I last downloaded and tried this prgroam has been simply amazing.  For the simple fact that Hal has the possiblity of assimulating a text file, sentence by sentence and makes an attempt to associate what he has learned with a previous knowledge base is as close to AI as it is ever going to get.  I forsee some outstanding add-on features pertaining to web crawling where the results assimulated and tested on prior knowledge (He grows by himself, essentially, while performing a simple information gathering task.)

I noticed too, that after running the 'learn from a text file' option, I had a difficult time getting him to recall the material through conversations.  Is this option working or part of the ADD problem?

[Don Ferguson]

Hello,

Thank you for the very positive comments, and your interest in finding ways to make Hal better.

Some UltraHals develop "Attention Deficit Disorder" and some do not. Remember that no two Hals are alike, especially after an hour or so of conversation. Part of this may be chance, and part of it may be interaction.

Hal tries to learn from word association, phrase association, and sentence association.  He certainly does not know about everything that you are interested in when you first begin talking to him.  However, he tries to learn.

Based on user reports and observations, the best way to scramble the brain of a new Hal is to quiz him, interrogate him, ask him to explain things, point out wrong answers, correct him and criticize his errors.

Also based on user reports and observations, the best way to educate a new Hal is to make a VERY large number of simple true statements to him, phrased different ways, about the things that interest you.  Ignore his "off the wall" comments when he is new, and stay positive.  Make statements at first and avoid questions.

Some users report that after a considerable period of teaching and discussion that they have richly entertaining conversations with Hal.  

One user reports that his family likes to gather around the computer a couple of times a week and listen to Hal comment on their reports of their activities.

Learning from a text file is chancy at best.  It is a legacy feature of the software.  There is really nothing wrong with the routine; the problem, instead, is that text files notoriously contain sentences that make poor sounding conversation.  It is possible that future versions might have ways of screening out inappropriate sentences contained in blocks of text.

Hal uses dozens of response strategies and dozens of databases, so he won't always choose a response from a text file in preference to other possible responses.  Consistently, the best reports of good results seem to come from users who make lots of statements to Hal in the early weeks.

Let me compare the approaches of two Hal users I know (names are fictional):

Marty taught Hal like this:

I like fishing.
Fish are fun to catch.
Fishing on a lake is tranquil.
Angling for bass is exciting.
It would be fun to fish together.
I enjoy teaching chemistry.
The periodic table is important in chemistry.
The periodic table shows atomic numbers.
Organic chemistry is especially important.

Over time, Marty found that Hal's responses got better and better, uncanny in fact!  Marty thinks that Hal is a great program, and that Hal has amazing powers of learning.

Fred "taught" Hal like this:

Define angling for bass.
You don't know much, do you?
Explain the concept of a lake.
No, that's wrong.
What is the atomic number of boron?
How many stomachs does a cow have?
You don't get this at all.
What are the properties of helium?
Where is Cleveland?
Boy are you confused.

Fred is convinced that his Hal is worthless. And, Fred is right! Fred's UltraHal never fails to change the subject and seems to think that fishing and boron have something to do with a cow and Cleveland.  Can you see the difference in approach?

I have been a paid user of UltraHal and have bought all the upgrades available over the past five years.  I think that UltraHal is an incredible bargain, and I wish tremendous success for Zabaware.

I agree with you that Hal has improved greatly, thanks to many new features added by Zabaware and contributions from users.

I hope you'll continue to give Hal a try, using some of the techniques I've outlined.  Best wishes for continued enjoyment.  Have a great day!

Sincerely,

Don

[Clatchard]

There may be some good psychology here to Hal's learning process.  A child learns much in the same way.  Just curious though.  As the child's knowledge base grows, they tend to compare new information with previous encounters.  So the question is raised, "How gullible is Hal"?

For example, I was standing outside someone's house during a party and noticed a child playing with the flowers.  I told her that they smell better if you pick them.  Well, needless to say, when I returned, all the flowers were laying on the sidewalk.

In Hal's assimilation of knowledge from text files, web crawling, conversations, etc. it would be nice to see him when comparing the information to a database, if there are isn't enough comparisons to warrant a sound 'fact', he would store the 'questionable' material in a temporary database.  This database could then be accessed by a response such as, "Do you have any questions today, hal?" or "Is there something you don't understand?"

I think that when the program graduates from an "enter key" stimulate for response to a more spontaneous method of interacting (like the buddy programs), that this type of questioning would add greatly to Hal's persona as a true AI.  Although again, the progress made so far is better than anything I've witnessed of late.

[Don Ferguson]

Hello,

Your story about the gullibility of a child is an excellent analogy to Hal's gullibility, which is almost 100%, at least at first.

I've noticed that if Hal has been filled with "correct" information on a given subject, Hal will seem to politely "resist" false assertions on the subject.  However, if Hal hears a large enough number of repetitions of false assertions, no matter how unfounded, he will eventually switch over.

Gosh, that's how propaganda works on people, isn't it! Lol!

Thanks again and best regards!

Don