Gabriela goes Brainhat, ChatGPT and Back

This BLOG entry discusses interaction with ChatGPT:

  1. I had Brainhat compile some simple English input into knowledge data structures (CCs),
  2. I had Brainhat translate the knowledge back into English and then gave that to ChatGPT to summarize,
  3. I gave ChatGPT's own summary back to ChatGPT and asked it to translate it into very simple English--one idea per line,
  4. I fed that back into Brainhat.

The purpose of the exercise was to demonstrate that Brainhat could talk to ChatGPT and ChatGPT could talk to Brainhat with little loss of information.

"Gabriela" is a data set derived from an English writing exercise by a young woman in Tapachula, Mexico, editted so that it contains one idea per line: 

My name is Gabriela.
I am a woman.
I am 17 years old.
I am from Mexico.
I am from a city.
The city's name is Tapachula.
The city is always hot.
We have a beach.
I have a beautiful family.
I have one brother.
I have one sister.
My father's name is Jorge.
My mother's name is Irma.
My sister's name is Irma.
My brother's name is Jorge.
He is married.
His wife's name is Alejandra.
She will have a baby in November.
The family is so excited.
I have a beautiful dog.
He is a Siberian Husky.
He is my baby.
He is very lovely.
I love my dog.
   .
   .
   .
I plan to work.
I plan to earn money.
I plan to go to Hawaii.
I plan to go to Europe.
I plan to return to mexico.
I plan to marry the correct person.
I plan to have three children.
I plan to make a wonderful family.

Brainhat reads the "Gabriela" data set and builds a context. The context is made of data structures that represent the knowledge from the original English input in reverse order.  Brainhat can translate the elements of the context back into simple English. This is how the first part of the context appears when translated: 

You plan You do make your pretty family.  You plan You do marry You. You
plan You do go to europe. You plan You do go to hawaii. You plan You
do make money. You plan You do work....
Brainhat shares with ChatGPT

I gave the first 20 elements of the context English translation to ChatGPT along with instructions to summarize:

Summarize the following, taking the sentences in reverse order: You
plan You do make your pretty family.  You plan You do marry You. You
plan You do go to europe. You plan You do go to hawaii. You plan You
do make money. You plan You do work. You plan You do finish your
major. You plan You do study your major. You plan You do return to
mexico. You plan You do learns english. You have You do plan. You like
your pretty dog belonging to You. You like a college will be perfect
for You. You may be with hugo. You may be with your pretty family. You
believe a college will be perfect for You. a college is described. You
are from tapachula. You want You to study your major.

This is what ChatGPT gave back:

Here’s a clear, simple summary of your passage — with the ideas
arranged in **reverse order** (starting from the last sentence and
moving backward):

You want to study your major and come from Tapachula.  You believe
college will be perfect for you and that you may be with Hugo and your
pretty family.  You like your pretty dog and think college will suit
you well.  You have plans — to learn English, return to Mexico, study
and finish your major, work, make money, travel to Hawaii and Europe,
get married, and build a beautiful family.

At this point, Brainhat has shared its context with ChatGPT and ChatGPT has eloquently described it. Next, I ask ChatGPT to translate its own summary into simple English that Brainhat would understand directly:

Using only words that you find in the following text, convert the text
into very simple sentences with one idea expressed per sentence: You
want to study your major and come from Tapachula.  You believe college
will be perfect for you and that you may be with Hugo and your pretty
family.  You like your pretty dog and think college will suit you
well.  You have plans — to learn English, return to Mexico, study and
finish your major, work, make money, travel to Hawaii and Europe, get
married, and build a beautiful family.

ChatGPT produced simple output--one idea per line, as requested:

Here is the text rewritten in **very simple sentences**, each with one clear idea, using **only the words from the original text**:

1. You come from Tapachula.
2. You want to study your major.
3. You believe college will be perfect for you.
4. You may be with Hugo.
5. You may be with your pretty family.
6. You like your pretty dog.
7. College will suit you well.
8. You have plans.
9. You learn English.
10. You return to Mexico.
11. You study your major.
12. You finish your major.
13. You work.
14. You make money.
15. You travel to Hawaii.
16. You travel to Europe.
17. You get married.
18. You build a beautiful family.
ChatGPT shares with Brainhat

I removed the statement numbering and fed this text from ChatGPT back into Brainhat:

You come from Tapachula.
You want to study your major.
You believe college will be perfect for you.
* You may be with Hugo.
* You may be with your pretty family.
You like your pretty dog.
College will suit you well.
You have plans.
You learn English.
You return to Mexico.
You study your major.
You finish your major.
You work.
You make money.
You travel to Hawaii.
You travel to Europe.
You get married.
You build a beautiful family.

This is how the context looks when Brainhat reads back the compiled context derived from the ChatGPT input:

I make a pretty family. I get married. I travel to europe. I travel to
hawaii. I make money. I work. I finish my major. I study my major. I
return to mexico. I learn english. I have I do plan. I like my pretty
dog. I believe college will be perfect for I. I want I to study my
major. I come from thing.

There were two issues: the first is that Brainhat treats statements in future imperfect tenses as opportunities to ask questions. So, "you may be with your pretty family" causes Brainhat to ask "am I with my pretty family?" This functionality will be made optional. The statements are marked above with '*'.

The second issue is that Brainhat records "I come from Tapachula" as "I come from thing". A fresh copy of Brainhat does not know 'Tapachula.' It was originally introduced in another part of the "Gabirela" data set like so:

I am from a city.
The city's name is Tapachula.

Since Brainhat did know about 'Tapachula', it substituted 'thing.'  Note that if 'Tapachula' were part of the start-up vocbulary, as is 'Mexico', naming the city would not be necessary.

Summary

Brainhat and ChatGPT can interact.  

Brainhat already has text interfaces so that it can use third-party software to help with language simplification or act as a resource for answers. Particularly, Brainhat interfaced with an AIML processor called Pandora some years back.  These interfaces can be used for LLMs, too, and will appear in a copy of Brainhat soon.

Powered by Publii