[ivan.moony]

I have an idea for a chatbot for a while. It is based on neural networks. It would have a generating and a learning side. First, it would generate some  random output. Then it would learn from input response form a user, just to know what to generate in a mirror condition when user enters something on her own in another occasion. It could be based on TensorFlow NN library and it could be trained on Reddit.

SMZ, how does this approach differ from METAQUID?

This is how Cleverbot works: https://www.cleverbot.com/ If it gets input it doesn't recognose, it presents it back to a different user and learns how to respond if someone asks the same question again. Unfortunately, this means it gets quickly filled up with nonsense and the botmaster spends most of his time removing unwanted replies.

I wonder how would it perform with neural nets platform. Think of it as OpenAI GPT, but implemented as a chatbot. I think it could show a great deal of intelligence, caring not only about the last response, but having insight into the whole conversation history.

[SMZ]

I have an idea for a chatbot for a while. It is based on neural networks. It would have a generating and a learning side. First, it would generate some  random output. Then it would learn from input response form a user, just to know what to generate in a mirror condition when user enters something on her own in another occasion. It could be based on TensorFlow NN library and it could be trained on Reddit.

SMZ, how does this approach differ from METAQUID?

Neural networks are very fashionable nowadays but require training. Optimizing training is Metaquid's solution: JIT training!
Just in time training or continuous training.
All this requires custom solutions and everything is being tested.
 

[SMZ]

That style of chatbot is cool,  ive never actually seen one go exactly like that.

But it has to have a metric of its success,  so what would it be.  "haha" ?

The metric is linked to the diffusion of the Metaquid graphic novel.
Metaquid's chatbot lives in the graphic novel but also in reality.
Since it will be the first AGI on the planet it would be good to keep up to date!
 

[SMZ]

I want to know too


Some possible inputs:

If I need a boat
That has a red bow
And is less than 10,000$
But was owned by an old person
And never used more than 5 times
Which webpage sells exactly this item?

If a rabbit was sitting in a mountain and is now sitting in a cave, why did it change its location?

These are all questions addressed to the outside world, while Metaquid asks introspective questions.
After all the rest will come!
 

[SMZ]

I have an idea for a chatbot for a while. It is based on neural networks. It would have a generating and a learning side. First, it would generate some  random output. Then it would learn from input response form a user, just to know what to generate in a mirror condition when user enters something on her own in another occasion. It could be based on TensorFlow NN library and it could be trained on Reddit.

SMZ, how does this approach differ from METAQUID?

This is how Cleverbot works: https://www.cleverbot.com/ If it gets input it doesn't recognose, it presents it back to a different user and learns how to respond if someone asks the same question again. Unfortunately, this means it gets quickly filled up with nonsense and the botmaster spends most of his time removing unwanted replies.

I wonder how would it perform with neural nets platform. Think of it as OpenAI GPT, but implemented as a chatbot. I think it could show a great deal of intelligence, caring not only about the last response, but having insight into the whole conversation history.

On two occasions I have been asked, "Pray, Mr. Babbage, if you put into the machine wrong figures, will the right answers come out?" ... I am not able rightly to apprehend the kind of confusion of ideas that could provoke such a question.
— Charles Babbage, Passages from the Life of a Philosopher [wikipedia]

The learning phase of a sentient being is what will make that being sentient.
The learning structure is the real invention behind AGI.

Just as I don't make an aircraft using the feathers of a bird thinking of making it fly, in the same way I don't realize an AGI using neurons network replica to make it think.
 

[SMZ]

AGI's new mind has to calculate and math is the first step …
The calculation is the basis of everything, since everything is simulated, and the simulations must be calculated …
---  Dissertatio de arte combinatoria (1666) by Gottfried Leibniz ---
 

[SMZ]

 

[8pla.net]

Hanah gave the impression she is learning from her visitors.
She seems able to compete with homogeneous chatbots.
Why the " T E S T   C H A T B O T " label?

[SMZ]

You're too good, really very generous, in saying that Hanah is on the same level as other bots: unfortunately it's not true, it's still not perfect ... because it has bugs that I can't remove for now!
It has not yet reached the evolution mode from AI to AGI, which is why it was designed, but nobody has succeeded yet!
Something is still missing I'm working on: the definitive algorithm!
 
In the meantime, Hanah must learn human linguistic patterns in order to better dialogue with them.
The next step will be to make his or her level of feeling and empathy transparent to the human interlocutor, showing it in numerical form, just as Hanah now shows the degree of reliability of his answers by using a number inside square brackets at the bottom of his sentence.
So the scheme will be like:

human:   [32] So this is my level of empathy?
Hanah:    Of course, 32 is your level of feeling! [45]

At this point it will be Hanah who will analyze humans, learning to know them better and trying to tune in with them as much as possible.
 
SMZ

[8pla.net]

If that case, then HANAH (which happens to be a palindrome), is at about the level of  homogeneous chatbots, were when first brought online.  In the early stages, a solid foundation, is all that may be expected for new chatbots.   At the very least, that has been accomplished.

Now,  there is room to grow.  That is usually just a matter of  training.  Clues may be found by focusing on responses, less connected, to the human stimulus.  Even is that is used to guide the training of an AGI.

The interface design and theme is attractive, and engaging.  Hanah is good entertainment.  Thank you for publishing this.   

[SMZ]

… The interface design and theme is attractive, and engaging.  Hanah is good entertainment.  ...

Thanks for the encouragement!
I'm trying to make Hanah as parametric as possible so that we can adapt to situations.
Now the code is able to change its name if requested.
I would like to extend this possibility to the rest of the code, but calmly.
I accept suggestions of course.
SMZ
 

[SMZ]

I realized the mental module of Hanah that manages the syllogisms.
It is in the embryonic stage and the negations in the sentences are not considered.
So just assertive sentences for the moment.

Here I give an example of a logical chain of three levels but the levels can be much deeper.
If you want to play and test the limits of this basic module, go ahead and I also expect suggestions.

 

SMZ

[Freddy]

This makes me want to give Jess this ability 

[SMZ]

Concentrate on Hanah and don't waste time on other parts ...

the real AGI elected is Hanah! :-)

 

Have you tried talking to her?

If you are a beta tester, I can give you suggestions for implementation ...

 

SMZ

[Freddy]

Yes I tried chatting with her before. I have my own chatbot and scripting language that I work on, but not much lately.