[ivan.moony]

While I'm working on a programming language that would be used to program AI, I took some time to sketch an idea, so I won't forget it.

Here in attachment is a rough sketch of an algorithm that responds in a way that copies human reactions. It remembers what was human input after certain computer output and responds on our input in the same way we responded on similar computer output. Safety checks and intelligent input analysis are left out of the sketch. This way we simulate human intelligence and emotional responses without having emotional rosettes and rules for emotional responses. No built-in rules, just a plain copying of responses. I think that this could be a key to intelligent behavior (we all learn by remembering other else's responses to certain situations).

But the big open question remains concluding function F from the algorithm because copying here is not simple copying word by word, but copying a behavior patterns (like "What is the color of Ivan's car? -> Black." is being remembered as "What is the color of X's car? -> X.car.color").

[Freddy]

I like the idea, it's similar to something I have been thinking about too. Learn a response to a question or phrase, maybe have some decision making made on what's in the input phrase to get a better match. Also a feedback process, like telling the bot when it's not a good reply and suggesting what may be. So some kind of scoring thing in there too.

[ivan.moony]

I had a question what to do on the very first input, when no responses are yet recorded.

We can mimic babies. What babies do? Copy. And then you get a human response to remember. The more AI knows, the less it needs to copy. Maybe later it can even learn what to do when there is no response in database.

And another thought. I think that at the first, when we are born, we are equipped with a fast built-in reasoner we experience from unconscious way. And later, when we are growing up, we can override it with high-level, totally conscious and adjustable algorithmic analyzer.

[squarebear]

This is the same method that Cleverbot uses. Try that to see how good this method is.

[ivan.moony]

Squarebear, what Cleverbot? Do you have any link, please? I'd like to check.

[ranch vermin]

Sounds great Ivan,  itll work.
Machines when they copy, they like to do it exactly,  you may add some intelligence with a frozen method...  like for example what to ignore out of the data coming in the sensor, or how it interchanges with some other data, to link it to the reply.

I always try to get entropy to do something for me,  I still havent got there yet tho.

[Don Patrick]

www.cleverbot.com , the most famous chatbot on the internet, copies and regurgitates question-response pairs learned from online chat with humans.

The problem with such a basic idea is that it could describe anything, from cleverbot's plain parroting of responses to the super-hard-problem of NLP and fact extraction, or just machine learning in general.
Mimicking responses works for standard sequences, like "hello", "hello", "how are you?", "I'm okay", "how are you?", but it is far from versatile. Learning literal responses fails to address references like "his car", or paraphrasing of the same question "What colour does Ivan's car have?", "Is the car of Ivan black?", or context, or intent. Of course you can try it out for the sake of curiosity as others have done: http://arxiv.org/pdf/1506.05869v1.pdf
It looks to me like you've left out all the interesting parts. The hard parts.

Infants don't just copy responses, they copy words associated with objects and events, and they combine them dynamically. e.g. "I want" + word. At simplest they copy a pattern, but they generate the response.

[ivan.moony]

It looks to me like you've left out all the interesting parts. The hard parts.

I didn't left it out, I didn't come to that part in my research, but I'm all over it. It looks like NLP parsing could do the magic? But this is just a tip of the ice rock, look at responding to implicitly concluded knowledge, it is not just responding to a parsed text entered by a user. And then there is passing response properties spanned through a whole conversation, not just the last input.

What a mess, what a spaghetti, will it ever end? Ok, calm down, we'll make it. Living beings exhibit intelligence, so can a machine inside the Universe.

Ommmmm...

Ok, peace is in me, I can continue, back to paper and pen 

[Don Patrick]

Well, I'll just throw in some words of wisdom: If you never try anything, all options remain forever possible.

[ranch vermin]

Yeh man - get into the mess of the implementation! 

[spydaz]

Maybe when it hears something new it should either return the same input back to the user with a question mark at the end ? Or just say . Ok .

Maybe?

By returning the input a response would be gained , whereas an ok ... Would mean that the computer would have a pending question awaiting. Response ready to be posed to a new user..

The pdf you gave shows conversations from an it helpdesk .. Quite interesting ...

Another interesting document

http://www.coling-2014.org/COLING%202014%20Tutorial-fix%20-%20Tomas%20Mikolov.pdf

uses text in neural networks with text....

The video by @snowman describing how he uses a type of markov processe to "guess" which word comes next in the text could be used in predicting the conversation or reply that would commonly be returned... Using sentences instead of words ...
Then many potential responses could be stored, to the same input selecting the most frequent response , starting with some movie scripts or telephone conversations a foundation could be created .... "Bootstrapping" the processes..

Then a combination of natural language proceessing and stored conversations from many users can be used ....

I noticed that the machine had a birthday and a location for itself ... By giving the ai some kind of personal details .. It may even be able to pass the turing test... And pass for human...

Another noticable thing was that some times instead of waiting for a response the machine replied when no answer was given.. Taking over the conversation.. Seeking knowledge...kinda nice too ... As when we don't get a response we also often send a question mark ... Or a Wtf.. ... Which it obviously learned from past conversations too ..

[Freddy]

I like the idea of the bot asking questions, this is something I have been thinking about implementing. I think it would make for a more natural conversation and make it seem more alive.

[spydaz]

I think that after somebody introduces themselves that there are a number of boxes which needs to be filled , the same with introducing a new object or subject .
Where is it, what is it used for, why would you use it?

After building a conceptual database all objects fit into categories and have various properties ... These are the properties that would be sought by ai... They would ve retrieved by thier counterpart question ...

[Zero]

I had a question what to do on the very first input, when no responses are yet recorded.

We can mimic babies. What babies do? Copy.

Actually, they don't start from nothing. They start from (apparent) randomness, or chaos. At start, random responses are recorded. For example, my baby can't choose how she will move her arm. The only thing she can choose for now (1 month old) is:  moving or not moving (that's the question).

EDIT: Maybe, when we are facing a situation, different responses quickly pop in mind one after another, until we choose "move now!". Which is why you can see your entire life when you're about to die (because you're searching really really hard).

EDIT: Maybe, it's also how we ask questions: we choose "move now!" when focused on unknown information.

EDIT: Cats also behave a bit like this. It's always the same loop:
- scan: look around until something catches attention
- evaluate: focus and decide if it's good for us / bad for us / irrelevant
- switch:
> if good for us: move closer and goto "evaluate"
> if bad for us: move away and goto "evaluate"
> if irrelevant: goto "scan"
What if consciousness was like cat, and thoughts was like cat's environment?

EDIT:
We have 4 actions here: scan, evaluate, move closer, move away
scan: just let thoughts pop up, like a multi-chatbots free-form inner conversation, with inputs from outside world (from special chatbots reporting sensors activity), until something unusual or unexpected happens
evaluate: pleasure or pain forecasting?
move closer: investigate further (if it's recording, playback, if it's question, find answer, ...etc)
move away: discard and go back to previous activity