Teaching an AI? Rather than programming it?

You can teach AI by telling it to learn from random responses. From those responses, possibly what kind of answers it can say to carry on a good conversation. It would be difficult, but entirely not impossible. Our minds run on statistics too...... the only difference is the formations of invisible facts that MAY exist to balance an equation that is to our knowledge, unexplainable (beliefs... religion)

 

Log in or Sign up to hide all adverts.

www.20q.net is fun!


it's been learning for 20 years!

 

Log in or Sign up to hide all adverts.

Different thread in the same forum regarding "chatbot" "AIs" http://www.sciforums.com/showthread.php?t=44608

G71, that was a very interesting if somewhat cryptic post. Do you have a link to the project you were referring to? I develop in .NET, primarily C# so it is of much interest to me. Thanks.

 

Log in or Sign up to hide all adverts.

/bump

 

Nexus: It's not online now. The doc needs some updates and there is a lot of code to write. We can discuss details through email if you are interested. jjelinek.us @ gmail.com

 

Hi, I'm new to this forum. The topics are quite interesting, so I joined.

I believe that the correct response to a question depends on the question AND the environment. The AIs that I know of do not take into consideration the environment. This could be things such as the time of day, the setting (for humans, anyway) and quite importantly, recent dialog/topics/facts (/moods?). Some factors in the environment are more important than others and the AI should know this.

Therefore, answer=f(question, environment)

I suspect that a true AI cannot be programmed using structured programming. If it can be done with a computer, I suspect it will be done using artificial neural networks or self-modifying code (which will become quite messy, very quickly).

 

It all depends on how you define the "true AI". Structured programming is being used within object oriented programming (OOP). I would be interested to know what part of the "true AI" functionality cannot be coded using OOP (?). OOP is IMO suitable for coding the core of the "true AI". Thinking is IMO primarily an object oriented activity. There are some tasks for which it makes a lot more sense to use artificial neural networks (ANN) rather than OOP (and other types of programming), but when talking about "true AI", I view those as lower-level sub-tasks. So ANN might be just a handy tool for an AI system (like a calculator for me - I just use it when it makes sense).
Whenever I had something to do with a self-modifying code, it was always the type of code which was changing itself during runtime (which is an extremely hard-to-read type of code and developers should stay away from that technique whenever possible!). AI should IMO never change its own code when it's running. It can play with copies of its components, do a lot of testing and then possibly replacing some of its original parts while being ready to switch back to the previous version if needed. Some of the core parts (code and data) should never be modified by the AI without an approval from the author (or an authorized subject). Our AI systems are being designed to work for us. It makes no sense to invest resources into development of an AI system which would self-modify so that its goals and our goals could/would become incompatible and could cause undesirable scenarios. So, self-modification - YES with limits, but I do not see it as a necessary thing for the first version of the "true AI". ANN - YES, but it may not be needed for the first version as well (depends on the environment/interface). "True AI" may initially have nothing to do with ANNs and self-modification techniques. Both things can come later, after the system gains significant amount of knowledge.

 

the field needs thinkers like you keep it up

 

Theoneiuse,
Data-driven AI can exceed the human level of intelligence and can get hardly predictable since it will see more links between concepts than humans.