How do I start programming a self learning basic AI?

You can Start Learning About Artificial Intelligence (A.I.) From Here - DialogFlow

This is the gazillion dollar question if such a named number exists! The straightforward answer is “I don’t know but I have some clues and hints”. Some of the answers below highlighted a few critical ingredients for self-learning AI. Below I will summarize these adding some of my own:

1. It needs to have some sensory capabilities like human beings. But obviously, “sensory capabilities” are not limited to those we have such as vision, hearing and the sense of small. In a abstract way, a sensory capability is simply a means to receive information about the outside world through measurements. In that sense, just about any means of physical measurement would do such as, e.g., receiving information from other computers in the form of current fluctuations we call bits!
2. It needs to be connected to its environment and that environment must be as large as possible. If the AI we are after is not mobile, e.g. not robotic and capable of physical motion, such connectivity could be through unfettered access to the Internet.
3. It needs to be able to make mistakes. Yes you read this correctly. If it doesn’t make mistakes, then the problem it is facing is already resolved and its operation can hardly be qualified as intelligence but more of an automaton.
4. It needs to be in contact with other forms of intelligence, artificial or human/animal. This is crucial because much of what we learn is due to our interaction with other or similar forms of intelligence. This is not as circular as it seems: humans, a higher form of intelligence, interact with other lower forms of intelligence such as animals so the initial interactions could be between the self-learning AI and other lower forms of AI that can perform specific tasks.
5. It needs to be able to replicate so as to create other forms of intelligence with at least the same potential as itself. Furthermore, the replication must be inexact, otherwise it becomes sterile: simply a copy of the same machine incapable of responding to stimuli differently from the mother template.
6. Last and by far the most important: it must be able to formulate its own hypotheses, i.e. ask questions it was not pre-programmed to ask or answer.

Requirements 1 is what most of machine learning is about nowadays: image and speech recognition etc. Requirement 2 is being addressed in a limited way by some people but in a haphazard way. Some learning methods such as genetic algorithms allow for requirement 3 as a matter of course but their methods are still crude. To my knowledge requirements 4 and 5 have not yet been addressed seriously while requirement 6 seems to be complete virgin territory.

Source - Quora