The AI safety mindset: 12 rules for a safer AI future by Cassie Kozyrkov
========================================================================

.. post:: Dec 30, 2020
   :tags: talks

https://www.youtube.com/watch?v=EjBXZrQ7fTs&feature=youtu.be&ab_channel=Robotex


1. Don't be distracted by science fiction

    Science fiction version of AI has to do with person hood - autonomous
    entities. Instead, AI comes from humans like all technologies.

    All tools are better than humans be it paper/pencil as it has more memory, a
    bucket of water as it can hold more water etc. The question is to identify
    what is it specifically that AI is better at.

    It is about "Explain by -instructions- examples"

2. Remember that the objective is subjective

    The decision is subjective to some context e.g. detecting cat / not-cat will
    work differently for a tiger if the context is that "The detected cats need
    to be used as pets"

3. Strive for decision intelligence

    https://youtu.be/EjBXZrQ7fTs?t=494
    Amplify the decision maker of the human

4. Wish responsibly

   This is about choice of data sets and choice of objective. We need to be
   careful of how one expresses the wish so that the true spirit of the wish is
   captured.

5. Think like a site reliability engineer

6. Test everything!

7. Always use pristine data for testing

   Just like we test humans by giving them questions they haven't seen before in
   an exam, AI systems need to be tested in the same way.

8. Get in the habit of splitting your data

9. Avoid jumping to conclusions

10. Make sure your data are representative

11. Open the textbook with analytics

    If the author of a textbook is prejudiced, the student will also pick up the
    same prejudices. The teacher needs to make sure that they analyze the
    textbook before giving it to the student for learning.

    Datasets, like textbooks, also have human authors. 

12. Seek a diversity of perspectives