A couple of funnies from the internet

 

Another: Why do meteorites always land in craters?

Adventures in Turing testing

I found a couple of vulnerabilities in the current Bing AI. One centers around the experience of peeling an orange. The other around AI's lack of free will.

I asked: why do you peel an orange? It gave a good answer.

I asked: how do you eat an orange? The first step was to wash the orange. Obviously that  is unnecessary since you are going to remove the peel.

I asked it: how do you peel an orange. It gave two answers - showing more information that a person would.

I asked: why would you peel an orange that way? It had no good answer. Human example: 

    cuz that is how my mother did it

    cuz I like to use my knife, and it makes less mess

    cuz I like to separate the fascia of the segments

I tried to get the AI to explain why it was not able to initiate an action. It get making excuses for not being designed to do that. This is an area that could be explored.

Carnitas recipe

The correct way to make this is the cook the ingredients together - see a typical recipe. In my case, I made a slow-cooked pork roast for direct eating, intending to make empanadas afterwards. Instead I decided to make burritos and adapted a carnitas recipe to start with pre-cooked pork. This is not spicy but you can use Jalapeno and add heat with eg Tabasco sauce.

Ingredients:

 2-3 cups of cooked pork, chopped, 

spices: black pepper, oregano, chilli powder, bay leaf, cumin, parsley sprig. I also used one dried non-spicy pepper.

small onion, garlic cloves

1/2 green pepper (or Jalapeno)

orange, 1/2 lime

lard

Start the onions at medium heat, use lard. Add meat, spices, then garlic and green pepper, add citrus fruit juices. Add 1 cup of water.

Simmer for 1-2 hours. Stirring and topping off water. End with a small amount of liquid.

Use this "stew" with rice, beans, guacamole, sour cream, in a burrito or tortilla. 

How to beat the AI in a Turing test

The Turing test has an obvious flaw: a failure to specify the intelligence and knowledge of the real person testing the computer. For example if the real person is a newborn child, it cannot ask ANY questions. Or, slightly more reasonably, if the real person is particularly dumb and asks extremely simple minded questions.

But never mind that. I am contemptuous of computer scientists who think "intelligence" is possible in a un-embodied system with no sensory/motor perceptions. The real trick is to ask the AI a question that has never been asked before and cannot be answered by cutting and pasting previous answers. To do that, I am imagining examples involving sensory/motor perception.

My first try was this: "How does it feel to peel an orange?". This did not quite work with Bing's AI because it put together various descriptions of an orange that were available to it. If you look at what a AI says to this question, it is clear that the answer is cut and pasted from every bit of information the AI had available. Here, you need the real person doing the testing to know a great deal about how other people answer the question. People answer simply but the computer goes to great lengths to add details that a person would not. So the tester must already know the human answer [and be careful to not record any such answers or the AI can cut and paste].

A second try was this: "You are in a room with a table and an orange on the table. What do you do if you want to eat the orange?" This is supposed to test the AI's ability to form a plan. This is not good test because a good AI will know how to create a plan.

So let us ask a new question, never asked before: "Can you eat pudding with chopsticks?"  [To be honest, Bing's AI did a reasonable job with this question. It knows the difference between solid and liquid.]

It is not OK to come in, after the fact, and say every possible action with every possible material could be learned by an AI. One can argue that there is a continuum of questions and no continuum of programmable facts. The only way to ensure the continuum is available to the AI is to give it a working body.

Humbler Pie: It seems that however clever I can try to be, coming up with a kind of question, 100 Google engineers can come up with it as well and add a solution to the AI function.