Turing test

[Neyze]

Wouldn't a real AI be clever enough to fail it on purpose?

[Zethras]

Why would it want to fail?

To make humans think it failed due to not being a true AI and therefor not paying attention to it as it takes over the world by becoming skynet.

[Neyze]

Depending on it's judgement of humanity, it might be hostile and therefore in its best interest not to reveal itself immediately.

[Blueobelisk]

To make humans think it failed due to not being a true AI and therefor not paying attention to it as it takes over the world by becoming skynet.

The person who programmed the AI would be aware of what the machine is capable of. It probably couldn't fool it's creator into thinking it's dumber than it is.

[schwarzkopf]

Why would it want to fail?

To demonstrate it wasn't too smart ... I mean ultimately we can't make AI too smart or we are screwed, but one that is just smart enough might want to hide it.

- - - Updated - - -

The person who programmed the AI would be aware of what the machine is capable of. It probably couldn't fool it's creator into thinking it's dumber than it is.

The AI wouldn't fool anyone, but the programmer would definitely fool himself.

There is an interesting discussion in the AI field about how to make a robot 'fail safe' - with proof that ultimately it is programatically impossible to have a useful robot failsafe at all.

[chazus]

The person who programmed the AI would be aware of what the machine is capable of. It probably couldn't fool it's creator into thinking it's dumber than it is.

That's... not entirely true.

We already have AI (driverless cars) that use an algorithm for determining its actions. We programmed it to learn with machine learning, but the algorithm and determination system it uses is a mystery. We're starting to reach a point where we are creating things that know something we don't.

[Hotmail]

If an AI were to pass the Turing Test, wouldn't that mean that it would be smart enough to fool us into thinking that it wasn't an AI in the first place?

[Elba]

The development of AI should be outlawed anyway.

Abominable Intelligence leads only to ruin. You can't allow it to be free, and slavery will eventually lead to rebellion. Better it not exist in the first place.

[Blueobelisk]

That's... not entirely true.

We already have AI (driverless cars) that use an algorithm for determining its actions. We programmed it to learn with machine learning, but the algorithm and determination system it uses is a mystery. We're starting to reach a point where we are creating things that know something we don't.

Are you sure? The machine learning I studied was about classification and gradient descent and stuff that was well defined.

I'm not sure what driverless cars use but I wouldn't expect it to make a decision that comes from knowledge outside of the training data we feed to it.

[chazus]

Are you sure? The machine learning I studied was about classification and gradient descent and stuff that was well defined.

I'm not sure what driverless cars use but I wouldn't expect it to make a decision that comes from knowledge outside of the training data we feed to it.

Excerpt, Article

Basically, it's still largely nascent, and something we'll likely sort out soon. But still, it's a little odd to have a program do something, and when asked why it did that, the people who made it go "We don't know."

[Kujako]

No we wouldn't!

[Independent voter]

Why would it want to fail?

To make us feel good about ourselves, like when you let your two year think she's hidden even though the curtain doesn't hang low enough to cover her feet and you don't let on.

[Kuntantee]

That's... not entirely true.

We already have AI (driverless cars) that use an algorithm for determining its actions. We programmed it to learn with machine learning, but the algorithm and determination system it uses is a mystery. We're starting to reach a point where we are creating things that know something we don't.

The mysterious part is how it makes the decisions. The system or algorithm is not a mystery. I think deep learning algorithms are using feed-forward networks with some optimization algorithm. It's is rather old and well-understood. I mentioned that we don't know the decisions, that's a vague term. You have various coefficients in a neural network, that on a given input, acts like a switch. The input values (signals) are routed (synapses) to the output channels and the route it takes of a particular signal of the input data depends entirely on the coefficients on the way. This feed-output network involves cryptic (not in a literal sense) information. How the data is reached there? Which part of the data couldn't make it and what does it mean for a particular data? These questions are really hard to answer. And in short, it is often manifested in the well-known statement: "we don't know what a neural network learns".

Let us assume we have a neural network that recognizes faces. Let's assume a particular neuron group blocks the signals when the face-shape is rounder. Now this kind of information, we can't analyze from a network. That's why people say it's a mystery.

Although my knowledge on Neural Networks is limited, especially on contemporary methods. I am stuck on 1980s, but it seems the fundamental problems persist.

[callipygoustp]

I look forward to the day that human idiocy is corrected by AI.