# Tag Info

6

The Control Problem is, in short, the idea that AI will eventually be much better decision-makers than humans. If we don't set things up correctly beforehand, we won't get a chance to fix it afterwards, because AI will have effective control. There are three main areas of discussion with regards to the Control Problem: Whether or not the problem is urgent. ...

5

There is a small survey of continuous states, actions and time in reinforcement learning in my thesis proposal. Regarding books, Reinforcement Learning: State-of-the-Art seems to be pretty up-to-date from the excerpts I've read.

5

Yes, to some of what you propose. No to some. Today corporations are granted rights: to own property, earn income, pay taxes, contribute to political campaigns, offer opinion in public, ad more. Even now I see no reason why an AI should not be eligible to incorporate itself, thereby inheriting all these rights. Conversely, any corporation already in ...

5

From the book Reinforcement Learning, An Introduction (R. Sutton, A. Barto): The term system identification is used in adaptive control for what we call model-learning (e.g., Goodwin and Sin, 1984; Ljung and S ̈oderstrom, 1983; Young, 1984). Model-learning refers to the act of learning the model (environment). Reinforcement Learning can be divided ...

4

Will Artificial Intelligence some day become a problem to humanity after learning human behaviors and characteristics? It can be answered in both ways, I think. Yes, they may become a problem. With the increasing integration of loads of apps and smart devices in our life, almost everything defining an individual human being is digitalised. For instance, ...

4

I find the concept of the a Turing machine useful. In one dimension, everything is a string. All of the parts that are "not you" are merely a substrate, a medium for the program your_mind runs on top of. The you, your identity, the "metaphysical" component we think of as the mind, is a result of running the algorithm that is your_mind on the bioware of your ...

4

In this case, $\pi$ has always been an $\epsilon$-greedy policy. In every iteration, this $\pi$ is used to generate ($\epsilon$-greedily) a trajectory from which the new $Q(s, a)$ values are calculated. The last line in the "pseudocode" tells you that the policy $\pi$ will be a new $\epsilon$-greedy policy in the next iteration. Since the policy that is ...

4

Typically the ramification of overfitting is poor performance on unseen data. If you're confident that overfitting on your dataset will not cause problems for situations not described by the dataset, or the dataset contains every possible scenario then overfitting may be good for the performance of the NN.

3

Murray Shanahan, in his book The Technological Singularity, makes the case that the rights of any being are determined by its intelligence. For instance, we value the life of a dog above that of an ant and likewise value human life above that of other animals. From here one could argue that a general artificial intelligence of equal intelligence to a ...

3

I think most answers that come to mind are going to be connected to the control problem in some way---e.g. sandboxing frameworks, AI behavior monitoring protocols, etc. However, if you aren't interested in those for now, that pretty much leaves two possible needs such technology could satisfy: the technical challenges of AI itself, and the need to optimize ...

3

Is overfitting always a bad thing? The answer is a resounding yes, every time. The reason being that overfitting is the name we use to refer to a situation where your model did very well on the training data but when you showed it the dataset that really matter(i.e the test data or put it into production), it performed very bad. This can never be good, ...

3

According to Ray Kurzweil, a prominent AI researcher, yes. In his book The Singularity is Near he predicts that AIs will take over developing other AIs in about 30 years, after which human intelligence will become marginalised.

3

Metaphorically: make it so depressed it commits suicide. As per my answer to this AI SE question, the idea is to feed it a sequence of inputs that will cause it to become (permanently) inactive. The technical details of how this might be achieved (and they are somewhat technical) can be found in this paper.

2

The Mars Rover is a highly successful example of the 'New AI' that emerged from work by Rodney Brooks in the 1990s. In a quote from Brooks: In 1984 I joined the faculty at MIT where I have been ever since. I set up a mobile robot group there and started developing robots that led to the Mars planetary rovers. Together with the 'Allen' paper, the ...

2

If an AI is developed by humans, we surely can create another one! Develop another AI agent without all the possible bugs that can make it go rogue to tackle the rogue AI, but more technically advanced than the previous one. Hardwire it with the sole purpose of disabling any rogue AI agent that can harm humanity and have it self-destruct in case it is ...

2

Nuke it from orbit - it's the only way to be sure If you want to be really sure you destroy everything of the AI, you'll need to launch an EMP (ElectroMagneticPulse) from the orbit (there are different ways to achieve this, one would be an atomic bomb, but there are better ones). EMPs will destroy every electronic device it hits without causing really much ...

2

No matter what rights it gets (as a company), it will still lack the right of not getting liquefied and all its properties transferred back to natural persons. This is of course if no laws are changed. To change the laws you will need to convince people that this machine is more "life" worthy than intelligent animals, and hope that people will deal with ...

2

As long as the output of the AI affects the world, the way in which it communicates makes no fundamental difference to the control problem. The AI might still be able, for example, to manoeuvre mankind into a situation, in which only the AI can save us. It might provide a seemingly inoccuous technological solution to global warming, but 50 years later it ...

2

I would say no due to the possibility of psychological manipulation of the messenger by the AI. Also, the LED communication constraints place severe limitations on the capabilities of the AI, as the usefulness of AI is likely predicated on its ability to learn quickly from a vast amount of information (e.g. using the internet). In some sense you may ...

2

Impending or Past? Niel Postman, in his book, Technopoly, argues the preemption of human centered culture to technically driven culture has already occurred. Jaques Ellul, in his book, Technological Society, heaped evidence behind the proposal that technology became autonomous centuries ago. Their arguments are convincing. Some think other criteria must ...

2

What you are talking about is known as the Control Problem. We have our own tag for this specific topic here, which you can use for this and similar questions. How to address the control problem is heavily discussed and still considered unsolved. Two of the important approaches are motivation control and capability control. Motivation control aims at ...

2

May not be quite what you’re looking for, but nonetheless helpful, I hope. The White House a year ago commissioned a report on AI that touches briefly on policy issues.

2

These kind of questions cannot be answered without looking at a particular project. Each algorithm has its particular strengths and weaknesses; and trade-offs in terms of use of resources (processing power and/or storage space, for example). If there was an objective answer, then the worse algorithm would surely fall in disuse. It also depends what you mean ...

2

Meaning of Low Level Goals in Data-Efficient Hierarchical Reinforcement Learning * * Ofir Nachum, Shixiang Gu, Honglak Lee, Sergey Levine Before explaining why a move or a push are goals, let's examine a statement from a prior paper. Although Sergy Levine (Berkeley) is well published and a respected contributor, we can see why Ofir Nachum (Google) is ...

2

Learning is possible without random thoughts and actions. Knowledge can be encapsulated in predetermined forms and passed through predetermined knowledge transfer mechanisms. Much of civilization is based on these predeterminations. Without them, humanity would be thrown back possibly 120,000 years. However, initial discovery requires trials and review of ...

2

I think that something like "strength" would be difficult to quantify in this context. I do think that formal experimentation around the "AI in a box" scenario could be interesting. I know that experiments have been done where a human plays the role of the AI, attempting to get naive test subjects to "release" him by interacting with them over a chat ...

1

Overfitting is almost always bad and hurts generalization. You say what we want from the NN is to say "take a slight left turn" or "turn right hard" if another ship comes slightly close on the right or rapidly close on the left, respectively. But what would you say if the NN learns to "take a slight left turn" only if the coming ship is small (...

1

Predicting what happens post-singularity is simply not possible as we cannot attempt to model let alone conceptualise a mind far more complex than ours. If that is a difficult concept to get your head around, consider how far an insect's central nervous system could go in understanding human behaviour. That fact alone is an argument against the likelihood ...

1

"Ego death" Ego death is a "complete loss of subjective self-identity." -"Ego death", Wikipedia The defining aspect of an ego is an exclusive sense of self. If an ego-dominated individual joined a greater mind and dissolved within it without really dying, then the only thing that it's really lost is its ego. Real world analogy: political states' ...

Only top voted, non community-wiki answers of a minimum length are eligible