When's the Singularity?

cyberwolf77

I speak to you of the technological edition. Some say it is coming, I feel though it already happened. We are using it to communicate now. Ladies, and gentlemen written language was a singularity of the technological stripe. Next up , will be the development of machine languages which allow the creation of artificial intelligence. Once that happens, we are doomed. Floor is open for debate.

AlphabetCards

Are you talking about the complete removal of the conscience from the biological substrate to the silicon variant?

Den14

I'd say that the interesting point of a singularity is exactly how we distinguish what is truly alive. Thoughts coming from a human vs replicated thoughts coming from a machine. They may match in character and personality to a tee but is the person really there? Transfer of brain info to a machine may take all info which is conscious information but will it also include a subconscious element. Can we in essence upload our souls?

cyberwolf77

AlphabetCards wrote: »

Are you talking about the complete removal of the conscience from the biological substrate to the silicon variant?

Essentially brain uploading, yeah.*




*Poster has no academic credentials, he just a big geek

cyberwolf77

Den14 wrote: »

I'd say that the interesting point of a singularity is exactly how we distinguish what is truly alive. Thoughts coming from a human vs replicated thoughts coming from a machine. They may match in character and personality to a tee but is the person really there? Transfer of brain info to a machine may take all info which is conscious information but will it also include a subconscious element. Can we in essence upload our souls?

Another question to consider is, " When we create true artificial intelligence will it have a soul?"

Black Swan

Programmer friend of mine suggests that ASI is already happening. Models two competing AI systems that challenge and learn from each other at a tremendous rate. Claims his fellow programmers that collaboratively work with ASI often encounter things that work that they do not understand. Only that they work.

cyberwolf77

That just unsettles me to no end.

AlphabetCards

There's quite a good book on this very subject called 'To Be a Machine" by Mark O'Connell which is really worth a read. It did unsettle me for a few days to be honest.

cyberwolf77

AlphabetCards wrote: »

There's quite a good book on this very subject called 'To Be a Machine" by Mark O'Connell which is really worth a read. It did unsettle me for a few days to be honest.

Always good to find a new book. Thanks for the suggestion.

Fathom

cyberwolf77 wrote: »

That just unsettles me to no end.

Ask Ex Machina (2014).

AlphabetCards

https://www.theguardian.com/books/2018/apr/30/transhumanism-movement-wins-wellcome-book-prize-mark-o-connell

Well, what can I say I must have impeccable taste...!

cyberwolf77

Indeed, I have it lined up after I finish the novella I'm on now.

cyberwolf77

Fourier

Black Swan wrote: »

Programmer friend of mine suggests that ASI is already happening. Models two competing AI systems that challenge and learn from each other at a tremendous rate. Claims his fellow programmers that collaboratively work with ASI often encounter things that work that they do not understand. Only that they work.

Is that not to be expected from algorithms that use statistics to "walk" a space of solutions? I mean biological evolution is similar, it has produced designs we don't comprehend via a "random walk" with some forcing. Anything that operates like this will be able to find solutions you don't understand because you didn't input them.

Fathom

Fourier wrote: »

I mean biological evolution is similar, it has produced designs we don't comprehend via a "random walk" with some forcing. Anything that operates like this will be able to find solutions you don't understand because you didn't input them.

Random walk. Rapidly evolving AI.

Fourier

Fathom wrote: »

Random walk. Rapidly evolving AI.

I don't understand I'm afraid.

cyberwolf77

I think Fathom is suggesting that if an AI was in command of a robot capable of locomotion it's evolution could be traced by how soon it started creating random movement patterns on it's own.

Fourier

cyberwolf77 wrote: »

I think Fathom is suggesting that if an AI was in command of a robot capable of locomotion it's evolution could be traced by how soon it started creating random movement patterns on it's own.

The "Random Walk" I'm referring to isn't physical motion, but the exploration of the space of algorithms conducted by an AI.

It walks, in the sense of moving between algorithms, abandoning one in favour of another. The walk is often random because it uses statistics, as opposed to a fixed decision theory check, to decide what parts of the algorithm to change.

Fathom

cyberwolf77 wrote: »

I think Fathom is suggesting that if an AI was in command of a robot capable of locomotion it's evolution could be traced by how soon it started creating random movement patterns on it's own.

Rapid, random interactions between competing algorithms. Later becomes targeted movements. Rather than random.

Fourier wrote: »

The "Random Walk" I'm referring to isn't physical motion, but the exploration of the space of algorithms conducted by an AI. It walks, in the sense of moving between algorithms, abandoning one in favour of another. The walk is often random because it uses statistics, as opposed to a fixed decision theory check, to decide what parts of the algorithm to change.

See Wallace's Wheel of Science. Beginning with big data. To analysis. To empirical generalizations (patterns discovered in data). Models created (algorithms). Contributes to creation or revision of theories.

Fourier

I know Wallace's wheel of science, I'm not sure of the relevance here I'm afraid (I may be missing something). The algorithms in AI change components of their structure by a Markov chain or some similar method, they don't do anything like having theories or hypotheses. They just have a utility function that they try to maximise.

Fathom

Fourier wrote: »

I know Wallace's wheel of science, I'm not sure of the relevance here I'm afraid (I may be missing something). The algorithms in AI change components of their structure by a Markov chain or some similar method, they don't do anything like having theories or hypotheses. They just have a utility function that they try to maximise.

Markov chain? Makes note. Look up. Thanks Fourier.

Fourier

No worries Fathom, if I get the time I'll say a bit more.

For now just to say that AIs make use of stochastic algorithms such as a Markov chain to alter themselves, although a given AI might use only a Markov chain, or a Markov chain with something else.

The basic idea is that, let's say a chess playing AI called Albert, has someway of evaluating itself, hypothetically it might play 1000 games of chess against against another AI, called Betty. It wins 340 times. It then makes a random alteration of its own code and plays again, this time it wins 280 times, so it rejects that alteration, undoes it and makes another alteration. This time it wins 409 times. Hence it accepts this alteration, which becomes its new state. It will keep doing this until it has collected enough alterations to win, say, 990 times against Betty.

When that happens, it replaces Betty with itself and starts the process again.

It's very easy to see how quickly the AI might do stuff you don't understand, as the code was altered randomly, with no actual guiding principal or logic.

You want a way of making random changes that:
(i) Only depends on what the AI is like now, i.e. no memory of earlier states. For example:
Update line 104 of the code to "B" if it is currently "A" 40% of the time. The probability only depends on the value now, it is not
Update line 104 of the code to "B" if it is currently "A" and was previously "C" 20% of the time.
This is mainly for simplicity and ease of memory use (it also has nice mathematical properties, but that'd be too technical to go into)

(ii) Doesn't get trapped. Making random changes can have the problem that the AI might basically get trapped in a loop of making and rejecting the same set of changes over and over again.

A Markov chain is something that satisfies (i), you usually also want (ii) which is known as detailed balance.

Fathom

Fourier wrote: »

For now just to say that AIs make use of stochastic algorithms such as a Markov chain to alter themselves, although a given AI might use only a Markov chain, or a Markov chain with something else. The basic idea is that, let's say a chess playing AI called Albert, has someway of evaluating itself, hypothetically it might play 1000 games of chess against against another AI, called Betty. It wins 340 times. It then makes a random alteration of its own code and plays again, this time it wins 280 times, so it rejects that alteration, undoes it and makes another alteration. This time it wins 409 times. Hence it accepts this alteration, which becomes its new state. It will keep doing this until it has collected enough alterations to win, say, 990 times against Betty. When that happens, it replaces Betty with itself and starts the process again.

Cool explanation. Very clear.

Black Swan

Useful Fourier.