[AgentSmith]

Hello all,

a crucial thing most of us share is without doubt the fascination for AGI. An AGI is some sort of general intelligence that enables an agent to perform potentially all tasks a human could perform given that the agent has sufficient physical capabilities.

To create an AGI it might be helpful to have examples of such tasks such that one knows what sort of problems he has to take care of. Additionally, such examples could be used to disprove the AGI property. To support the work of all who are working on AGI, I would like to ask for good examples of things an AGI must be capable of. The examples can be abstract to make generalization easier. On the other side, the examples should also not be too abstract as these might be too difficult to be covered by premature AGI concepts and thus less helpful. It might also make sense to refine/specialize provided examples later if they are too abstract or to generalize them if they are too concrete.

I will make a start with the following example: An AGI should be able to improve its decision-making on the base of past experiences.

[ivan.moony]

Maybe to create a true life?

[infurl]

I'd settle for an intelligent conversation.

[AgentSmith]

Maybe to create a true life?

Well, if an AGI itself can be considered to be a "true life", it just has to reproduce itself to create such one, right?
In this case, your example could also be reduced to the problem of self-reproduction.

[unreality]

I don't think mainstream has detailed the definition of "life" enough to decipher when AI becomes alive. Everyone has their general feel for that. On youtube there's a dog robot that's learned how to walk. In the video their creators are kicking it around, making it fall. This has caused a movement to such a degree that a group has filed for AI rights. IMO present AI hasn't reached the level where I'd consider it alive, but we're not far away from that. Definitely within a decade when AI is extremely complex like humans, consisting of terabytes of complex analysis, which will definitely give each AI their own personality, way of expression & desires (emotions).

BTW it's said that human brain might have 1 or 2 petabytes, of which only a very small percentage is long-term memory. That's what I call sloppy memory. It's not equivalent to AI computer memory, unless we're talking about AI neural networking, which I'm not. So 100 TB of uncompressed non-NN AI memory will most likely be equivalent to at least a few petabytes of human brain memory. Compressed memory holds even more. Also lets not forget that humans don't have conscious access to that memory. The world record for the most digits a human has remember pi is 67890 digits, which is equivalent to about 55 KB (uncompressed) in hex.

[keghn]

 It can be done at AAI, Artificial Animal intelligence. It will scale with better firmware and hardware. For sure.


 Meaning of life for me is a energy using device is to survive to the end of time. Since life is so bad at predicting the future
it make copies of its self and spread out. Surviving copy share genetic information to build better copies.
 Since smaller life replicated faster and will defeat bigger life. Two surviving life forms that having the anti microbe betting
attributes. That will be enhance to a level faster then microbe can adapt. Just like two point to give a leap frog direction to make the anti
microbe immune system in the child.
 the two adult immune factors will be close but the difference will give direction to improve.?

 Large silicon life may not have this problem.
 Silicon micro life may need too much energy to start small. Maybe only ant size and up? Unless there is better solvent for
silicon chemicals.

Notes:
orthosilicic acid
https://en.wikipedia.org/wiki/Venus%27_flower_basket

https://www.scientificamerican.com/article/sea-sponge-inspires-bette/
https://www.newscientist.com/article/dn11329-algae-skeletons-made-into-silicon-components/
http://greenedge-expeditions.com/diatoms-and-silica-a-shell-of-glass/

[ranch vermin]

I'd settle for an intelligent conversation.

even a fairly dumb one that retains attention span context would be amazing already.

[Korrelan]

The human ‘brain’ is the only ‘object’ that has ever named it’s self.

So perhaps… when an AGI chooses a relevant name for it's self.

 

[AgentSmith]

Maybe to create a true life?

Well, if an AGI itself can be considered to be a "true life", it just has to reproduce itself to create such one, right?
In this case, your example could also be reduced to the problem of self-reproduction.

After some thinking I come to the conclusion that self-reproduction is actually not a property of intelligence, since even micro-organisms are able to do this without any need of intelligence. Creation of life itself relies on intelligence if it refers to the creation of a sketch of life rather than the creation of life itself.

The human ‘brain’ is the only ‘object’ that has ever named it’s self.

So perhaps… when an AGI chooses a relevant name for it's self.

 

If I write a simple program displaying "I am a simple program with the name " + randomlyGeneratedString + ".", would this already fulfil your requirement?


Another requirement of an AGI I want to state is the employment of some sort of internal optimizer that optimizes the program of an AGI such that it improves over time. It needs to be cleared how optimizers work and how they can be used to improve programs.

[Korrelan]

The human ‘brain’ is the only ‘object’ that has ever named it’s self.

So perhaps… when an AGI chooses a relevant name for it's self.

 

[Art]

Perhaps it will not even Have to name itself. It will know what it is and it's capabilities and it will know all things. Why bother with a name at all? Unless for the puny little humans who cower before it wishing to have some name with which to address it in the most humble of manners and with the utmost respect so as not to offend nor anger "It who knows all things". I'm sure if the time comes, It will provide a name for those humans to use.

[Korrelan]

Yeah! Something catchy like...

Korr-Tex... haha...

 

[AgentSmith]

The human ‘brain’ is the only ‘object’ that has ever named it’s self.

So perhaps… when an AGI chooses a relevant name for it's self.

 

So the meaning of relevant is defined by some humans and not the AGI. What, if these humans do not agree with their definitions of this term?

"It who knows all things".

Exactly like humans, an AGI cannot know everything and will always have limitations (maybe it is worth another thread). Lets not treat AGIs as hypothetical artificial gods that actually cannot be realized due to logical impossibilities and so on. An AGI just has to prove that it can achieve (and maybe even surpass) certain levels of human intelligence. Achieving this is clearly realistic and powerful.

-

A new AGI requirement from my side: Ability to define subgoals in order to achieve a high-level goal.