Interesting mailing lists

  • 16 Replies
  • 4605 Views
*

ivan.moony

  • Trusty Member
  • ************
  • Bishop
  • *
  • 1723
    • mind-child
Interesting mailing lists
« on: August 10, 2021, 03:57:17 pm »
Anyone knows about interesting tech mailing lists worth of tracking?

I'm subscribed to
  • opencog@googlegroups.com (an AGI attempt)
  • open-nars@googlegroups.com (an AGI attempt)
things become interesting now and then, but not very often.

*

LOCKSUIT

  • Emerged from nothing
  • Trusty Member
  • *******************
  • Prometheus
  • *
  • 4659
  • First it wiggles, then it is rewarded.
    • Main Project Thread
Re: Interesting mailing lists
« Reply #1 on: August 10, 2021, 10:06:59 pm »
The topic box AGI mailing list, I had shared it before.

https://agi.topicbox.com/groups/agi

I still want to know how all these non-GPT projects are going to predict the rest of images like they do on openAI.com. Imagine NARS has some AGI behaviour maybe, great but it will never, ever compete in efficiency at predicting the rest of images unless it uses Transformer architecture. You can use NARS but when it comes to dreaming paths of thoughts, you can't beat GPT. Except for perhaps some very evil project that is like GPT, not really saw any that do compete though. LSTMs are ok etc but again these are not NARS etc.

My point is they should at least say they will combine NARS with GPT.....because if not how will they predict image so good? Efficiently?
Emergent          https://openai.com/blog/

*

MagnusWootton

  • Replicant
  • ********
  • 634
Re: Interesting mailing lists
« Reply #2 on: August 11, 2021, 11:49:48 am »
thanks for the links.   ;D

*

ivan.moony

  • Trusty Member
  • ************
  • Bishop
  • *
  • 1723
    • mind-child
Re: Interesting mailing lists
« Reply #3 on: August 11, 2021, 02:30:42 pm »
My point is they should at least say they will combine NARS with GPT.....because if not how will they predict image so good? Efficiently?

There are ways... see algorithm synthesis.

thanks for the links.   ;D

I trust you, guys, won't leave this place on that account.  :knuppel2:
« Last Edit: August 11, 2021, 03:28:24 pm by ivan.moony »

*

LOCKSUIT

  • Emerged from nothing
  • Trusty Member
  • *******************
  • Prometheus
  • *
  • 4659
  • First it wiggles, then it is rewarded.
    • Main Project Thread
Re: Interesting mailing lists
« Reply #4 on: August 14, 2021, 04:31:18 am »
My point is they should at least say they will combine NARS with GPT.....because if not how will they predict image so good? Efficiently?

There are ways... see algorithm synthesis.

But it isn't a working way. And brute force is unlikely to work, it's the cherry on top (1-5% useful). Those without a sure alternative should admit to the need to use GPT/Transformers for now. The reason they don't is because the years they put in would look like they have failed.
Emergent          https://openai.com/blog/

*

MagnusWootton

  • Replicant
  • ********
  • 634
Re: Interesting mailing lists
« Reply #5 on: August 14, 2021, 06:46:33 am »
My point is they should at least say they will combine NARS with GPT.....because if not how will they predict image so good? Efficiently?

There are ways... see algorithm synthesis.

But it isn't a working way. And brute force is unlikely to work, it's the cherry on top (1-5% useful). Those without a sure alternative should admit to the need to use GPT/Transformers for now. The reason they don't is because the years they put in would look like they have failed.

Brute force is the only Ai where the computer comes up with a decision itself (albiet based apon a spoon fed goal, how it adapts it is fully original and u dont need to code it, the computer does by itself.),    Its brother "BACK PROPAGATION" or "HEBBIAN LEARNING" isnt as magic, but it can be useful when you dont have as much processing power, but still a shitload is still required for both, really.

*

HS

  • Trusty Member
  • **********
  • Millennium Man
  • *
  • 1175
Re: Interesting mailing lists
« Reply #6 on: August 14, 2021, 09:50:49 pm »
My point is they should at least say they will combine NARS with GPT.....because if not how will they predict image so good? Efficiently?

There are ways... see algorithm synthesis.

But it isn't a working way. And brute force is unlikely to work, it's the cherry on top (1-5% useful). Those without a sure alternative should admit to the need to use GPT/Transformers for now. The reason they don't is because the years they put in would look like they have failed.

Brute force is the only Ai where the computer comes up with a decision itself (albiet based apon a spoon fed goal, how it adapts it is fully original and u dont need to code it, the computer does by itself.),    Its brother "BACK PROPAGATION" or "HEBBIAN LEARNING" isnt as magic, but it can be useful when you dont have as much processing power, but still a shitload is still required for both, really.

Input - output might not even be the only way to conceptualize AI. You could think of it in terms of the degree of environmental symbiosis and rate of useful novelty generation. Each input can exist in the context of all previous events, and each output may concern the totality of foreseeable futures.

Even really simple rules can do this crazy thing and self-generate a Turing complete search pattern, ala The Game of Life. Being tethered to this chaotic principle of nature might give it that certain life force, which we notice to be absent from present-day AI. Even tiny, isolated, closed loop bits of the universe are often incalculable, so I think the soundest method, really the only sane method, is to have your system build itself on real world results.

Now give the fundamental rule-set the ability to evolve, and you might be able to create a search pattern which optimizes search patterns. Then things like explorer modes and parallel attempts in the form of imagination may eventually develop to speed up an evolutionary approach. It may have the highest startup cost or activation energy, but I think it's got the potential to be the most powerful approach.

This got me thinking about a background search pattern for optimal neural behavior. Each neuron could build up a database of how it's behavior coincided with events, essentially building up its DNA as it goes. If you fed a game of life type rule set into, to randomly take an example completely at random, a cube made of 27 or more artificial neurons, I imagine you'd get an evolving self-sustaining pattern going. If you went a step further and enabled the rules which generate the pattern to evolve as well, then such a system might really get somewhere.

*

MagnusWootton

  • Replicant
  • ********
  • 634
Re: Interesting mailing lists
« Reply #7 on: August 15, 2021, 10:46:47 am »
Input - output might not even be the only way to conceptualize AI. You could think of it in terms of the degree of environmental symbiosis and rate of useful novelty generation. Each input can exist in the context of all previous events, and each output may concern the totality of foreseeable futures.

Everything can be put in terms of input -> output, so itself is a little ambiguous what it means.
Because of hidden variables, i.e. you can't know if the cat is dead or alive until you open the box,  you would have to concern yourself with both possibilities past that point?

This got me thinking about a background search pattern for optimal neural behavior. Each neuron could build up a database of how it's behavior coincided with events, essentially building up its DNA as it goes. If you fed a game of life type rule set into, to randomly take an example completely at random, a cube made of 27 or more artificial neurons, I imagine you'd get an evolving self-sustaining pattern going. If you went a step further and enabled the rules which generate the pattern to evolve as well, then such a system might really get somewhere.

Sounds good, definitely could work, I wish u good luck!   It would be alot easier if we had more powerful computation,    if we had computers that could compute astronomical fluid sims, even simple ones,  Ai would be easier as well.

*

HS

  • Trusty Member
  • **********
  • Millennium Man
  • *
  • 1175
Re: Interesting mailing lists
« Reply #8 on: August 15, 2021, 01:15:36 pm »
Because of hidden variables, i.e. you can't know if the cat is dead or alive until you open the box,  you would have to concern yourself with both possibilities past that point?

Exactly, most things are a Schrodinger's something or other, we just don't tend to see it that way because the chances aren't always 50/50. If an intelligence is general enough, I think it will use close to it's entire (unique to itself) dataset to interpret an input, and generate an output weighed against the full range of probable/impactful effects.

Once an AI has reached that level, how can someone with a different perspective (this will be everyone else) effectively judge the intelligence of its outputs relative to the inputs... At some point, I think we'll just have to go by the results. Which leads to the question, what results should we care about? For me at least, the two main ones would have to be environmental symbiosis and novelty generation.

Sounds good, definitely could work, I wish u good luck!

Thanks, I'm gonna need it lol. Even my brother who's in elementary school recently told me ''No offence, but [edit]: ''I'' don't think you're gonna be able to make artificial intelligence. I mean, it's artificial intelligence! That's literally like the hardest thing there is!"

Edit 2: Just to set the record straight, my brother and I are best friends. He recently bought me Skyrim with his own money, after asking me what video games I thought were interesting. He'd been earning it by making arts and crafts at school for the past month. I just think some of his passing comments are quite funny because of his honesty, so I wanted to share.
« Last Edit: August 16, 2021, 05:55:59 pm by HS »

*

MagnusWootton

  • Replicant
  • ********
  • 634
Re: Interesting mailing lists
« Reply #9 on: August 15, 2021, 02:36:54 pm »
Another example of the shrodinger's cat problem,  is a locked door,  there is no visible way to tell if the door is locked or open until you test the handle.
So the robot is this example will simulate the door as if it was unlocked,  and the door as if it was locked, but theres got to be some factor as well for having it barred not knowing the result,  before it bothers to even pick these chains because it actually doesnt know if the door is locked or not.
Maybe it decides its more important to get the key,   or maybe it needs to at least test the door before it knows to get the key, etc etc. ad absurdum. :)
Its easier if you dont bother at all and only deal with 100% guaranteed  8) routes.  But if you dont detect it at all, it would be like a fly against a window,  no effect from stimulus. =)

But thats not all,   it happens to an embodied robot with cameras for eyes all the time!    anything outside its view becomes a hidden variable,  so it cant know for sure whats there, or what was there before is still there,  or moved.

So maybe the "hidden geometry" problem is very important, if the robot is projecting for the truth in front in time, and I might put some more thought into it.

Even making something well below human intelligence could probably blow the world up,  dont worry about non believers, they are just busy not believing in everyone including themselves.

*

MagnusWootton

  • Replicant
  • ********
  • 634
Re: Interesting mailing lists
« Reply #10 on: August 15, 2021, 04:05:22 pm »
environmental symbiosis and novelty generation.

u mean by environmental symbiosis,   it deriving a model which agrees from the sensory report from the environment right?

And Novelty generation is it using the environmental symbiosis in its favour - to get what it wants?

If thats the case, it is also what I'm doing,  but if it isn't it would be cool if you could clarify.

*

HS

  • Trusty Member
  • **********
  • Millennium Man
  • *
  • 1175
Re: Interesting mailing lists
« Reply #11 on: August 15, 2021, 06:07:23 pm »
environmental symbiosis and novelty generation.

I've got quite a busy day ahead (yeah it's the weekend, but ''life'' happened), but I'll think about how to best explain it, and get back to you when I can.  O0

*

HS

  • Trusty Member
  • **********
  • Millennium Man
  • *
  • 1175
Re: Interesting mailing lists
« Reply #12 on: August 16, 2021, 11:44:14 pm »
When I think about it, these are probably the two things I value the most; first that the environment is basically agreeable, providing sentient life with a fair chance at having their basic needs met, and second that cool new stuff still happens. So I’m trying to figure out how to get an AI to exemplify/instantiate those qualities.

By environmental symbiosis I mean that an intelligence would, in a sustainable manner, contribute positively to the sum total of experience within its sphere of influence. You know, the whole range of things which we find important; from things like not polluting the oceans with harmful plastics, to being a good friend to others, to being good to yourself.

I explored how this might be accomplished in my Human Style AGI thread, though I called it environmental integration, but it's the same thing really. Like you said, it's based on accurate environmental modeling, specifically the cataloging of the positive/arbitrary/negative effects (possibly I could do it like ivan’s idea of -1,0,1 receptors which later combine) on the system(s), which can lead to perception, moods, imagination, and the enactment of narrative progressions.

By novelty generation I mean the capacity to produce genuinely new things, not just recombine existing things. I think this would be for environmental symbiosis. It would probably require harnessing of chaos/entropy/disorder to create new forms of order. The mere recombination of previously existing things seems insufficient to explain all that we have here.

At this point I can see a flaw in my theory… All writing for example, on one level, is just a recombination of existing letters, and more fundamentally, everything new could be considered as just a recombination of atoms. But I think there is a distinction to be made between recombination and creation. The existence of one uninspiring perspective does not preclude the existence of other more interesting and functionally representative ways of looking at things.

I believe true originality, novelty, and growth cannot be achieved through fully controlled methods, because current mainstream conceptions of intelligence appear self-limiting in that way. They are still below a certain singularity, a self-generative threshold. I think it's a similar boundary that people working on constructor theory are attempting to cross.

The destructive/chaotic/entropic principle may be necessary to create entirely new puzzle pieces for intelligent processes to then work with. Quickly evolving forms of nature might be places specifically designed for chaos and entropy to occur, so that the surrounding order can then make use of it’s outputs to complexify itself.

The resulting increase in complexity then creates more entropic potential, and on it goes, with the two principles propelling each other for as long as they can maintain balance. For an AI to have the same type of internally generated novelty in its thinking could depend on incorporating this dichotomy into its cognition. This is what an evolving rule set for neural behavior/connectivity might help with.
« Last Edit: August 17, 2021, 03:26:05 am by HS »

*

MagnusWootton

  • Replicant
  • ********
  • 634
Re: Interesting mailing lists
« Reply #13 on: August 17, 2021, 06:10:37 pm »
So it is a form of fixed function goal/motivation system for an AGI?

If so,  you need a function that will score the situation of the robot 0-1000,   You need this function to be able to get it to work for real.

*

HS

  • Trusty Member
  • **********
  • Millennium Man
  • *
  • 1175
Re: Interesting mailing lists
« Reply #14 on: August 17, 2021, 08:26:44 pm »
I'm trying to get the necessary starting elements into place so all of this will be the result of an unconstrained motivational system. I'm thinking that if the kind of AGI I'm envisioning has certain fundamental sensory capabilities and neural reactions, the rest of the functions should develop naturally without the need for instructions.

 


OpenAI Speech-to-Speech Reasoning Demo
by ivan.moony (AI News )
Today at 01:31:53 pm
Say good-bye to GPUs...
by MikeB (AI News )
March 23, 2024, 09:23:52 am
Google Bard report
by ivan.moony (AI News )
February 14, 2024, 04:42:23 pm
Elon Musk's xAI Grok Chatbot
by MikeB (AI News )
December 11, 2023, 06:26:33 am
Nvidia Hype
by 8pla.net (AI News )
December 06, 2023, 10:04:52 pm
How will the OpenAI CEO being Fired affect ChatGPT?
by 8pla.net (AI News )
December 06, 2023, 09:54:25 pm
Independent AI sovereignties
by WriterOfMinds (AI News )
November 08, 2023, 04:51:21 am
LLaMA2 Meta's chatbot released
by 8pla.net (AI News )
October 18, 2023, 11:41:21 pm

Users Online

296 Guests, 0 Users

Most Online Today: 346. Most Online Ever: 2369 (November 21, 2020, 04:08:13 pm)

Articles