Re: [agi] Re: Call for Models: Working Memory Modelathon 2020

[Ben Goertzel]

Gary Marcus's article explains quite clearly why and how GPT2 fails to
approach human-like AGI,

https://thegradient.pub/gpt2-and-the-nature-of-intelligence/

He also explains the fallacy of simplistically claiming that
prediction = understanding

The merits or demerits of OpenCog are a different question.   If I had
none of my own ideas about AGI, or OpenCog did not exist, etc.,
Marcus's critique of GPT2 as relates to AGI would still be very apt
and I would still agree with it.

Whether GPT2 is "Closer to AGI" than other current AI systems is a red
herring IMO.   It's like discussing whether a blimp or a glider is
closer to a starship ... or whether a cat or dog is closer to human
intelligence.   Who cares?   (And if we wanted to pursue this metaphor
-- OpenCog is then like a partly-built starship that is designed
according to a proper theory of starships, but doesn't yet fly as high
as the blimp or glider.   But as I said, I don't want to make this a
comparison of OpenCog vs. GPT2 which are completely different sorts of
animals.   The two things should be discussed separately.)

Chollet has made similar points in a style designed to be palatable to
folks from the contemporary ML/DL school of AI,

https://arxiv.org/pdf/1911.01547.pdf


-- Ben G


On Tue, Jul 7, 2020 at 10:29 PM  wrote:
>
> Make sure to read my above post.
>
> Really? You don't see how Blender (or my improvement above) is closer to AGI 
> than GPT-2 is? Or that GPT-2 is close-ish to AGI? Do you have something 
> better? Does it predict text/images better? What does OpenCog AI do if it 
> can't compare to OpenAI's showcase!? Either you have better results showing 
> it can think more like AGI than my vision of Blender or you haven't coded it 
> yet and can explain it instead, but as far as I can see, OpenCog AI isn't as 
> "AGI" as Blender or my vision of Blender. Explain OpenCog. Why doesn't it 
> just recognize a sentence/image patch and predict the next item? What does it 
> do? I can't even find out. AGI is just recognition, prediction, attention, it 
> has to create/predict the future thoughts/discoveries it desires. Prediction 
> is 90% of AGI and all new AGI mechanisms simply improve it or allow it to do 
> interesting manipulation on data ex. logic AND OR or when you edit a Paper or 
> delete things or clone paragraphs, or count how many letters in a sentence or 
> how many times 's' appears in a sentence, etc etc. These things are AGI like 
> 'behavior', but still connected to Prediction and can be showcased.
>
> GPT-2 is obviously very close to the AGI we're looking for. And Blender is 
> even better because 1) it knows how to finish its replies (because of Byte 
> Pair Encoding), 2) is trained not just on wikipedia but actual human-like 
> chat log conversations, 3) is forced to talk about/stick to some 
> domain/question and not just some nutty prompt like unicorns (which changes 
> over time too) which is not about cancer or immortality (and my improvement 
> is to set it to force it to talk about Survival, AND evolve it as explained 
> how to let it find the right sub domains of sub domains as it thinks. This 
> let's it much better answer the installed question/goal). There's no better 
> than this. It's closer to natural AGI than all others.
> Artificial General Intelligence List / AGI / see discussions + participants + 
> delivery options Permalink



-- 
Ben Goertzel, PhD
http://goertzel.org

“The only people for me are the mad ones, the ones who are mad to
live, mad to talk, mad to be saved, desirous of everything at the same
time, the ones who never yawn or say a commonplace thing, but burn,
burn, burn like fabulous yellow roman candles exploding like spiders
across the stars.” -- Jack Kerouac

--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T604ad04bc1ba220c-M05d283814bf9533ee64f28f7
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Re: [agi] Re: Call for Models: Working Memory Modelathon 2020

[immortal . discoveries]

Make sure to read my above post.

Really? You don't see how Blender (or my improvement above) is closer to AGI 
than GPT-2 is? Or that GPT-2 is close-ish to AGI? Do you have something better? 
Does it predict text/images better? What does OpenCog AI do if it can't compare 
to OpenAI's showcase!? Either you have better results showing it can think more 
like AGI than my vision of Blender or you haven't coded it yet and can explain 
it instead, but as far as I can see, OpenCog AI isn't as "AGI" as Blender or my 
vision of Blender. Explain OpenCog. Why doesn't it just recognize a 
sentence/image patch and predict the next item? What does it do? I can't even 
find out. AGI is just recognition, prediction, attention, it has to 
create/predict the future thoughts/discoveries it desires. Prediction is 90% of 
AGI and all new AGI mechanisms simply improve it or allow it to do interesting 
manipulation on data ex. logic AND OR or when you edit a Paper or delete things 
or clone paragraphs, or count how many letters in a sentence or how many times 
's' appears in a sentence, etc etc. These things are AGI like 'behavior', but 
still connected to Prediction and can be showcased.

GPT-2 is obviously very close to the AGI we're looking for. And Blender is even 
better because 1) it knows how to finish its replies (because of Byte Pair 
Encoding), 2) is trained not just on wikipedia but actual human-like chat log 
conversations, 3) is forced to talk about/stick to some domain/question and not 
just some nutty prompt like unicorns (which changes over time too) which is not 
about cancer or immortality (and my improvement is to set it to force it to 
talk about Survival, AND evolve it as explained how to let it find the right 
sub domains of sub domains as it thinks. This let's it much better answer the 
installed question/goal). There's no better than this. It's closer to natural 
AGI than all others.
--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T604ad04bc1ba220c-M3e2fdd339d11df1387c470aa
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Re: [agi] Formal Language Theory Has Its Head Up Its Ass

[John Rose]

Need something akin to an LED, a Qualia Emitting Diode (QED), bidirectional... 
like a DIAC to communicate direct consciousness using a concept language, time 
symmetric. An array of quantum dots as conscious connect coupled to nervous 
system excited states... dots to DIAC to DIAC to dots... or some circuit like 
that. Essentially transmit concept consciousness between parties, computer or 
human. But is that considered language since language implies time asymmetric 
serialization with some sort of symbol virtualization verses deeply linked and 
directly coupled concept graphs.

Similar to two people staring each other in the eye, speaking without words 
which natural language has reduced the need for in our recent evolution...

Though when you think about it the best would be partially entangled molecular 
structure between parties so thoughts are immediately linked with minimal 
latency. There would still be a serialization I suppose since it would be a 
subset of operating atoms entangled... When DNA splits does it leave any 
sustained entanglement?

But partially entangled molecules would prefer a particular communication 
structure that would optimize on it. The symbol set would virtualize up from 
there but be compressed down from higher up for maximally efficient 
transmission and complexity representation... though IMO optimal multiparty 
communication among general intelligences is multichannel. Perhaps an optimal 
language is "multichannel" capable, or multicontextual/multitargeted and.. 
multidirectional.  IOW allow communication to many targets in many contexts in 
the same concept "sentence" multiduplexed... essentially eliminate transmission 
medium, you think it and we all know... kind of like... Sheldrake's Morphic 
Fields. The optimal language is ... no language???
--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T044a258b4a99cfd7-Ma27b146705a3041a9f48adc7
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Re: [agi] Experimental Testing of CIC (the Compression Information Criterion)

[James Bowery]

On Tue, Jul 7, 2020 at 2:31 PM Matt Mahoney  wrote:

> Why bother with a CIC training and test set? Compression evaluates every
> bit as a test given the previous bits as training. Even if the compression
> algorithm doesn't explicitly predict bits, it is equivalent to one that
> does by the chain rule. The probability of a string is equal to the product
> of the conditional probabilities of its symbols.
>
> You can see this effect at work in http://mattmahoney.net/dc/text.html
> The ranking of enwik8 (first 100 MB) closely tracks the ranking of enwik9.
> Most of the variation is due to memory constraints. In small memory models,
> compression is worse overall and closer to the result you would get from
> compressing the parts independently.
>
> Occam's Razor doesn't necessarily hold under constrained resources. All
> probability distributions over an infinite set of strings must favor
> shorter ones, but that isn't necessarily true over the finite set of
> programs that can run on a computer with finite memory.
>

Yes, and that is the most vocal of Ben's critiques of what we're calling
now (I guess) *The COIN Hypothesis* (however much of a stretch it is to get
to the memetic advantage of that acronym).  To reiterate that hypothesis
with a little more nuance and refinement including emphasis on resource
constraint:

*The COIN (COmpression Information criterioN) Hypothesis is about the
empirical world *and is *both* that among:

   1. existing models of a process, the one producing the smallest
   executable archive of the same training data and within *the same
   computation constraints*, will *generally* also produce the smallest
   executable archive of the test data, AND
   2. *all* model selection criterion, will do so *more* *generally*.

More to the point, whenever I discuss COIN as *the* model selection
criterion, the obvious fact that it isn't...

   - a mathematically provable aspect of "the unreasonable effectiveness of
   mathematics in the natural sciences"
   - empirically tested (although it seems measurable)
   - widespread, nor even minority use

...people react in one of 3 ways, in order of frequency:

   1. Huh?  Wha?  Fuhgeddaboudit.
   2. Where's the empirical evidence?
   3. Minimum Description Length Principle is just the Bayesian Information
   Criterion.
   4. You're just plain wrong because _insert some invalid critique_.

Indeed, the research program I set forth should be pursued if for no other
reason than to rank order the general practicality of various model
selection criteria.

On Tue, Jul 7, 2020 at 2:31 PM Matt Mahoney  wrote:

> Why bother with a CIC training and test set? Compression evaluates every
> bit as a test given the previous bits as training. Even if the compression
> algorithm doesn't explicitly predict bits, it is equivalent to one that
> does by the chain rule. The probability of a string is equal to the product
> of the conditional probabilities of its symbols.
>

The practice of dividing training and test data is standard industry
practice.  Why should it not be pursued in this instance since the point is
to convince people of the truth or falsity of COIN?

--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Ta901988932dbca83-M9076d4fa7ac88c4402052595
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Re: [agi] Re: Call for Models: Working Memory Modelathon 2020

[Ben Goertzel]

Yeah we @ SingularityNET have been using Blender, and conditioning
Blender on other specialized corpora, in some application work.
However I don't see how this is directly useful for AGI, though it's
cool for narrow-AI application work...

On Tue, Jul 7, 2020 at 5:27 AM  wrote:
>
> Have you seen PPLM, CTRL, and Blender? They all do the same thing but are an 
> improvement on GPT-2. Blender is the farthest, it both controls the 
> generation, plus is trained on chat logs, wiki, and empathy, plus finishers 
> its reply to you.
>
> I can build on Blender. No one yet has realized my achievement so I'll only 
> explain it if you actually really want AGI.
> Artificial General Intelligence List / AGI / see discussions + participants + 
> delivery options Permalink



-- 
Ben Goertzel, PhD
http://goertzel.org

“The only people for me are the mad ones, the ones who are mad to
live, mad to talk, mad to be saved, desirous of everything at the same
time, the ones who never yawn or say a commonplace thing, but burn,
burn, burn like fabulous yellow roman candles exploding like spiders
across the stars.” -- Jack Kerouac

--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T604ad04bc1ba220c-M6399eeaa263f3b102e2fb90b
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Re: [agi] Experimental Testing of CIC (the Compression Information Criterion)

[Matt Mahoney]

Why bother with a CIC training and test set? Compression evaluates every
bit as a test given the previous bits as training. Even if the compression
algorithm doesn't explicitly predict bits, it is equivalent to one that
does by the chain rule. The probability of a string is equal to the product
of the conditional probabilities of its symbols.

You can see this effect at work in http://mattmahoney.net/dc/text.html
The ranking of enwik8 (first 100 MB) closely tracks the ranking of enwik9.
Most of the variation is due to memory constraints. In small memory models,
compression is worse overall and closer to the result you would get from
compressing the parts independently.

Occam's Razor doesn't necessarily hold under constrained resources. All
probability distributions over an infinite set of strings must favor
shorter ones, but that isn't necessarily true over the finite set of
programs that can run on a computer with finite memory.

On Tue, Jul 7, 2020, 12:13 AM James Bowery  wrote:

> On Fri, Jul 3, 2020 at 6:53 PM Ben Goertzel  wrote:
>
>> ...Under what conditions is it the case that, for prediction based on a
>> dataset using realistically limited resources, the smallest of the
>> available programs that precisely predicts the training data actually gives
>> the best predictions on the test data?
>
>
> If I may refine this a bit to head off misunderstanding at the outset of
> this project:
>
> The CIC* (Compression Information Criterion) hypothesis is that among
> existing models of a process producing an executable archive of the same
> training data within the same computation constraints, the one that
> produces the smallest executable archive will in general be the most
> accurate on the test data.
>
>
> Run a number of experiments and for each:
> 1 Select a nontrivial
> 1.1 computational resource level as constraint
> 1.2 real world dataset -- no less than 1GB gzipped.
> 2 Divide the data into training and testing sets
> 3 For each competing model:
> 3.1 Provide the training set
> 3.2 Record the length of the executable archive the model produces
> 3.3 Append the test set to the training set
> 3.4  Record the length of the executable archive the model produces
> 4 Produce 2 rank orders for the models
> 4.1 training set executable archive sizes
> 4.2 training with testing set executable archive sizes
> 5 Record differences in the training vs test rank orders
>
> The lower the average differences the more general the criterion.
>
> It should be possible to run similar tests of other model selection
> criteria and rank order model selection criteria.
>
> *We're going to need a catchy acronym to keep up with:
>
> AIC (Akaike Information Criterion)
> BIC (Bayesian Information Criterion)...
> ...aka
> SIC (Schwarz Information Criterion)...
> ...aka
> MDL or MDLP (both travestic abuses of "Minimum Description Length
> [Principle]" that should be forever cast into the bottomless pit)
> HQIC (Hannan-Quinn Information Criterion)...
> KIC (Kullback Information Criterion)
> etc. etc.
> *Artificial General Intelligence List *
> / AGI / see discussions  +
> participants  + delivery
> options  Permalink
> 
>

--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Ta901988932dbca83-Md0356e370e6fb9b9f01847eb
Delivery options: https://agi.topicbox.com/groups/agi/subscription

[agi] Re: Call for Models: Working Memory Modelathon 2020

[immortal . discoveries]

You will get that in my upcoming guide but for now try this explanation (2 
parts to it):


ROOT FORCE: I'll trust yous already know GPT-2 and the even cooler Blender. My 
discovery to improve Blender is: These AIs collect lots of diverse/general data 
(explores), but lots of it doesn't answer it's main question(s) or "domain of 
choice". Your AGI should have a domain like cancer research questions for 
example (preferably survival), forced (as they do in Blender), which causes it 
to talk most the day about cancer and therefore [generate data about cancer 
from the context around it] and also [collect cancer data off the internet] 
(like a cancer dataset). Exploitation. This is akin to feeding your AGI more 
data, because it helps you answer your question of focus better.

EVOLVING THE ROOT: Your AGI should begin life with just a few words Forced (as 
they call it in Blender). Like food, sex, survival. For example "I will get 
maximum food by ". But it has to specialize more to better answer them. It 
won't get food if it doesn't become immortal, and food/reproduction is just 
trying to do that anyway. I mean it needs to invent new questions, not how do I 
get food/sex but how do I get money or job or immortality or AGI. The way it 
happens is food/sex is recognized semantically (by shared contexts) as money. 
Food/sex=money. Now it Forces the word/phrase "money"! It has a new focus/goal. 
It wants data around this context "money" now! Now it can further specialize by 
recognizing money node as job node by shared contexts. What this is is updating 
like a checkpoint where to pay attention to. It is translation, like semantics, 
you use similar experiences to answer your question, BUT, it's changing the 
question and changing what data it collects, from a specific domain source.


If you look at robot RL where they learn how to walk etc, they update/tweak 
their favorite moves that give them most acceleration/prediction accuracy. At 
some point it doesn't gain much at each new tweak and may even get stuck in 
local optima. Same for what I presented above.

When humans try real world experiments (lab, or even at home: living is an 
experiment!), we do it just to collect data! From a specific focused domain we 
desire. Our bodies are a feedback loop, we evolve where we collect data from, 
we update the test, we may change from cancer lab to AGI lab. Well, the brain 
can do the same thing. You don't need a body or world. Just data. Lots of data. 
You can find new discoveries in data. Once you do, you can update your Forced 
Goal dialog and search there to collect more new hidden data!

Robots that learn to walk using say, Transformers, is nothing compared to 
something like GPT-2, because text/images model/describe Earth much more 
expressively than walking does. So stay away from robots learning to walk 
And the real lab tests mentioned, same, stay away, it is not they key to AGI, 
it is not efficient or as powerful or flexible. I can plan to go to Mars in my 
brain, faster, safely, increase my tool size, shape, etc all in my brain 
images/text. I don't need lab to make discoveries, just big data and data 
extraction. Our universe/world only has a few patterns, so enough data can 
capture the whole world.
--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T604ad04bc1ba220c-Md38f878ceb52b4a02f031dfe
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Re: [agi] Experimental Testing of CIC (the Compression Information Criterion)

[Ben Goertzel]

COIN CON ;)

On Tue, Jul 7, 2020 at 9:22 AM James Bowery  wrote:
>
> A slight modification to avoid idioms like "the HIV virus":
> COINC = COIN Criterion
> COIN = COmpression Information criteriON
>
> Otherwise it would be good memetics if its psychological appeal vs memetic 
> drift reaches the selective regime in time to achieve fixation against the 
> psychological appeal of the *IC prior.
>
> On Tue, Jul 7, 2020 at 1:42 AM Ben Goertzel  wrote:
>>
>> The COIN Criterion ... sounds like money, it's got to be good...
>>
>> On Mon, Jul 6, 2020 at 9:13 PM James Bowery  wrote:
>> >
>> > On Fri, Jul 3, 2020 at 6:53 PM Ben Goertzel  wrote:
>> >>
>> >> ...Under what conditions is it the case that, for prediction based on a 
>> >> dataset using realistically limited resources, the smallest of the 
>> >> available programs that precisely predicts the training data actually 
>> >> gives the best predictions on the test data?
>> >
>> >
>> > If I may refine this a bit to head off misunderstanding at the outset of 
>> > this project:
>> >
>> > The CIC* (Compression Information Criterion) hypothesis is that among 
>> > existing models of a process producing an executable archive of the same 
>> > training data within the same computation constraints, the one that 
>> > produces the smallest executable archive will in general be the most 
>> > accurate on the test data.
>> >
>> >
>> > Run a number of experiments and for each:
>> > 1 Select a nontrivial
>> > 1.1 computational resource level as constraint
>> > 1.2 real world dataset -- no less than 1GB gzipped.
>> > 2 Divide the data into training and testing sets
>> > 3 For each competing model:
>> > 3.1 Provide the training set
>> > 3.2 Record the length of the executable archive the model produces
>> > 3.3 Append the test set to the training set
>> > 3.4  Record the length of the executable archive the model produces
>> > 4 Produce 2 rank orders for the models
>> > 4.1 training set executable archive sizes
>> > 4.2 training with testing set executable archive sizes
>> > 5 Record differences in the training vs test rank orders
>> >
>> > The lower the average differences the more general the criterion.
>> >
>> > It should be possible to run similar tests of other model selection 
>> > criteria and rank order model selection criteria.
>> >
>> > *We're going to need a catchy acronym to keep up with:
>> >
>> > AIC (Akaike Information Criterion)
>> > BIC (Bayesian Information Criterion)...
>> > ...aka
>> > SIC (Schwarz Information Criterion)...
>> > ...aka
>> > MDL or MDLP (both travestic abuses of "Minimum Description Length 
>> > [Principle]" that should be forever cast into the bottomless pit)
>> > HQIC (Hannan-Quinn Information Criterion)...
>> > KIC (Kullback Information Criterion)
>> > etc. etc.
>> > Artificial General Intelligence List / AGI / see discussions + 
>> > participants + delivery options Permalink
>> 
>> 
>> --
>> Ben Goertzel, PhD
>> http://goertzel.org
>> 
>> “The only people for me are the mad ones, the ones who are mad to
>> live, mad to talk, mad to be saved, desirous of everything at the same
>> time, the ones who never yawn or say a commonplace thing, but burn,
>> burn, burn like fabulous yellow roman candles exploding like spiders
>> across the stars.” -- Jack Kerouac
>
> Artificial General Intelligence List / AGI / see discussions + participants + 
> delivery options Permalink



-- 
Ben Goertzel, PhD
http://goertzel.org

“The only people for me are the mad ones, the ones who are mad to
live, mad to talk, mad to be saved, desirous of everything at the same
time, the ones who never yawn or say a commonplace thing, but burn,
burn, burn like fabulous yellow roman candles exploding like spiders
across the stars.” -- Jack Kerouac

--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Ta901988932dbca83-M58ea1fc2de53183a853f3b1b
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Re: [agi] Experimental Testing of CIC (the Compression Information Criterion)

[James Bowery]

I'm sure the Hollywood meme machine isn't worried about such quibbles as
the "HIV virus" idiom populating its marquees.

On Tue, Jul 7, 2020 at 1:43 AM Ben Goertzel  wrote:

> On Mon, Jul 6, 2020 at 11:41 PM Ben Goertzel  wrote:
> >
> > The COIN Criterion ... sounds like money, it's got to be good...
>
> Maybe we can fund the competition by making a Hollywood-style thriller
> about some Bitcoin criminals... lots of potentials here...
>
> >
> > On Mon, Jul 6, 2020 at 9:13 PM James Bowery  wrote:
> > >
> > > On Fri, Jul 3, 2020 at 6:53 PM Ben Goertzel  wrote:
> > >>
> > >> ...Under what conditions is it the case that, for prediction based on
> a dataset using realistically limited resources, the smallest of the
> available programs that precisely predicts the training data actually gives
> the best predictions on the test data?
> > >
> > >
> > > If I may refine this a bit to head off misunderstanding at the outset
> of this project:
> > >
> > > The CIC* (Compression Information Criterion) hypothesis is that among
> existing models of a process producing an executable archive of the same
> training data within the same computation constraints, the one that
> produces the smallest executable archive will in general be the most
> accurate on the test data.
> > >
> > >
> > > Run a number of experiments and for each:
> > > 1 Select a nontrivial
> > > 1.1 computational resource level as constraint
> > > 1.2 real world dataset -- no less than 1GB gzipped.
> > > 2 Divide the data into training and testing sets
> > > 3 For each competing model:
> > > 3.1 Provide the training set
> > > 3.2 Record the length of the executable archive the model produces
> > > 3.3 Append the test set to the training set
> > > 3.4  Record the length of the executable archive the model produces
> > > 4 Produce 2 rank orders for the models
> > > 4.1 training set executable archive sizes
> > > 4.2 training with testing set executable archive sizes
> > > 5 Record differences in the training vs test rank orders
> > >
> > > The lower the average differences the more general the criterion.
> > >
> > > It should be possible to run similar tests of other model selection
> criteria and rank order model selection criteria.
> > >
> > > *We're going to need a catchy acronym to keep up with:
> > >
> > > AIC (Akaike Information Criterion)
> > > BIC (Bayesian Information Criterion)...
> > > ...aka
> > > SIC (Schwarz Information Criterion)...
> > > ...aka
> > > MDL or MDLP (both travestic abuses of "Minimum Description Length
> [Principle]" that should be forever cast into the bottomless pit)
> > > HQIC (Hannan-Quinn Information Criterion)...
> > > KIC (Kullback Information Criterion)
> > > etc. etc.
> > > Artificial General Intelligence List / AGI / see discussions +
> participants + delivery options Permalink
> >
> >
> >
> > --
> > Ben Goertzel, PhD
> > http://goertzel.org
> >
> > “The only people for me are the mad ones, the ones who are mad to
> > live, mad to talk, mad to be saved, desirous of everything at the same
> > time, the ones who never yawn or say a commonplace thing, but burn,
> > burn, burn like fabulous yellow roman candles exploding like spiders
> > across the stars.” -- Jack Kerouac
> 
> 
> --
> Ben Goertzel, PhD
> http://goertzel.org
> 
> “The only people for me are the mad ones, the ones who are mad to
> live, mad to talk, mad to be saved, desirous of everything at the same
> time, the ones who never yawn or say a commonplace thing, but burn,
> burn, burn like fabulous yellow roman candles exploding like spiders
> across the stars.” -- Jack Kerouac

--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Ta901988932dbca83-M198483120a43ae72ba94809a
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Re: [agi] Experimental Testing of CIC (the Compression Information Criterion)

[James Bowery]

A slight modification to avoid idioms like "the HIV virus":
COINC = COIN Criterion
COIN = COmpression Information criteriON

Otherwise it would be good memetics if its psychological appeal vs memetic
drift reaches the selective regime in time to achieve fixation against the
psychological appeal of the *IC prior.

On Tue, Jul 7, 2020 at 1:42 AM Ben Goertzel  wrote:

> The COIN Criterion ... sounds like money, it's got to be good...
>
> On Mon, Jul 6, 2020 at 9:13 PM James Bowery  wrote:
> >
> > On Fri, Jul 3, 2020 at 6:53 PM Ben Goertzel  wrote:
> >>
> >> ...Under what conditions is it the case that, for prediction based on a
> dataset using realistically limited resources, the smallest of the
> available programs that precisely predicts the training data actually gives
> the best predictions on the test data?
> >
> >
> > If I may refine this a bit to head off misunderstanding at the outset of
> this project:
> >
> > The CIC* (Compression Information Criterion) hypothesis is that among
> existing models of a process producing an executable archive of the same
> training data within the same computation constraints, the one that
> produces the smallest executable archive will in general be the most
> accurate on the test data.
> >
> >
> > Run a number of experiments and for each:
> > 1 Select a nontrivial
> > 1.1 computational resource level as constraint
> > 1.2 real world dataset -- no less than 1GB gzipped.
> > 2 Divide the data into training and testing sets
> > 3 For each competing model:
> > 3.1 Provide the training set
> > 3.2 Record the length of the executable archive the model produces
> > 3.3 Append the test set to the training set
> > 3.4  Record the length of the executable archive the model produces
> > 4 Produce 2 rank orders for the models
> > 4.1 training set executable archive sizes
> > 4.2 training with testing set executable archive sizes
> > 5 Record differences in the training vs test rank orders
> >
> > The lower the average differences the more general the criterion.
> >
> > It should be possible to run similar tests of other model selection
> criteria and rank order model selection criteria.
> >
> > *We're going to need a catchy acronym to keep up with:
> >
> > AIC (Akaike Information Criterion)
> > BIC (Bayesian Information Criterion)...
> > ...aka
> > SIC (Schwarz Information Criterion)...
> > ...aka
> > MDL or MDLP (both travestic abuses of "Minimum Description Length
> [Principle]" that should be forever cast into the bottomless pit)
> > HQIC (Hannan-Quinn Information Criterion)...
> > KIC (Kullback Information Criterion)
> > etc. etc.
> > Artificial General Intelligence List / AGI / see discussions +
> participants + delivery options Permalink
> 
> --
> Ben Goertzel, PhD
> http://goertzel.org
> 
> “The only people for me are the mad ones, the ones who are mad to
> live, mad to talk, mad to be saved, desirous of everything at the same
> time, the ones who never yawn or say a commonplace thing, but burn,
> burn, burn like fabulous yellow roman candles exploding like spiders
> across the stars.” -- Jack Kerouac

--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Ta901988932dbca83-M6c288e33fc799150fe51c972
Delivery options: https://agi.topicbox.com/groups/agi/subscription

[agi] Re: Call for Models: Working Memory Modelathon 2020

[keghnfeem]

 Pleas explain it like i am a fiver year old.
--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T604ad04bc1ba220c-Mc75ef5f73e57ee87f61224b3
Delivery options: https://agi.topicbox.com/groups/agi/subscription

[agi] Mentifex Communicat 2020-07-07

[A.T. Murray]

Yesterday on Cable News Network (CNN) the president of my Alma Mater was
talking about the one hundred thirty-seven (137) fratboys on Greek Row
who recently caught the coronavirus during their incessant parties and
keggers and Animal House orgies. She bragged about how back in March 2020
she was first in the nation to close down a major American university for
the virus and how all the other universities followed her lead. I was on
campus that day on Friday 2020-03-06 and I had no idea that I would soon
have to give up the sweet lifestyle of re-living my college years. Now four
months later there is no end in sight to the coronavirus pandemic and all I
have
is things like the https://agi.io AGI project to sustain me.

Amazing things did happen last week and "si vacat et placidi rationem
admittitis, incipiam," which means "if there is time and if you placidly
accept reason, I will begin." It is a perhaps garbled quote from a book in
Latin which
I bought as a teenage boy and really loved reading in Latin. That book, the
Satires of Juvenal, had absolut-and-I-don't-mean-vodka-ly priceless
expressions such as "Difficle est, satyram non scribere" ("It is difficult
NOT to write a satire") and "Orandum est, ut sit mens sana in corpore sano"
which became the motto of every American high-school athletics coach.

You see, we are all bored to death during this pandemic, and so we are
doing the Internet equivalent of passing notes in school by posting to the
preeminent AGI Mail-List in these parsecs of the Universe. The list owner
Ben Goertzel or the list moderator John Rose catches me folding up a piece
of paper and sternly orders me to read the Liebesbrief to the entire class.

"Dear (codename) Daisy (choker) Russian spy: The University closed down so
quickly that I could not get your e-mail or telephone number or dead-drop
box-address to stay in touch with you. So please read between the lines in
these purported AGI messages."

In the sheer boredom of waiting out the coronavirus, I was posting to
Reddit so heavily last week that I started to get 800 or 900 page-views per
day on my Latin-language AGI pages. I would log in and see "You have new
messages" pop up on-screen as an alert to me, but in eight years of
Redditing I have never once opened a Private Message (PM) for fear of
seeing a bone-chilling threat from some evildoer's anonymous throw-away
account. There could
be people eager to collaborate with me on AGI, but I don't dare open a
single Private Message. I mean, when I worked the night shift at a hotel
twenty years ago and the phone would ring at 2:00 a.m, I learned to say
nothing at first, and it would thwart and stymie the guy who was saying to
me, "I'm going to cut you. I'm going to drink your blood." There may be
similar Private Messages
waiting for me on Reddit, but my ignorance of them is bliss.

So last week was Latin Week, and now this week is Russian Week. I recently
spent two days translating my Russian AI User Manual back into its original
English. Seven years ago I had paid my Russian translator to express my
ideas in perfect Russian, but he had changed some English expressions into
more natural Russian expressions. So last week I translated his Russian
version back into http://ai.neocities.org/RuAiUser.html so that my English
wording would match his Russian wording. I also used the most up-to-date
links within the English and Russian versions of the AI User Manual.

Even on days when I post nothing to Reddit, the Live Traffic Feed report
shows that people around the world are now randomly accessing either the
Latin AGI pages or the Russian AGI pages. On Usenet, there are newsgroups
where Mentifex-bashers were pouncing viciously until quite recently,
whenever I dared to post about Mentifex AI. Now there is an awesome silence
and I get away with posting the most outrageous ESC -- Extraordinary
Scientific Claim.

http://medium.com/p/e78d959117af -- 2019-11-01 Fri.
How each Mentifex Strong AI Mind Thinks with Conjunctions

http://www.mail-archive.com/agi@agi.topicbox.com

--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T98ab1a42c270a017-M0958c3a3c919ab05c18a084f
Delivery options: https://agi.topicbox.com/groups/agi/subscription

[agi] Re: Call for Models: Working Memory Modelathon 2020

[immortal . discoveries]

Have you seen PPLM, CTRL, and Blender? They all do the same thing but are an 
improvement on GPT-2. Blender is the farthest, it both controls the generation, 
plus is trained on chat logs, wiki, and empathy, plus finishers its reply to 
you.

I can build on Blender. No one yet has realized my achievement so I'll only 
explain it if you actually *really *want AGI.
--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T604ad04bc1ba220c-Md789fffa84067cfad261dd2b
Delivery options: https://agi.topicbox.com/groups/agi/subscription

[agi] Call for Models: Working Memory Modelathon 2020

[ARAKAWA Naoya]

(Please allow me to 'call' here.)

The Whole Brain Architecture Initiative, the NPO, is planning the 5th WBA 
Hackathon 
with the theme of working memory, which is considered to be the keystone for 
'fluid' 
cognitive functions such as planning, and thus for AGI.

Prior to the hackathon, it holds a modelathon to call for computational 
cognitive 
neuroscientific models of working memory.  Models that are evaluated highly by 
an academic review will be awarded and grants up to 50,000 yen will be offered.
Deadline: September 30th
https://wba-initiative.org/en/15968/


-- Naoya Arakawa@WBAI



--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/T604ad04bc1ba220c-Mf1b909aacad3291ee6671df8
Delivery options: https://agi.topicbox.com/groups/agi/subscription

Re: [agi] Formal Language Theory Has Its Head Up Its Ass

[Ben Goertzel]

Wow, fascinating history!

I know both Dean Radin and Ed May ... I haven't been active in
empirical psi research but as you may know I edited the book "Evidence
for Psi" with Damien Broderick a few years back, and have been lurking
around the parapsychology community for quite a while...

This paper

https://arxiv.org/abs/2005.04589

which is presented as AGI theory, actually originated in some
discussions at the Psi Theory workshop held last year in Paris before
the Parapsychology Association conference ... Dean was there though
not Ed ... one question raised there was: If this physical universe is
part of a higher-dimensional trans-physical world, what might be the
"Hamiltonian" or other broadly-physics-ish dynamics of that broader
world (which I call "euryphysical").   The algorithmic-information-al
dynamics described in the above paper was originated in that direction
though it is also relevant I think to AGI as in manifests in the
physical universe...

Starting at 49:00 or so of

https://www.youtube.com/watch?v=JzlCYCAaMNw

I gave a crude sketch of a path leading from Spencer Brown / Kauffmann
"distinction" toward AGI theory, with the above stuff as part of that
path

Regarding your comments on the primacy of time vs. distinction, I
think my intuition is more similar to Lou's ...

Principles 13-16 as articulated here

https://journals.sfu.ca/jnonlocality/index.php/jnonlocality/article/view/65

roughly outline how I suspect time and space emerge from underlying
phenomena of distinction, pattern, information and so forth...

-- Ben




On Sun, Jul 5, 2020 at 11:49 AM James Bowery  wrote:
>
> BTW:  As long as we're going down the "formal tools" rabbit hole, an 
> important bit of history bears mentioning.
>
> About the time HP kicked Tom and me out of Fiorina's $500M "Internet Chapter 
> 2" fiasco, Federico Faggin stepped up to the plate to fund the work.  
> Although Faggin wasn't directly aware of what Tom did under my sponsorship at 
> HP, he _was_ aware of the need for better formal tools in VLSI design, and 
> found imaginary logic an intriguing approach.  So he endowed The Boundary 
> Institute, which Richard Shoup founded and Shoup hired Tom as the theorist.  
> This work became side-tracked when Dean Radin and Edwin May showed up to 
> pursue psi research from the empirical side.  That was most unfortunate, not 
> because there can be no interpretation of QM under which "paranormal" 
> theories may be tested empirically, but because they paid little attention to 
> Tom's output and sucked all the air out of the room, so to speak.  So they 
> merely added more to the corpus of "evidence" for psi phenomena, without 
> doing anything to ground such experimental work on (meta)physical theory.  
> There was only one exception to this -- an experiment designed by Tom based 
> on his approach to QM -- but the experiment was postponed until Faggin's 
> original endowment was running out, not to be renewed due to inadequate 
> pursuit of VLSI tools.  By then Radin and May were looking for psi funding 
> with Radin ending up bolting for The Institute for Noetic Sciences, where it 
> was easier to raise such funds.  All in all it was a very bad decision Shoup 
> made letting those guys come in.  Both Tom and Dick were left without 
> material support pretty much for the rest of their lives. They both passed 
> away several years later.  I was barely able to salvage Tom's ANPA West 
> journal archives, and one last paper -- bringing everything together -- 
> progress on which stopped when dementia started setting in after his wife 
> passed and he followed her.
>
> What might the world look like today if Faggin had used his founding role at 
> Intel to introduce imaginary logic VLSI tools there?
>
> On Sun, Jul 5, 2020 at 12:32 PM James Bowery  wrote:
>>
>>
>>
>> On Sun, Jul 5, 2020 at 11:09 AM Ben Goertzel  wrote:
>>>
>>> As you perhaps know I am a big fan of imaginary logic, having started in 
>>> this direction due to some correspondences w/ Lou Kauffmann (G. Spencer 
>>> Brown's collaborator) in the mid-1980s ...
>>
>>
>> I'm not at all surprised which is why I said NOR DCGs should ring all kinds 
>> of bells with you.  But I am a _bit_ surprised I don't find you on the Laws 
>> of Form yahoo group mailing list, which I administer (although that mailing 
>> list is not very active).
>>
>> Lou Kauffman is prolific but has not, to the best of my knowledge, pursued 
>> Tom's construction of imaginary logics from real valued spinors.  I know he 
>> and Tom were very well acquainted because I had dinner with him at Tom's 
>> house in Palo Alto circa 2000, and Tom was the original editor of the 
>> Alternative Natural Philosophy West's journal, in which Lou published some 
>> articles.  Lou told me that he didn't think my "obsession" with "time" was 
>> on target regarding imaginary logic -- which struck me as quite odd given 
>> his obvious affinity for GS Brown. However, this may have been due to the 
>> 

Re: [agi] Experimental Testing of CIC (the Compression Information Criterion)

[Ben Goertzel]

On Mon, Jul 6, 2020 at 11:41 PM Ben Goertzel  wrote:
>
> The COIN Criterion ... sounds like money, it's got to be good...

Maybe we can fund the competition by making a Hollywood-style thriller
about some Bitcoin criminals... lots of potentials here...

>
> On Mon, Jul 6, 2020 at 9:13 PM James Bowery  wrote:
> >
> > On Fri, Jul 3, 2020 at 6:53 PM Ben Goertzel  wrote:
> >>
> >> ...Under what conditions is it the case that, for prediction based on a 
> >> dataset using realistically limited resources, the smallest of the 
> >> available programs that precisely predicts the training data actually 
> >> gives the best predictions on the test data?
> >
> >
> > If I may refine this a bit to head off misunderstanding at the outset of 
> > this project:
> >
> > The CIC* (Compression Information Criterion) hypothesis is that among 
> > existing models of a process producing an executable archive of the same 
> > training data within the same computation constraints, the one that 
> > produces the smallest executable archive will in general be the most 
> > accurate on the test data.
> >
> >
> > Run a number of experiments and for each:
> > 1 Select a nontrivial
> > 1.1 computational resource level as constraint
> > 1.2 real world dataset -- no less than 1GB gzipped.
> > 2 Divide the data into training and testing sets
> > 3 For each competing model:
> > 3.1 Provide the training set
> > 3.2 Record the length of the executable archive the model produces
> > 3.3 Append the test set to the training set
> > 3.4  Record the length of the executable archive the model produces
> > 4 Produce 2 rank orders for the models
> > 4.1 training set executable archive sizes
> > 4.2 training with testing set executable archive sizes
> > 5 Record differences in the training vs test rank orders
> >
> > The lower the average differences the more general the criterion.
> >
> > It should be possible to run similar tests of other model selection 
> > criteria and rank order model selection criteria.
> >
> > *We're going to need a catchy acronym to keep up with:
> >
> > AIC (Akaike Information Criterion)
> > BIC (Bayesian Information Criterion)...
> > ...aka
> > SIC (Schwarz Information Criterion)...
> > ...aka
> > MDL or MDLP (both travestic abuses of "Minimum Description Length 
> > [Principle]" that should be forever cast into the bottomless pit)
> > HQIC (Hannan-Quinn Information Criterion)...
> > KIC (Kullback Information Criterion)
> > etc. etc.
> > Artificial General Intelligence List / AGI / see discussions + participants 
> > + delivery options Permalink
>
>
>
> --
> Ben Goertzel, PhD
> http://goertzel.org
>
> “The only people for me are the mad ones, the ones who are mad to
> live, mad to talk, mad to be saved, desirous of everything at the same
> time, the ones who never yawn or say a commonplace thing, but burn,
> burn, burn like fabulous yellow roman candles exploding like spiders
> across the stars.” -- Jack Kerouac



-- 
Ben Goertzel, PhD
http://goertzel.org

“The only people for me are the mad ones, the ones who are mad to
live, mad to talk, mad to be saved, desirous of everything at the same
time, the ones who never yawn or say a commonplace thing, but burn,
burn, burn like fabulous yellow roman candles exploding like spiders
across the stars.” -- Jack Kerouac

--
Artificial General Intelligence List: AGI
Permalink: 
https://agi.topicbox.com/groups/agi/Ta901988932dbca83-Md530fea79242fc832bfd87f6
Delivery options: https://agi.topicbox.com/groups/agi/subscription