What is NARS's prediction score?

[Immortal Discoveries]

What is NARS's predictive score (either Lossless Compression or its shadowy cousin Perplexity)?

If NARS doesn't have a score, what is your evaluation of it getting closer to AGI? How do you check it I mean?

Also, do you have any completion results like openAI.com's GPT-2 and DALL-E made?

[Pei Wang]

The objective of NARS is different from the projects you mentioned, so cannot use their evaluation method. For discussions on how to evaluate such systems, see The Evaluation of AGI Systems.

Regards,

Pei

--
You received this message because you are subscribed to the Google Groups "open-nars" group.
To unsubscribe from this group and stop receiving emails from it, send an email to open-nars+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/open-nars/115981fd-5457-4eec-9872-11477e52ae58n%40googlegroups.com.

[Immortal Discoveries]

There is only 3 ways to test for AGI:

1) The algorithm/body works like us. This won't tell us so much that it works good though.

2) Scores/compresses good on a large diverse set of tasks. This does tell us.

3) Generates useful outputs on a large diverse set of tasks ex. it generates good looking text/image/or actual real tool/job completions. This too, just as much sometimes tells us.

Wang seems to confuse 2 things as being different, and tries to combine them too:

"Now we see that the empirical approach and theoretical approach of evaluation actually depend on each other for the meta-evaluation."

No, generating solutions, and generating the actual procedure in real life to /do/ it (collect data/ find a cure/ give needle/ etc), are the same thing, we recognize them as a good thing, theories and working cures for humans made by AI are BOTH as useful, a cure for cancer of AI may seem untested crap until proven by a working invention but this isn't the case, ideas can be really accurate if you try. Ideas and inventions are same thing. This evaluation is the #3) I listed above! It's subjective, it sounds great as a completion, but YOU - not Lossless Compression, have to check it, very finely, to make sure it is the best solution, and must check thousands of tasks (solving building higher towers, solving cancer, inventing better storage devices). This is very wobbly, especially if it has no body to actually implement a full working idea to solve cancer and can only make semi-discoveries. That's, why we need to test for SOME accuracy, not that it can solve cancer fully. Way #3) is very wobbly, it's good but not actually telling you it is finding patterns in data.

We recognize/ predict the same way. Humans wrote text, so it therefore is extremely patterny as humans are, because we predict text and our generated text we wrote is what we recognize/ predict. It is the inner patterns of a mind written out, or drawn if use images made by us. The world too does this, physics makes patterns. Our data is actually more advanced in evolution though.

"NARS is based on the belief that “intelligence” is the capability of adaptation with insufficient knowledge and resources. This belief itself is justified empirically — the human mind does have such capability"

This is just pattern finding. you take only some snapshots of the world using only some diverse sensors and so on and can get something much more approx./ like a full atom by atom brute force simulation.

Intelligence is using patterns/ experiences and making your homeworld into a pattern fractal to increase prediction accuracy and becoming a pattern (immortality), things that die die and things that are smart clone and last longer and outlive others. The future world will be a cooled-down metal dark less-dense airless gravityless place that only acts when there is danger from external inputs. Life is whatever acts like a statue/ metal, rocks are Life, not aliens only. Women and donkies and rocks are not sexy, we are born with the reward to predict/ see/ be near them because we need clones to help us with work and survive deaths of "employees" in the homeworld, parent DNAs combine to try likely good traits ex. man+wings, instead of just random mutations.

[robe...@googlemail.com]

1) A intelligence doesn't need a body once it's teached, Hawking is a good example for that. The mind can also and will work differently than the human mind, because AI is not human. One similarity is that the human mind works under AIKR just like NARS.

2) compression is not the core of intelligence, it's just another tool.

3) is domain specific, contemporary ML use a model for task A in domain D, another model for task B in domain D, another model for task C in domain E, etc. A NARS/proto-AGI system will be different, it basically just needs modalities to ingest data and a reasoner core to reason about the data. This is fundamentally different.

Testing for domain and task specific tasks is just that, to narrow and task-specific.

There is always the danger that the system is either designed or optimized for a few tasks, thus leading to non-general "intelligence".

---

NARS is a scientifc theory and there are some implementations which can do some stuff already.

Yes it wasn't shown that a NARS(like) system can "invent" stuff, but this will be eventually solved and shown.

All without requiring outsourced indian workforces to label data, ideally an proto-AGI can discover the regularities itself (we are working on that too).

Hopefully all without human supervision to check if the solutions are good solutions, etc.

Also note that contemporary ML algorithms will never scale to anywhere on super-intelligent level (curing cancer, inventing better storage technology, etc. ). Such a notion is ridiculous and will weed itself out with time in decades in the usual way, overpromise, AI-winter etc. We are currently seeing it with self-driving cars.

It's unclear how far AIKR and NAL based systems can go, probably beyond human level (whatever that is) in a long timeframe >50 maybe >100 years (?)!

Note also that NARS theory and implementation and research is done by a very small team for any notion, expecting us to cure cancer single handed (quite literally) is just that ... ridiculous.

Note also that NARS is being developed since 30 years, so it did exist before companies like DeepMind or OpenAI came into existence. Companies come and go.

To view this discussion on the web visit https://groups.google.com/d/msgid/open-nars/bc28c076-afdf-48a5-9b95-d243cadec336n%40googlegroups.com.

[Immortal Discoveries]

What is an example of you or wang or other doing an evaluation? What does such a moment look like for you/other I mean? For example I'll look at the generated text or compression score and may say "wow, it's text completion it made really completes the rest of the story" or "wow, the compressed file is really small, or Perplexity score is really good". What is your wow moment like? What is it? Which you evaluate/ look for?

[Robert w]

Basically lots of things for which a paper or forum entry is to small and the wrong medium. Books could be literally written about the evaluation of proto-AGI systems. Other authors already have written papers about the evaluation of proto-AGI systems, not just Pei.

I personally don't have a "single" wow moment I am looking for, it's more of a spectrum in tasks/domains I deem interesting and valuable such as NLP/NLU/Vision/Reasoning/Program Understanding/Automated Programming etc. Basically all of my interests of AI/proto-AGI engineering+research.

I guess Pei was very impressed with what his system(s) had done already.

As I said, NARS isn't compressing anything into a model like a ML model is doing. A NARS is also not built on compressionistic ideas, like ML and AIXI approximations. NARS can use compression and ML as mere tools, but that's a different story.

You seem to be to influenced/hyped by GPT-X. GPT-X is to me just a lot of marketing which is misleading and damaging to the field of ML and companies engaging in the bets of OpenAI.

Here is a anti-GPT video where they test GPT for a lot of claims OpenAI made: https://youtu.be/iccd86vOz3w?t=13799
It's really "devastating" because GPT-3 is neither doing NLP or NLU or reasoning, it can't be used for automated programming etc.

[stephen clark]

I always thought the best metric was " I think therefore I am" but writing that into an algorithm has proven to a bit difficult.

To view this discussion on the web visit https://groups.google.com/d/msgid/open-nars/f519a2ff-ecbd-4ad4-86d6-2aba242798f8n%40googlegroups.com.

[Immortal Discoveries]

"I personally don't have a "single" wow moment I am looking for, it's more of a spectrum in tasks/domains I deem interesting and valuable"

I didn't mean a single task of course. I meant what what such a moment looks like. When yous work on NARS, what do yous check for to know it is getting closer to AGI? Do you: check text/ image completions, or simply the fact that it is doing processes like the brain ex. forgetting, sparse memory storage, translating (word2vec) - even if you get NO results on completions/ compression of data? If the latter, then there is no clue those mechanisms are working correctly, nor if its completions/ solutions are on par to ours.

"You seem to be to influenced/hyped by GPT-X"

Jukebox made these completions on my friend's first run, check them all: https://www.youtube.com/watch?v=6Q3V238JmNI

GPT-2 does this too, so does DALL-E for images. Iv'e tried all these but the full DALL-E, they really do generate novel correct completions. And get top benchmarks for compression/ Perplexity evaluations.

All intelligence is is Prediction, all we can do is code in pattern finding mechanisms. AI can answer many problems with outputs /even as soon as you add/ a simple pattern finder like in Prediction by Partial Match (using Backoff mixing). When you add something like word2vec it works together with other pattern systems and can answer even more questions. Our job is to add enough of these until it can use them together and come up with its own by using memories to modulate them.

As far as I'm aware, GPT-3 already has (which may look a bit different due to optimizations) memories, relational embeds, forgetting, Byte Pair Encoding parts learnt, holed matches (Dropout/ Random Forests), Recency energy boosting (say cat more likely if saw cat recently lots and now see ca>? (if predict next letter in this example)), bordism (deciding to predict less 'cat' as the article passage is probably going to soon stop talking about cats, so it too should stop), and Blender/ PPLM/ etc use Persona reward forcing (but no updating (pass reward to related nodes ex. food>AGI)) to make it say some word or phrase or letter domain more likely as it generates. Etc. That's really human brain like. They are also trying sparse memories like the eye retina to capture lnger memories. And DALL-E uses multi-sensory.

[Pei Wang]

On Sun, May 2, 2021 at 11:02 AM Immortal Discoveries <immortal.d...@gmail.com> wrote:

"I personally don't have a "single" wow moment I am looking for, it's more of a spectrum in tasks/domains I deem interesting and valuable"

I didn't mean a single task of course. I meant what what such a moment looks like. When yous work on NARS, what do yous check for to know it is getting closer to AGI? Do you: check text/ image completions, or simply the fact that it is doing processes like the brain ex. forgetting, sparse memory storage, translating (word2vec) - even if you get NO results on completions/ compression of data? If the latter, then there is no clue those mechanisms are working correctly, nor if its completions/ solutions are on par to ours.

Our understanding of "intelligence" is different. I don't see it as the ability to solve a certain set of problems, but a meta-level capability that is independent of domain. See http://sciendo.com/article/10.2478/jagi-2019-0002

Regards,

Pei

 

To view this discussion on the web visit https://groups.google.com/d/msgid/open-nars/48d23e87-dcf0-429c-b080-b7ba36fecb46n%40googlegroups.com.

Virus-free. www.avast.com

[robe...@googlemail.com]

> All intelligence is is Prediction

This is where the NARS theory and philosophy and engineering and involved entities(mostly people) all disagree.

For us intelligence is about doing work under assumption of insufficient knowledge and resources (AIKR).

What does this mean for an implemented system?

* there has to be a way to calculate how much new knowledge (about the environment) is weighted against existing knowledge, none of the work of ML does this correctly as all NARS implementations do

* the system has to be always open to new knowledge, there is no seperation between training(learning) and inference(disabled learning in a deployed system). Even autoregressive models like GPT still have this massive massive flaw

* the system has limited resources - storage and compute. There is never enough time to compute all effects in the environment, there is not enough compute time to retrieve data from a database of all perceptions for most cases, thus the system has to select what to forget and what to keep. There have been attempts to let ML models learn how to access memory for read/write to solve problems such as NTM, but NTM ran into scaling problems which make it impractical for anything useful. NARS is not AIXI and not engineered as an AIXI approximation. My opinion is that a proto-AGI will at some point behave as if it were an AIXI-approximation, but this is a long way off.

Note that it doesn't matter much which techniques were used when designing a AIKR system or if the system was written by humans at all by handcrafting. Most if not all ML models don't implement AIKR systems, thus such models can't be AGI!

Pei wrote a paper about the difference between current contemporary ML models and NARS https://cis.temple.edu/~pwang/Publication/conceptions-of-learning.pdf

GPT is just a better signal processing filter, but intelligence is more than just building a filter.

You can also see GPT as a better pocket calculator, intelligence is using pocket calculators to boost it's performance, we are doing it all the time. A pocket calculator isn't intelligent :)

>As far as I'm aware, GPT-3 already has (which may look a bit different due to optimizations) memories

It has no way to store and retrieve anything beyond it's sequence horizon which are encoded as the input for one forward pass. It doesn't have any long term memory and thus no ability to learn anything at inference time :) . Such a system will never be AGI.

> AI can answer many problems with outputs /even as soon as you add/ a simple pattern finder like in Prediction by Partial Match (using Backoff mixing).

Your talking about ML, not AI!

DL is in ML, ML is in AI.

NARS is a reasoning system which is in AI.

a lot of NARS systems can predict the next symbol too. Of course not as effective as a specialized generic ML model which is trained for it, but it can do it. Main difference is that these NARS systems all have agency (as an agent), most ML systems don't have that.

>That's really human brain like.

No it's not. That's like saying that human brains use electrons for intelligence thus anything that is using electrons is automatically intelligence because the human brain does so too.

No one understands what the human brain is doing, we only understand a tiny fraction of it (hierachical architecture, RL mechanism, motor feedback loops, etc.)

Plus reasoning about anologies is almost completely missing in GPT-2 and GPT-3, it can't do (mathemtical) reasoning at all!

We believe that the brain works under AIKR too, thus we are doing at least something right, even when our implementations can't scale up to some (par human) expectations yet.

The brain is a terrible compressor too, because it's doing way more than just compression :)

Text is already highly compressed from all of human stimulus and the bias which was engrained into the brain by evolution. A lot of the inferences which are done to write and understand text is missing from the communication as (already compressed) text.

GPT-3 is doing something fundamentally different than the brain on the level of reasoning about text. That's why it's not understanding text on par-human level and will never do so.

>They are also trying sparse memories like the eye retina to capture lnger memories.

People tried a lot of AI since the 40s, DL is from the 60s https://builtin.com/artificial-intelligence/deep-learning-history :)

Entities can throw huge amounts of money (or all of their budget) at a problem with the wrong approach without any usable results. We have seen it before. Companies come and go :)

Gwern thinks that GPT-X will scale to par human performance (on text prediction!) with 2.200.000x compute of GPT-3, good luck with that :D

https://www.gwern.net/newsletter/2020/05

>If we irresponsibly extrapolate out the WebText scaling curve further, assume GPT-3 has twice the error of a human at its current WebText perplexity of 1.73 (and so humans are ~0.86), then we need 2.57 · (3.64 · (10³ · x))^-0.048 = 0.86, where x = 2.2e6 or 2,200,000× the compute of GPT-3.

This is all only for a "text modality" without other modalities. OpenAI is basically trying to emulate what evolution did over billions of years on a supercomputer which we are calling earth. Good luck trying to solve these problems with tiny slices of silicon :D

To view this discussion on the web visit https://groups.google.com/d/msgid/open-nars/CAK0i%3DhjJqfX5nHhEBFCnMm34fNS6yzPhCikSU72ZKqDb4XJoew%40mail.gmail.com.

[Immortal Discoveries]

"It has no way to store and retrieve anything beyond it's sequence horizon which are encoded as the input for one forward pass. It doesn't have any long term memory and thus no ability to learn anything at inference time :) . Such a system will never be AGI."

It is not that hard to change it so it can at inference time, there is a few ways to do this. This does disqualify GPT even one bit. It's a design flaw, it isn't a totally different thing because of this.

"NARS systems can predict the next symbol too. Of course not as effective"

"Main difference is that these NARS systems all have agency"

Oh boy, this is philosophical sounding. Agency? And what are these decisions based on? Past experiences? Please explain. All that can matter in the universe is patterns, it has to find patterns, generate patterns, and make its homeworld into a predictable pattern including its life (immortality, i.e. stay the same, be a pattern).

I want to mention 'in' right here an example of how we do a math problem using simple pattern finder rules (I mention them again in the paragraph below). Say we have the prompt "3, 7, 15, 29, 78, ?" If the AI doesn't have a confident match (yes, I can tell if it is confident by looking at if the longest match has many counts, summed over various views of the prompt), and still not enough confident, I can make it match a memory by ex. translation and then ex. link sequentially to a memory (if not trans) that says the rule how the number grows in the prompt, ex. times itself is 3, 9, 27, and so on, so the memory here is #*self(translated/ primed as the same #) and it can try these out on the prompt then until it finds a memory that generates and matches the ones in the prompt, and assumes the predictions is too going to match then. A bit much here but this can be made to work based on just the pattern finder systems we know of.

"Plus reasoning about anologies is almost completely missing in GPT-2 and GPT-3, it can't do (mathemtical) reasoning at all!"

First of all, "too many feet in these shoes" may be overlooked by GPT but this is due to not translating it properly to get the insinuation meant. Secondly, of course it can't do math, there is a billionz of problems it can't solve accurately, but it can most of them using a handful of pattern finding mechanisms (matches, sparse matches, Dropout, Recency+bordism, Seq2Seq, categories, multi-sensory, forgetting, More Data, optimizations ex. backpropagation, etc), it is now our job to make it uses these to allow it to on its own come up with its own pattern rules and from reading text come up with them, ex. one such pattern rule may be to always take the 5th letter of any sentence and predict it 100 letters away, it can do this using the pattern mechanisms we give it, simulate this pattern mechanism, see? Getting to work is the big next problem, but since the pattern mechanisms we know of are IF-THEN rules, it is essentially not far fetched to allow it to run on these...Anyway we let it atain the rest of the pattern finders that would cost us too much time to code (notice the score on Perplexity gets only a tad lower as you add more code). So, to get it lower, you need to let it do the coming up with rules, or better use the ones we gave it. For example the following uses translated matches (recognizes similar tuples), said multiple times as priming, and needs a 3rd item predicted which is a translate too "cat wind dog, moon build, mars, rice book carrots, loop ice hole, dad lint ?"

"A lot of the inferences which are done to write and understand text is missing from the communication"

Go on, explain these AI rules...  though I'm not looking for all that, well, intros and references and rare words which are found in papers at top/ bottom/ and throughout, rather direct simple answers that we both can instantly relate to. Just say their names, and explain them in 30 words. Please? Explain the major NARS systems. I will attach mine below, it isn't the full 17 page AGI Guide but a tidy summary version that you should find easy to read and understand a ton of what I know quickly.

[robe...@googlemail.com]

> Oh boy, this is philosophical sounding. Agency?

Agency as an "agent" in control theory (I belief), it perceives the environment and picks actions on it, RL is an example of a mechanism for an agent. Every intelligent animal is an agent and has agency.

> A bit much here but this can be made to work based on just the pattern finder systems we know of.

Yes Schmidhubers Optimal Ordered Problem Solver can solve simple tasks like that.

Been there, done that. It's not intelligent, just a better pocket calculator.

That a model with 5 million US$ of optimization pressure can learn to induce a sequence is not impressive at all to me. I didn't need 5$ to do my OOPS experiments.

Sure a GPT-X learns millions of these tiny models and it also learns to vote these into the next "prediction", but that's not reasoning, it's just inductive inference. :)

> Go on, explain these AI rules

I wasn't talking about AI rules, humans burn a lot of "thinking cycles" when reasoning about what to say next and how to put it so that the receiver has a chance to unpack it and reconstruct the intended meaning of the sender. Thats way harder than just predicting the next symbol 500 times :)

AI has to reproduce a lot of these processes in the correct way to be useful and make sense of anything. GPT-3 can't do that. OpenAI will be unable to bruteforce this into existence without the right architecture which they wont find in their limited lifetime of the company.

My point there was actually that a lot of background knowledge isn't communicated in written text. When I say "cat" I usually refer to the animal "cat", not the linux program "cat". ML algorithms can learn a algorithm which looks right on the surface to do this disambiguation, but it quickly falls apart when it's closer examinated. Another point is that written text is already compressed because the background knowledge is left out. Else it would read something like this:

A X (where I refer to X here as an animal which we are calling a cat: it's a animal with fur and so on) jumped  (yes I mean the action jumping here) over(relation which indicates a spatial relationship) the brown(the color brown) fox(the animal fox, not the company "fox").

A ML algorithm has no chance to guess the semantic meaning and relationships from already compressed text, that's one reason why GPT-3's responses don't make much sense. The information isn't anywhere or it are just 0.0001 bits or something like that. A compressor can't magically invent the missing information. Compute can't replace this (missing) information, no matter how hard one squeezes the model.

Best may be to just read the paper from Pei about NN's vs NAL which was already linked in this discussion.

---

It's a bit pointless to discuss about software which is not accessible or opensource etc.

All I can tell you is that compression is not intelligence. It's your problem if you can't understand the reasons why this is true.

Pei has actually encoded it in his papers, you just need to find and read the right places which takes time.

---

I didn't read all of your text but just one more thing

>The best evaluation for AGI is Lossless Compression.

This is not true, must come from Hutter. He fell into the same reductionistic trap. See the Hutter price, it won't lead to AGI in my opinion (but it will be useful for AGI as a tool).

I already mentioned the gist of my criticism against this pure compressionistic view.

The hutter prize and GPT-3 are actually closely related, both have the same premise, that intelligence can emerge from just compute+data+some model, this is not the case.

To view this discussion on the web visit https://groups.google.com/d/msgid/open-nars/db3c4c88-a881-40d1-94e7-cacbfd5a02c0n%40googlegroups.com.

[Immortal Discoveries]

"Sure a GPT-X learns millions of these tiny models and it also learns to vote these into the next "prediction", but that's not reasoning, it's just inductive inference. :)"

An AGI, given a Problem/ Question, now has a context. What can it generate? To solve the issue? It checks many memories/ experiences for various matches of the problem. It gives it predictions of what probably comes next. Using reward that favors certain things it could say (fries/ AGI/ lasers), it outputs the Next Word (or letter/ bit). My point is what it can decide has to be based on the context, a known memory, and it combines its experiences to get a novel prediction, it can generate never seen before new discoveries, just like GPT-3 does. I showed you Jukebox, that was never seen before and it successfully completed the techno, like a boss. Any "reasoning" you are posturing about here has-to-be pattern finding, IOW doing memory matches from the prompt and combing the thousands of predictions (or dozens if you intelligently narrow them down). If NARS isn't great at predicting text or image like GPT is, or IOW compression per Hutter Prize and LTCB, it has no actual use other than appearing to function like a brain.

See you say this below yet little predictive ability is the end result:

"unpack it and reconstruct the intended meaning of the sender. Thats way harder than just predicting the next symbol 500 times :)

AI has to reproduce a lot of these processes in the correct way to be useful and make sense of anything. GPT-3 can't do that."

"Agency as an "agent" in control theory (I belief), it perceives the environment and picks actions on it, RL is an example of a mechanism for an agent. Every intelligent animal is an agent and has agency."

I hope you don't mean this is what is AI, this is primitive AI, it learns to walk or whatever but in the end this will not think/ simulate scenarios using many many images and text and sound, it can't think about how to improve hard drives. The brain does specialize where to collect data, though it can do this in the brain alone too, but the data generation is key, the body is only a data collector and an implementer.

"My point there was actually that a lot of background knowledge isn't communicated in written text. When I say "cat" I usually refer to the animal "cat", not the linux program "cat". ML algorithms can learn a algorithm which looks right on the surface to do this disambiguation, but it quickly falls apart when it's closer examinated. Another point is that written text is already compressed because the background knowledge is left out. Else it would read something like this:

A X (where I refer to X here as an animal which we are calling a cat: it's a animal with fur and so on) jumped  (yes I mean the action jumping here) over(relation which indicates a spatial relationship) the brown(the color brown) fox(the animal fox, not the company "fox").

A ML algorithm has no chance to guess the semantic meaning and relationships from already compressed text, that's one reason why GPT-3's responses don't make much sense. The information isn't anywhere or it are just 0.0001 bits or something like that. A compressor can't magically invent the missing information. Compute can't replace this (missing) information, no matter how hard one squeezes the model."

Yes cat is ambiguous until you see context, same for the brown, quick, fox, jumped, over, as you said. And it is somewhat by default disambiguated at start ex. cat usually means a kitty cat animal, the way my/GPT AI solves this is by either context or simply when it sees cat, the next letter or word it predicts is usually (more probably) going to be kitty related, 3% the time i'll say "the cat lifted the rocks over the other crane".

The brain really does invent novel predictions, based on input context and memory contexts. Let me show you. Place 16 windows on the last 16 letters of the prompt: "walking down the stre?" ex. "walki[n[g[ [d[o[w[n[ [t[h[e[ [s[t[r[e]]]]]]]]]]]]]]]]? The short matches to memories are most confident but don't look at more context, the longer memories do but are rarely seen, so we combine all 16 sets of predictions of the Next Letter. So match1 says ex. it saw come next as the next letter a 3 times, b 7 times, z 1 time, etc. It's probability based, see a is saw many times?, the node represents a strength of 3. And we also merge these 16 layer sets as said, too. We also merge translated contexts ex. cat activated dog (because they share contexts), so we get its predictions too. We also look at recently seen letters and words, so if we saw cat lots and now see ca>?, we predict t lots more. I have this working and get approx. 20.3MB on the enwik8 compression test. I don't yet have the translation implemented. Anyway, this is how you get new novel answers to some Problem, you do it by merging patterns and combing the predictions. No, other, way, just more pattern mechanisms.

[robe...@googlemail.com]

GPT never made new "discoveries". I didn't investigate NARS based discovery systems because I had no time to do it.

Sure AGI is supposed to do this and that, the only problem is how one realizes a useful system which is useful to expand to more general systems to expand even further etc. until the system can solve complicated real world tasks as if there is no tomorrow.

All under AIKR because it blows up before its of any use.

Using only compression is way more primitive than using even today's DL.

DL is way more primitive than even a simple NARS.

Yes the inference steps seem to be primitive but the point of NARS is that the system is supposed to learn what knowledge is relevant in which context. It also can add new knowledge as required. GPT can't.

Deployed NARS systems do this all the time already.

Maybe you can solve some simple problems but not any further because the time to compress the task increases with the size of the compressed text to solve the task.

It can't solve tasks in realtime for realistic datasets and useful problems.

Another problem is that you have to do the inferences how your "AGI" solves a problem by just applying compression, this is weak AI.

You will not get to AGI by just doing text compression, it's a dead end.

Same applies to compressing images and video, that takes even more resources.

>If NARS isn't great at predicting text or image like GPT is, or IOW compression per Hutter Prize and LTCB, it has no actual use other than appearing to function like a brain.

There are actually certain deployed NARS in the real world as described in some AGI literature.

NARS doesn't have to be great at a specialized task, it just has to be able to use and create tools to accomplish its goals. Text compression is just such an tool. GPT-3 can't even use tools.

Sure there are a lot of things missing in our implementations, it will improve over time.

This is a NARS forum, not a compressionist or ML forum. Your still trying to cast ML solutions as something else, proto-AGI, which it is not.

You received this message because you are subscribed to a topic in the Google Groups "open-nars" group.
To unsubscribe from this topic, visit https://groups.google.com/d/topic/open-nars/bO4DU78skf4/unsubscribe.
To unsubscribe from this group and all its topics, send an email to open-nars+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/open-nars/27d77088-20c2-4725-b564-f281191b1012n%40googlegroups.com.

[robe...@googlemail.com]

One has to start at sub human capability. 

Starting at superhuman level to narrow it down to human level or sub human would be very strange indeed :D .

There is not enough compute in today's 1000$ computers to drive an human level AGI anyways, just mouse level according to Moravec.

I would be happy with a rat level or crow level proto-AGI. Yes rats can't create technology but it will be useful!

[Immortal Discoveries]

"supposed to learn what knowledge is relevant in which context."

My and GPT AI does this...It's how all AI works.

"Another problem is that you have to do the inferences how your "AGI" solves a problem by just applying compression, this is weak AI."

I already know how my AI works, there are several mechanisms, same for GPT, I can know ahead of time what it can do see? - there is no need to further study how it uses these to arrive as answers after my code is made as I already know how it can use these. Sure that's useful but only for limited tests, and I can do those, though it would be timeful for any discourse of thoughts to be anaylzed.

"Maybe you can solve some simple problems but not any further because the time to compress the task increases with the size of the compressed text to solve the task.

It can't solve tasks in realtime for realistic datasets and useful problems."

The AI I work on/ know of can generate in real time, and the time to read in a dataset is usually slower but still much faster than human. No clue why you say these things. There is no "compress the task", just an AI that stores data and predicts better. There is no barrier, giving it 20GB of text is not the only thing you do to improve AI.

-------------------------------------------------------------------------

I read Pei's paper that compared his and ML's. Pei says his is a semantic grounded AI. ML's AI and my AI both do semantics.... and they are grounded by data by merging patterns, it's that simple, you can see in my Summary.txt above. And mutli-sensory will improve that capsule making too. There is no truth value in a brain, only patterns and reward like in Facebook's Blender, like in robots that learn to walk etc you can find on the internet.

ML and my AI already can answer all sorts of problems not programmed for nor are answered in the data it trained on. And as we add more pattern finding mechanisms like word2vec and beyond (new algorithms), it'll answer even rarer prompts more accurately i.e. using frequency/ recency/ word2vec/ categories/ multi-sensory/ persona reward as in Blender/PPLM/ and more I don't know yet perhaps.

"in NARS every level of generalization is produced by inference rules and verified independently. Even though the system may make mistakes due to insufficient knowledge and resources, it will not make incomprehensible mistakes. "

:) I can test how good my AI is on just 1MB of data, not the full Hutter Prize 100MB. It's score it predictable on larger data. I need not cost money to test my predictor. There's 2 things to ML and my AI as the above quote mentions, more data, and better/ more pattern finders make prediction more accurate. We all know this, it is not specially only in NARS. I can say the same thing quoted above, my AI/ML AI may not be good on small data but it compresses 10KB better than before I added some new Pattern Finder mechanism.

-------------------------------------------------------------------------

Anyway, look, this is going nowhere here, I'll leave it be but I recommend you share how NARS works in just 300 words, common English words, no references, no fancy bullets, just how to find patterns and combine predictions to make accurate discoveries. If you can't, you are not that easy to compress what you know, it is not so impressive to work with you. I can explain AI in a day, as above in Summary.txt.

[robe...@googlemail.com]

> No clue why you say these things.

because I wrote a lot of compressors where I observed these effects.

I tried some folded tree datastructure by datamining from the raw text. This was very slow even for relativly little data like 50MB, because it had to count every occurrence of a pattern with the help of a hashtable for all indices inside the data.

I just don't want to waste valuable compute with this stuff, even if I could compress say 1GB in 10hours. Then I am short of #CPU*10hours CPU time which I can never get back.

I also tried a less efficient incremental compression scheme with a not so good compression ratio for a old AGI-conf paper.

Both algorithms get n times slower or worse with n times as much data to compress, there is no way around it!

One has to compress the data if the data changes. I am not talking about the query/prompt here, because the compression of the data relevant for the task is the bottleneck.

>ML's AI and my AI both do semantics

I am sorry but you are mistaken with the assumption that syntax=semantics and compression=intelligence

>There is no truth value in a brain

It has to be encoded somehow because humans can clearly reason about confidence and uncertainty etc.

> I need not cost money to test my predictor.

There is always a cost, every cycle costs a tiny bit of money and time. Just like this conversation. Everything has a price.

> I recommend you share how NARS works in just 300 words

So you are asking me to explain a actual proto-AGI in 300 words without using scientific language. That's actually impossible because it's based on some complicated concepts which are interwoven with logic and psychology etc.

It's actually not my job nor my task to convince people that NARS is the right approach, it's in fact pointless because people either see it by themself or not.

There is also not any point in working with people who don't even want to invest the time to read the introduction and summary of some relevant NARS papers.

There is also no point in framing NARS as only a sequence predictor, it can predict the next event, but it's also doing way way more than just prediction.

How am I supposed to compress knowledge which is spread over 700 pages from Pei in just 300 words? Thats like asking to "compress" what all instructions of the x86 architecture do in 300 words :)

I don't take this impossible task, maybe someone else wants to do it.

---

To view this discussion on the web visit https://groups.google.com/d/msgid/open-nars/16ce7abb-a0e0-4ff9-a7cd-5e615bb7f76cn%40googlegroups.com.

[Immortal Discoveries]

" One has to compress the data if the data changes. I am not talking about the query/prompt here, because the compression of the data relevant for the task is the bottleneck. "

No. Eventually it will cover all domains. It learns a little bit about cats, factories, Mars, and then it can fill in all the cracks of knowledge and know about farming even though know little on it, without storing much on farming. The G in AGI...

"Truth has to be encoded somehow because humans can clearly reason about confidence and uncertainty etc."

We say what is probable. We don't need a Conscious_Value just because we 'think' we are conscious now do we? No. Probable is truth aka physics. Our brains are great prediction machines despite not simulating all atoms. We model patterns. Semantics is only one such.

[robe...@googlemail.com]

Here is a short description of what NARS is doing:

Take a thought.
Thoughts have to be processed. They have to get completed in a sequence.
A thought has to get processed in limited time.
Thoughts can access memory or get stored into memory, also in limited time.
Memory is of course limited.
Thoughts can be about things to do or known things.
Thoughts can lead to actions done in the world.
Perceptions can invoke new thoughts which need to get processed.

To view this discussion on the web visit https://groups.google.com/d/msgid/open-nars/bbb994b7-bd7b-4446-89ba-4da54ca96e04n%40googlegroups.com.

[Immortal Discoveries]

Everyones AI does all those.