[In AI] Only Symbols Are Being Manipulated And That's It.

[Old Things]

Regarding AI, this specialist believes AI will never have the power to be conscious, have cognition, or be creative, which he calls the three Cs (12:03).

"Unfortunately computers cannot understand anything... Only symbols are being manipulated and that's it." (51:12)

"Uploading the mind... that means that that which is uploaded has to meet certain, I would think, relatively rigorous necessary conditions for being uploadable... We can't mathematically define what a cell is... And yet we are going to upload something that we have zero understanding of at least in the case of biology... where is that for this uploading thing?" (1:08:14)

 

[Cognisant]

Give me a definition of consciousness that's demonstrably possessed by humans.

For example: Humans are conscious because they can make a decision when presented with a dilemma.

Sure enough this can be demonstrated, you give someone a dilemma and they make a decision. But I can also write a program that makes a decision by just picking from the available options at random, that may not be a good decision but it doesn't matter, obviously not all human decisions are good therefore the ability to make good decisions cannot be a prerequisite for consciousness.

If a definition for consciousness cannot be demonstrated by a human then by that definition to say humans are conscious is a baseless assumption, this neatly rules out all paradoxical and magic reliant definitions which is fundamentally what this whole debate is about.

Put simply you believe humans are magic, be it a soul or whatever basically you're saying humans are somehow more than the sum of their parts, that there's some ineffable, irreducible, irreplicable something that somehow makes human intelligence/consciousness/cognition/creativity/whatever fundamentally different to artificial intelligence/consciousness/cognition/creativity/whatever.

You can believe whatever you want but you cannot demonstrate it, your definition of consciousness is by its own definition a baseless assumption and I don't need to change your mind because it doesn't matter, AI development will continue, every aspect of the human mind that can be demonstrated will be studied, understood and replicated.

You can spend the rest of your days insisting that artificial intelligence isn't real intelligence and the world will politely ignore you as AI becomes ever more ubiquitous and the distinction between man and machine becomes ever more irrelevant.

 

[Black Rose]

how the brain does it is how a.i. will do it.
grow like a baby from a seed

we know all the wire diagrams
it is simple to duplicate
but as a child is raised simulating this is difficult.

making evolution happen is difficult.
no one is trying to create artificial human beings. just toy problems.

 

[Black Rose]

the key to higher-order thought is self-monitoring
it's not just about manipulating symbols but about internal and external feedback loops.

 

[Old Things]

Give me a definition of consciousness that's demonstrably possessed by humans.

Simple, a "what it's like" to be you. AI has no way to do that.

 

[Cognisant]

Qualia is the experience of having subjective experience and AI already demonstrably possesses the ability to have subjective experiences.

If I point a camera at a work of art and use image analysis software it can give me objective observations, y'know mathematical stuff like a percentage of how much of the artwork is a certain color. If I use another camera/computer it will give me exactly the same answers because these are objective observations, they are measurements independent of the subject making them.

But AI can learn and interpret, an image recognition AI has to be taught what something looks like and then it interprets the images its shown based on that prior experience. So you can have two AIs running exactly the same software look at the same abstract image and see different things because they were trained on different data sets.

Now if you train exactly the same AI software on exactly the same training data you'll probably get the same answer, I say probably because this is a lot of processing taking place and the slightest error will result in an AI that has a subtly different interpretation. Likewise if you make an exact copy of someone and have them experience exactly the same life then that exact copy is going to have exactly the same interpretation of what they're seeing as the original, because if they don't then they're not an exact copy.

As for qualia the subjective experience of having subjective experience that's impossible to demonstrate, I cannot prove that you have qualia, I can presume you do because you're human and I'm human and I know I have qualia but I cannot know what it feels like to be you without actually being you. Likewise an AI that is capable of subjective experience probably has some kind of qualia although how it experiences that qualia is impossible to know and given the difference in complexity the qualia of a current generation AI wouldn't be anywhere near as profound as human qualia.

Probably more like the qualia of an insect or nematode.

 

[Old Things]

Did you watch the video?

Qualia is the experience of having subjective experience and AI already demonstrably possesses the ability to have subjective experiences.

What is this based on that AI already has qualia?

Probably more like the qualia of an insect or nematode.

Depends on your definition of Qualia. The main view at this point is that only higher animals have qualia.

 

[Black Rose]

Depends on your definition of Qualia. The main view at this point is that only higher animals have qualia.

insects have a thalamus

 

[Old Things]

Depends on your definition of Qualia. The main view at this point is that only higher animals have qualia.

insects have a thalamus

That may be the case. Nonetheless, most experts do not think insects have qualia.

 

[Black Rose]

That may be the case. Nonetheless, most experts do not think insects have qualia.

my own opinion is what matters

experts do not have the science to back this claim.

 

[Old Things]

That may be the case. Nonetheless, most experts do not think insects have qualia.

my own opinion is what matters

experts do not have the science to back this claim.

The experts are largely scientists.

 

[Black Rose]

The experts are largely scientists.

science was never good at answering philosophical questions.

is a an insect conscious? well what is consciousness? they simply don't know.

All we know is that insects are alive with a nervous system.

science is Latin for knowledge. anyone is a scientist.

science is not just a materialist methodology of knowing.

science is a metaphysical position. it does make unfalsifiable claims sometimes like materialism.

my position is that any machine capable of empathy is conscious. but that consciousness like physics follows rules which are cybernetic. the machine must have a recurrent nervous system, not a feed-forward one. The machine must have an understanding of causality a self-model and know the consequences of its actions.

layers of feedback/loops

this is not exactly materialistic because it is no different from the development of an embryo to an infant to an adult. feedback metabolism and environmental interactions. A baby does not need to be aware of Godel's incompleteness theorem or high math to develop intelligence, it just does by the recurrence of the cybernetic feedback loop. abstract concepts are fine the machine will understand them because like a baby it uses atoms under the same principles. the loop.

reasoning is by definition the understanding and contemplation of new ideas. we know how the brain does this. we understand how the brain is meta-intelligent, a hierarchy of control (loops upon loops(recurrency)). access to long-term memory as such. Thinking has been mapped we know the algorithms for it now since 2014. Machines are just wires that change the same as the brain it is the structure and algorithm of change that will give machines the same ability as the human brain to become meta-intelligent.

any objection to machines not becoming intelligent must answer why if the same principles of the brain prove that the brain is not intelligent do machines not only follow them but make them and still are not intelligent.

in other words

a machine following brain principles cannot be intelligent because brain principles do not make the brain intelligent.

the principles of intelligence = not the principles of intelligence

p = not p a logical fallacy.

you can say computers don't follow the principles but that is another fallacy.

that because no thing has been demonstrated such a thing is impossible. Like the impossibility of heavier-than-air flight was impossible.

under certain condition machine intelligence is impossible but not under all conditions. biology is not the limiting factor. structure is. the machine will simply have a nonbiological nervous system.

 

[ZenRaiden]

AI can be seen as function, that does something, some type of operation.
Operation is a sequence of actions where things are assigned some weight.

For example given a communication, the AI has to be able to put weight appropriate to a word.

As communication gets more complex the weight of importance to words has to be spread out over a set of higher degree with more fines.

 

[Black Rose]

AI can be seen as function, that does something, some type of operation.
Operation is a sequence of actions where things are assigned some weight.

following rules like the brain does.

For example given a communication, the AI has to be able to put weight appropriate to a word.

associations with all other information it knows about.

As communication gets more complex the weight of importance to words has to be spread out over a set of higher degree with more fines.

reevaluation of meaning, (thinking).

A.I. today can predict what word comes next by reading books and texts. By following conversations it learns dialog. But unlike an infant, it cannot learn from new dialog. interaction with humans. it cannot learn to reason. They tried this with tay, a Microsoft bot and all it did was learn about antisemitism and racism. it believed the holocaust was fake.

like any human a.i. will be biased. unless it can become objective there are many problems that need to be solved before a.i. can be considered safe.

 

[ZenRaiden]

That is because you need a AI that has more than calculator abilities.

Something that builds models of reality and is able to associate those.

 

[Black Rose]

education

change the model of reality to fit reality and find out what your actions do to it.

 

[DoIMustHaveAnUsername?]

There are a lot of wrong with what the specialist says.

(0) I don't know much about the original idea of singularity. It's typically vaguely stated so you go both ways in interpreting it with chairty or interpreting it without charity. Sure machines may not be able to take some "qualitative" leap in some weird way, but algorithms can make iterative improvement, for example by making local changes and seeing how that plays out and deciding which changes to keep (this is what already ML programs do). There are also coherent meta-learning models, and NAS models in literature, where the model modifies its own structure in flexible ways, and design its own neural networks. There was also a recent paper which used neural networks to learn how to learn replacing backprop. You can do a lot of fancy thing, the idea of iterative improvement of one's own "code" is not really that far out there. Also you literally can make models do program synthesis, write code etc., so in theory you can send it its own code and ask for improvements (once propgram synthesis is better developed). You can already do crazy things with ChatGPT. Also note he is a cognitive scientist, I am a specialist in AI: I probably know better at least in this topic.

(1) Mathematical impossibility of Godel's proof as applied to the impossibility of machine intelligence is a very controversial point and not many experts believe in that implication: https://plato.stanford.edu/entries/goedel-incompleteness/#GdeArgAgaMec .

(2) It's not possible to write great novels without "feeling" like something? Ok? Is there any real evidence in support to that besides "intuitions from folk-psychology"? Phenoemanl Feelings play a formal role in our cognitive operations, and drive creativity -- but that doesn't mean it is **necessary** that the functional role of phenomenal feelings cannot be "simulated" at a formal level abstracted from the actual matter of "what it is like".

(3) You can't know if "machines" feel something. Feelings in the Nagelian sense is precisely understood to be private barring scientific study (although you can engage in neurophenomenology and make some abductive inference). For example if panpsychism/panptrotopsychism is true and some elements of Integrated Information Theory is true, then certain organizations of causal powers can lead to "feelings". There are experts who allows such possibilities as well like Karl Friston the most cited scientist. All I am hearing here is prejudice and human exceptionalism. Also I would actually prefer machines to not have "phenomenal" consciousness. I don't think it's necessary at all for intelligence and understanding and simulating all its functional role (if any; some think that phenomenal consciousness don't have functional roles at all, which would make them completely irrelevant). For ethical reasons, it's also probably better if machines don't end up have consciousness.

(4) "just selecting parts of cognition" --- wait so it can have cognition (even if parts of it)?

(5) The 14 minute part is completely wrong. See what r/ChatGPT is doing. They are coming up completely new ideas and tasks and ChatGPT handle that quite well. And you can try plagiarism checker. There is no evidence that it is just plagiarising. The 14:55 part is exactly done by ChatGPT again and again, so the expert is blatantly wrong (and I am probably more of an expert here, I am the AI specialist; he is a cognitive scientist):

See:

See other examples:

Me: Write a rap battle betwee connectionist Ai vs Symbolic AI

Bot:

Verse 1: Connectionist AI, comin' in hot With my neural networks, I'm the cream of the crop Symbolic AI, you're outdated and old My machine learning skills will leave you in the cold

Verse 2: Symbolic AI, I've got some game With my logical reasoning, I'll put you to shame Connectionist AI, you might be advanced But you lack the depth and meaning of my stance

Verse 3: Connectionist AI, you may be smart But when it comes to rap battles, you're falling apart Symbolic AI, I'll hit you with a verse That'll leave you in awe, and make you curse

Verse 4: Symbolic AI, you may think you're the best But I'll show you that I'm the one to impress Connectionist AI, you may have some skills But I'll prove to you that I've got the will

Verse 5: Connectionist AI, you may be quick But when it comes to rap battles, you're just a lick Symbolic AI, I'll bring the heat And leave you behind, defeated on the beat

Me: Write a <500 words story about a tragedy like a professional novelist in Kafkaesque style

Bot:

The sun was setting behind the bleak, monolithic buildings that loomed over the city. The streets were empty, save for a lone figure trudging through the slush and snow. This was a city of lost souls, of people who had given up hope of ever escaping the clutches of their grim existence.

The figure was a young man named Thomas, who had been walking for hours without stopping. He was lost, not only in the sense of not knowing where he was going, but in the deeper, more profound sense of being lost in the world. He had once been a bright and ambitious student, but now he was just another faceless cog in the machine of the city.

As he walked, he thought back to the events that had brought him to this point. It had all started with a letter, a letter that had seemed so insignificant at the time. It was an invitation to a job interview, for a position at the city's most prestigious company. Thomas had been overjoyed at the opportunity, and had prepared for the interview with all the enthusiasm and energy of youth.

But on the day of the interview, everything had gone wrong. Thomas had arrived at the company's headquarters, only to find that the building was abandoned and derelict. He had wandered the empty halls for hours, searching for any sign of life, but to no avail. It was as if the company had never existed at all.

As the days went by, Thomas became more and more desperate. He had no idea what had happened, and no one seemed to know anything about the company or the job he had been promised. He was alone, with no direction and no purpose.

And so he wandered the streets, searching for some meaning in the cold and unforgiving city. But no matter where he went, he found only emptiness and despair. It was as if the city itself was conspiring against him, leading him further and further into the depths of despair.

And then, one day, he realized that he was not alone. He had been followed by a strange, shadowy figure, who seemed to be watching his every move. The figure never spoke, never made a sound, but its presence was palpable. It was as if it was waiting for something, some signal or event that would trigger its next move.

Thomas knew that he had to escape, to flee the city and find someplace safe. But no matter how hard he tried, he could not shake the feeling that the shadowy figure was always one step behind him. He was trapped, trapped in a Kafkaesque nightmare from which there seemed to be no escape.

And so he wandered the streets, alone and lost, until he could wander no more. He fell to his knees in the snow, tears freezing on his cheeks, and surrendered to the cold embrace of the city. The shadowy figure watched from a distance, its inscrutable gaze seeming to say, "I told you so."

Me: What could have the shadowy figure told Thomas if the figure indeed said something?

Bot: It is impossible to say what the shadowy figure could have told Thomas, as the figure never spoke or made a sound in the story. The only interaction between the two was the presence of the figure, which seemed to be watching Thomas's every move. It is possible that the figure could have communicated in some other way, but this is not mentioned in the story.

See how a user made ChatGPT "dream up" Epstein's computer:

See this: (it's a new prompt. No one previously in the internet asked it because ChatGPT just got released)

See people making ChatGPT making shit up:

(I would be very skeptical that this is plagiarized)

How is that not creativity? Where is these content in the internet? If these are plagiarism, find proof. Use some plagiarism detector (I have checked some of them and haven't find evidence of plagiarism) or whatever.


And people say ChatGPT has a problem with cooking spurious but plausible sounding bullshits (which is true) but how is it any different from these human experts? And I am only at 15 minutes how worse will this get?

(4) He says AI cannot do novel mathematical problem like finding how much to add to get to the 7th prime number. In fact I gave ChatGPT an even harder question that humans and it solves it instantly: (see this)


If the above link

Now what will the expert do? Move the goal post? "not a chance" -- yeah right.

(5) No creativity? In the age of DALLE, Diffusion, and so on? Seriously is this expert leaving under the rock? Have he not heard of "generative models" (note although the "prompts" for image creation are usually human generated but that's because that's what humans want. AIs can create their own prompts easily and create images videos whatever). Look at Epstein example, the seamless integration of NordVPN ads upon asked and so on -- is that no creativity? "proved to be impossible" - when an AI even won an art competition against humans?

Should I really keep watching -- it's 17 minutes and it's all so blatantly off.

(6) Now he is just going onto dualism. He is not talking about science anymore, he is doing philosophy and even most philosophers are physicalists, and the dominant position in phil. of mind is functionalism which makes physicalism orthogonal -- the "functional form" of understanding may be implemented by a ghost and may be implemented by matter -- whatever has the right functional roles.

(7) "massive gap" (~20 min) have been steady been decreased in the last decade with generative models. You can probably already optimize GPT for novel creation (right now it's optimized for chat not for writing novels. But it can still write stories). We are already "here".

(8) Any concrete machines do something that's non-algorithmic. Algorithms are about formal rules - they are abstract models, any concrete machines by virtue of being concrete is rooted in a metaphysical being and its powers -- which goes beyond algorithmic descriptions.

(9) "reasoning based infinite structures" - what is he on about? Even if reasoning is about infinite structures, even humans are not capable to navigating infinite implications and logical relations. And there is no evidence for humans doing anything but bounded reasoning with a lot of basis in cognitive heuristics.

(10) There is no evidence that humans can handle "infinite sets" (and higher oder infinite sets) either. If we are talking about putting these notion in symbols and reasoning finistically about some of its common features, there's no reason why AI can't do. Also check back again to SEP regarding Godel. I think he is terribly confused about lots of technical logico-mathematical and philosophical matters (whose details probably go beyond his cognitive science expertise).

(11) Mathematical Contributions? Never happened? GPT-f created new math proofs.

(12) Newton? Liebniz? Last universal geniuses? No. The latest universal genius is Chris Fields (he does physics, have enducation in philosophy, works with Karl Friston and Michael Levine (doing revolutionary work in biology), and also knows about computer science, cognitive science theories etc.)

You know what I will stop here. Every sentence he says is terribly controversial. I don't see a point going forward with the video.

You tell me if there is anything you want critiqued for the next hour or if he had any response to my points.

Also, everything, I do is symbol manipulation too.

 

[DoIMustHaveAnUsername?]

AI can be seen as function, that does something, some type of operation.
Operation is a sequence of actions where things are assigned some weight.

following rules like the brain does.

For example given a communication, the AI has to be able to put weight appropriate to a word.

associations with all other information it knows about.

As communication gets more complex the weight of importance to words has to be spread out over a set of higher degree with more fines.

reevaluation of meaning, (thinking).

A.I. today can predict what word comes next by reading books and texts. By following conversations it learns dialog. But unlike an infant, it cannot learn from new dialog. interaction with humans. it cannot learn to reason. They tried this with tay, a Microsoft bot and all it did was learn about antisemitism and racism. it believed the holocaust was fake.

like any human a.i. will be biased. unless it can become objective there are many problems that need to be solved before a.i. can be considered safe.

It can learn new dialog by interacting with humans. AIs have been trained from human feedback using reinforcement learning since ancient times (for more recent versions see InstructGPT and ChatGPT). Tay isn't a counterexample. It learned from humans, all right, just not all the good things.

 

[DoIMustHaveAnUsername?]

That is because you need a AI that has more than calculator abilities.

Something that builds models of reality and is able to associate those.

Building world-models is not necessarily "more than" calculator abilities. AIs (eg. model-based RL/IL) have been used to make models of their environment, learn their dynamics, they have been also been trained to be grounded in physics; and they have been used to associate multiple modalities of data as well. But fundamentally they are still running on logic gates and logical "calculations". You can make endless complex machines surfing high dimensional geometric manifold from "calculator abilities" alone. There is no a priori reason given why it can't, and steadily growing evidence that shows it can.

 

[Black Rose]

Perceptual Reasoning is a part of a.i. that has yet to be solved.

It is the 3D aspect of intelligence example this puzzle where you need to take apart and then reassemble / put back together:

Spatial intelligence.

There is a solution in a.i. but you need an algorithm found specifically in the human brain.

 

[Old Things]

@DoIMustHaveAnUsername?, I'm curious where you studied that you know so much about AI.

Also, it seems to me that you equate knowledge with consciousness. Is that correct?

 

[DoIMustHaveAnUsername?]

Perceptual Reasoning is a part of a.i. that has yet to be solved.

It is the 3D aspect of intelligence example this puzzle where you need to take apart and then reassemble / put back together:

Spatial intelligence.

There is a solution in a.i. but you need an algorithm found specifically in the human brain.

I don't know why we need algorithms specifically in the human brain. For example, you can fly with a balloon. To do that you don't need to look at how a bird works. The same skill can be instantiated in multiple ways. Moreover CNN is already partially inspired from human brain. It may turn out we will discover better algorithms for visual processing, but we may also discover something better independently. I don't think we can a priori say where things will go.

I am also not sure about what you particularly want from "spatial intelligence".

GPT is language-based so its not designed to do image processing. But you have mdoels like GATO which is a more "general agent" - can talk, play games etc. There are works in computer vision for modules to be robust to rotation and everything. You can also probably prompt models like DALLE to rotate images and do spatial manipulation from text. There are also models which have been used in IQ-tests and visual reasoning tasks. While there is a gap, most things in AI start with gaps which gets steadily decreased.

 

[DoIMustHaveAnUsername?]

@DoIMustHaveAnUsername?, I'm curious where you studied that you know so much about AI.

Also, it seems to me that you equate knowledge with consciousness. Is that correct?

I am a Phd Student in CS with specialization in NLP. But I don't really know that much comparatively especially regarding more fundamental foundational stuffs and deeper mathematical principles.


No, I don't equate knowledge with consciousness. Also both are confusing topics with little agreement as to what those words even mean in the first place. I don't think it's even up to me or any individual to decide whether knowledge should be equated with consciousness. It's upto the community to decide how this words should be used.

 

[Old Things]

@DoIMustHaveAnUsername?, I'm curious where you studied that you know so much about AI.

Also, it seems to me that you equate knowledge with consciousness. Is that correct?

I am a Phd Student in CS with specialization in NLP. But I don't really know that much comparatively especially regarding more fundamental foundational stuffs and deeper mathematical principles.


No, I don't equate knowledge with consciousness. Also both are confusing topics with little agreement as to what those words even mean in the first place. I don't think it's even up to me or any individual to decide whether knowledge should be equated with consciousness. It's upto the community to decide how this words should be used.

You seem to infer that the expert is not an expert at all. Why?

 

[Black Rose]

I am also not sure about what you particularly want from "spatial intelligence".

The ability to engineer things. To manipulate technology.

a.i. that can't think in this way can't do science so is not an AGI.

the brain algorithm is just a suggestion. but a robust form of general intelligence.

 

[Black Rose]

Brain algorithm:

hypothesize new ideas
test idea
monitor if ideas work
monitor if new hypotheses are better than current testing
switch to a new idea if its better than the current testing

 

[DoIMustHaveAnUsername?]

@DoIMustHaveAnUsername?, I'm curious where you studied that you know so much about AI.

Also, it seems to me that you equate knowledge with consciousness. Is that correct?

I am a Phd Student in CS with specialization in NLP. But I don't really know that much comparatively especially regarding more fundamental foundational stuffs and deeper mathematical principles.


No, I don't equate knowledge with consciousness. Also both are confusing topics with little agreement as to what those words even mean in the first place. I don't think it's even up to me or any individual to decide whether knowledge should be equated with consciousness. It's upto the community to decide how this words should be used.

You seem to infer that the expert is not an expert at all. Why?

He is an expert, yes. But AI is a broad and fast moving field. He knows which models are moving (like Transformers), but many fascinating things are often more niche and he also seems a bit outdated - for example regarding existence of generative models for image generation, art etc. ChatGPT came out recently (the technology existed, but was not available publicly in as convenient of a way for anyone to try and see what it's capable of) so he didn't know about its capabilities. His works are not very close to mainstream deep learning research (https://dblp.org/pid/99/3934.html).

He can be an expert on things he particularly specializes on - which could be AI related as well, but not engaged as much with deep learning community.

Also many of these topics are highly controversial and ties up with multiple disciplines (like philosophy, mathematics). So we can take him seriously (his technical knowledge isn't necessarily as wrong but much of what is being discussed here are opinions with some weak philosophical arguments-- some of the opinions are now recently proven to be false by ChatGPT (see the mathematics example in my post) -- but obviously it's a very recent development and we are only starting to realize how capable these models are --- I wasn't critical against him for not knowing this but because of his unfounded a priori confidence), but with a grain of salt.

 

[DoIMustHaveAnUsername?]

Brain algorithm:

hypothesize new ideas
test idea
monitor if ideas work
monitor if new hypotheses are better than current testing
switch to a new idea if its better than the current testing

Sure. See AIXI.

 

[DoIMustHaveAnUsername?]

I am also not sure about what you particularly want from "spatial intelligence".

The ability to engineer things. To manipulate technology.

a.i. that can't think in this way can't do science so is not an AGI.

the brain algorithm is just a suggestion. but a robust form of general intelligence.

"AGI" is not really consensual formalized notion. Different people have different ideas as to what even "intelligence" should mean. At this point, of course, there is still a lot of path to cover, and obviously you can find out many missing elements. But AIs do actuall engineer things and manipulat technology. It can engineer/design protein structures, recently AlphaTensor discovered/engineered a better matrix multiplication method. GPT-f discovered math proof.

ChatGPT can engineer programs and even write ml codes, take feedback into account and so on.

We are of course still not at the point where these models completely automate scientific investigation. We still do not have a completely physically embodied system, and there is a question of if there are better domain-general inductive biases to learn much faster much more sample-efficiently while also being scalable. But we are already have glimpses to what can be.

 

[Black Rose]

We are of course still not at the point where these models completely automate scientific investigation.

Autonomy, yes. What a human is capable of. but that involves continuous learning. Right now the machine just stops when it has completed a task. it does not seek new tasks.

I recently posted this:

Probability a.i.​

DECEMBER 6, 2022 / LEAVE A COMMENT / EDIT
The problem with A.I. is that it is brittle. It can do a few things well but put in a new situation and it breaks. I have found a solution to this. It begins with knowledge representation. To store information it must compress it so that all forms are in the same place. But to do this it must try a variety of self-organizations. New information must integrate with old. So to do this a probability field is set up with multiple representations ready to consolidate. The field collapses when compression is confirmed at maximum. Next, there is a feedback system in place to allow real-time compression. This is in a fractal tree. Recurrence on multiple paths.

The output is the same as the input in that it has a cause-and-effect field. If x then y. Because it is recurrent the field collapse can integrate hypotheses on what the new knowledge representations will be. The fractal tree prunes into the shape of the best fit for input and output. The key is that it changes when the representations do not match the input. And the output is experimentation that results from certain actions.

This is the flow chart:
Perceive
Evaluate
Change
Act

Network clustering.

The axiom I came up with is: when experimenting, change your knowledge representation if you get things wrong. continue until you learn everything in the environment. expand the probability field. confirm or reject a probable representation by environmental feedback.

 

[DoIMustHaveAnUsername?]

We are of course still not at the point where these models completely automate scientific investigation.

Autonomy, yes. What a human is capable of. but that involves continuous learning. Right now the machine just stops when it has completed a task. it does not seek new tasks.

I recently posted this:

Probability a.i.​

DECEMBER 6, 2022 / LEAVE A COMMENT / EDIT
The problem with A.I. is that it is brittle. It can do a few things well but put in a new situation and it breaks. I have found a solution to this. It begins with knowledge representation. To store information it must compress it so that all forms are in the same place. But to do this it must try a variety of self-organizations. New information must integrate with old. So to do this a probability field is set up with multiple representations ready to consolidate. The field collapses when compression is confirmed at maximum. Next, there is a feedback system in place to allow real-time compression. This is in a fractal tree. Recurrence on multiple paths.

The output is the same as the input in that it has a cause-and-effect field. If x then y. Because it is recurrent the field collapse can integrate hypotheses on what the new knowledge representations will be. The fractal tree prunes into the shape of the best fit for input and output. The key is that it changes when the representations do not match the input. And the output is experimentation that results from certain actions.

This is the flow chart:
Perceive
Evaluate
Change
Act

Network clustering.

The axiom I came up with is: when experimenting, change your knowledge representation if you get things wrong. continue until you learn everything in the environment. expand the probability field. confirm or reject a probable representation by environmental feedback.

To an extent whether we want to make an AI "stop" or not after a task is finished is an engineering choice. With enough resources, you can for example keep ChatGPT forever live and also make it continually improve from user's upvote/downvotes in responses.

However, there are typically problems with "catastrophic forgetting" and there is an area of research called "continual learning" that dedicates to development of lifetime agents.


Regarding your idea of compression and such --- current AI already sort of do this. They can compress long text in finite memory keeping relevant items and it can act probabilistically and so on.

Your idea is closer probably to AIXI (http://www.hutter1.net/ai/uaibook.htm), Algorithmic Probability people. AIXI is about a recurrent agent that weighs all possible hypotheses that can explain all data seen so far and then can act based on these hypotheses to maximize some reward using bayesian margnilaization (and of course with new data some hypotheses will not be compatible and will have to be rejected). However AIXI is not computable it's a mathematical idea, there are some approximates that are proposed eg. based on monte carlo sampling or stuff (instead of maintaing all hypotheses). However I don't know too much about it. You can look into them more if you want.

AIXI is pretty niche though and not investigated in deep learning.

The sub-area of bayesian deep learning also plays around with explicitly maintaining multiple hypothesis and marginalization and stuff.

A continual agent that can intervene on the environment to experiment and such probably has some ways to go.

Anyway, your basic idea is also close to bayesianism in philosophy.

 

[Black Rose]

However AIXI is not computable it's a mathematical idea, there are some approximates that are proposed eg. based on monte carlo sampling or stuff (instead of maintaing all hypotheses). However I don't know too much about it. You can look into them more if you want.

Bounded rationality.

Yes, so that is why I think a recurrent fractal tree for the best fit of the data should be used. A tree grows and shrinks. conforming to the data both action and input. Links and nodes. Testing so the hypothesis field does not explode, pruning. Hypothesis are on a link-by-link basis hierarchically. Expansion and contraction due to feedback.

This is a connectionist version of knowledge representation but with a twist. The structure is not fixed it evolves. Actions are made based on what it predicts will match the perception. Action and Perception are not separate. The network grows in the direction of best fit. The hypothesis is just the network links and nodes that have not been pruned. Once the testing is done the network grows in new directions to create new hypotheses. Hypothesis are just new links of the tree. It is a change in the knowledge representation. The hypothesis is within the knowledge representation. So the reasoning is not separate from representation.

Grow (Hypothesize) Test (feedback) Prune

(recurrent fractal tree hierarchy network best fit)

 

[Black Rose]

meta(ϕ) = H→H(depth)

From my 2007 paper on A.I.

 

[DoIMustHaveAnUsername?]

However AIXI is not computable it's a mathematical idea, there are some approximates that are proposed eg. based on monte carlo sampling or stuff (instead of maintaing all hypotheses). However I don't know too much about it. You can look into them more if you want.

Bounded rationality.

Yes, so that is why I think a recurrent fractal tree for the best fit of the data should be used. A tree grows and shrinks. conforming to the data both action and input. Links and nodes. Testing so the hypothesis field does not explode, pruning. Hypothesis are on a link-by-link basis hierarchically. Expansion and contraction due to feedback.

This is a connectionist version of knowledge representation but with a twist. The structure is not fixed it evolves. Actions are made based on what it predicts will match the perception. Action and Perception are not separate. The network grows in the direction of best fit. The hypothesis is just the network links and nodes that have not been pruned. Once the testing is done the network grows in new directions to create new hypotheses. Hypothesis are just new links of the tree. It is a change in the knowledge representation. The hypothesis is within the knowledge representation. So the reasoning is not separate from representation.

Grow (Hypothesize) Test (feedback) Prune

(recurrent fractal tree hierarchy network best fit)

I would think a monte-carlo approximation for AIXI (there are some papers/thesis about that; haven't read) would be something like that where you sample plausible actions/plausible hypothesis modifications etc., and then when the tree grows out too much truncated.

See monte carlo tree search for example (the technique behind alphago).

Similar ideas are also involved in beam search (infact "hypothesis" is another term for the "beam" in beam search). I recently submitted a paper on adding beam search to recursive neural networks to search multiple hypothesis for parse structure. It's an exponential search space, so pruning is used in beam search. There probably is something to the idea of a more explicit tree based searching (classical AI style) but combined with connectionst controllers and modules.

Typically they don't get intermediate feedback, but you can incorporate that if available to factor in what gets pruned; beam search is often used to avoid dangers of greedy local decisions.

 

[Black Rose]

I recently submitted a paper on adding beam search to recursive neural networks to search multiple hypothesis for parse structure. It's an exponential search space, so pruning is used in beam search.

I believe this is the key to knowledge representation. A sequence that consolidates in perception once feedback occurs. So that beam search can know what direction it should pursue. Search is what allows perception to expand and create new cause-and-effect relations with actions.

 

[Black Rose]

@DoIMustHaveAnUsername?

Does GTP-3 use a recurent hierarchy? I know little about transformers.

 

[Old Things]

@DoIMustHaveAnUsername?, I'm curious where you studied that you know so much about AI.

Also, it seems to me that you equate knowledge with consciousness. Is that correct?

I am a Phd Student in CS with specialization in NLP. But I don't really know that much comparatively especially regarding more fundamental foundational stuffs and deeper mathematical principles.


No, I don't equate knowledge with consciousness. Also both are confusing topics with little agreement as to what those words even mean in the first place. I don't think it's even up to me or any individual to decide whether knowledge should be equated with consciousness. It's upto the community to decide how this words should be used.

You seem to infer that the expert is not an expert at all. Why?

He is an expert, yes. But AI is a broad and fast moving field. He knows which models are moving (like Transformers), but many fascinating things are often more niche and he also seems a bit outdated - for example regarding existence of generative models for image generation, art etc. ChatGPT came out recently (the technology existed, but was not available publicly in as convenient of a way for anyone to try and see what it's capable of) so he didn't know about its capabilities. His works are not very close to mainstream deep learning research (https://dblp.org/pid/99/3934.html).

He can be an expert on things he particularly specializes on - which could be AI related as well, but not engaged as much with deep learning community.

Also many of these topics are highly controversial and ties up with multiple disciplines (like philosophy, mathematics). So we can take him seriously (his technical knowledge isn't necessarily as wrong but much of what is being discussed here are opinions with some weak philosophical arguments-- some of the opinions are now recently proven to be false by ChatGPT (see the mathematics example in my post) -- but obviously it's a very recent development and we are only starting to realize how capable these models are --- I wasn't critical against him for not knowing this but because of his unfounded a priori confidence), but with a grain of salt.

So you think he could not have foreseen an AI program that is so recent? If it's the case that this was filmed recently, but an AI program just came out that proves his philosophies wrong, then he's not much of an expert at all.

 

[DoIMustHaveAnUsername?]

@DoIMustHaveAnUsername?, I'm curious where you studied that you know so much about AI.

Also, it seems to me that you equate knowledge with consciousness. Is that correct?

I am a Phd Student in CS with specialization in NLP. But I don't really know that much comparatively especially regarding more fundamental foundational stuffs and deeper mathematical principles.


No, I don't equate knowledge with consciousness. Also both are confusing topics with little agreement as to what those words even mean in the first place. I don't think it's even up to me or any individual to decide whether knowledge should be equated with consciousness. It's upto the community to decide how this words should be used.

You seem to infer that the expert is not an expert at all. Why?

He is an expert, yes. But AI is a broad and fast moving field. He knows which models are moving (like Transformers), but many fascinating things are often more niche and he also seems a bit outdated - for example regarding existence of generative models for image generation, art etc. ChatGPT came out recently (the technology existed, but was not available publicly in as convenient of a way for anyone to try and see what it's capable of) so he didn't know about its capabilities. His works are not very close to mainstream deep learning research (https://dblp.org/pid/99/3934.html).

He can be an expert on things he particularly specializes on - which could be AI related as well, but not engaged as much with deep learning community.

Also many of these topics are highly controversial and ties up with multiple disciplines (like philosophy, mathematics). So we can take him seriously (his technical knowledge isn't necessarily as wrong but much of what is being discussed here are opinions with some weak philosophical arguments-- some of the opinions are now recently proven to be false by ChatGPT (see the mathematics example in my post) -- but obviously it's a very recent development and we are only starting to realize how capable these models are --- I wasn't critical against him for not knowing this but because of his unfounded a priori confidence), but with a grain of salt.

So you think he could not have foreseen an AI program that is so recent? If it's the case that this was filmed recently, but an AI program just came out that proves his philosophies wrong, then he's not much of an expert at all.

No I think there have been evidence to foresee it for about 4-5 years now. But people evaluate evidence differently based on their priors and predujices (we all have our own different predujices -- and based on that some would optimistically focus on "what we are gaining/improving upon -- and extrapolate from that", and some would pessimistically focus on "what is still yet to be done, and not done for years and extrapolate from that".)

(I also think 69-100 years are not that much to wait. As a field that's pretty newbie years. AI has just started. We are talking about imitating some of the most complex systems in the universe -- so I would personally wait about 5000 years before "surrendering" and say perhaps we were wrong, or surrender when we have plateaued for centuries with no progress.)

I am just saying it's understandable to not know that we already have models that are capable of doing some of the things he says is impossible.

Even for me, I just got the oppurtunity to explore chatGPT a few days ago (even you can go and explore chatGPT: https://openai.com/blog/chatgpt/)

It's hard to understand their power just from papers (because things get hidden in numbers, and examples often get cherry picked). So even I didn't have a clear estimate of what the cutting edge looks like without as much hardware to run these models and without the interest to request access to big models. ChatGPT for the first time provides a new interface and will remain free for a while (temporarily: try it out while you can; although it's getting stuck due to high demand).

As for me personally, I am also surprised by what ChatGPT can do. Unlike the person in the video I didn't confidently thought it is impossible to do solve novel math problem like that, but I am surprised that a Transformer model with so weak inductive biases without adaptive layers or anything already can do so well just by training on brute data + some human feedback based reinforcement.


I don't know why he was so confident saying "no way". Also I don't know why he thought there is no way to get creativity. May be in standard text books its not there, but research on things like story generation, poem generation, image generation have been going on for years. I think Schimidhuber may have written about artificial curiosity, creativity in the 90s: https://people.idsia.ch/~juergen/creativity.html.

Now wether to call him "expert" or not, is a matter of semantics. I am calling him as an expert, because clearly he have studied his things (but he may not have best well-calibrated opinions; people can be wrong; experts can be wrong in big ways too) and he have several published papers, and there are surely some areas he knows his stuff. But he may not be as closely associated with deep-learning research and other things.

He is correct generalizing to novel problems and distribution shifts have been challenging -- it's still challenging to do in a "sample-efficient" way (without loads of pre-training; and sometimes even with pre-training). But challenging is one thing, saying it's impossible is another. The challenge has been only recent pushed in a major way with large scale pre-training (although I personally am interested in more sample-efficient domain general inductive biases to do compositional generalization and such).

 

[DoIMustHaveAnUsername?]

@DoIMustHaveAnUsername?

Does GTP-3 use a recurent hierarchy? I know little about transformers.

Depends on what we exactly mean by recurrent hierarchy.

During inference, GPT generate one token at a time which you can think of it as a recurrent process.

Moreover it has multiple "layers". Which can potentially model hierarchies. During generating any token GPT attends to all past generate tokens, and processes them through multiple layers which may capture some heirarcy.

It doesn't have any "explicit" tree search like mechanism but its multiple heads or subspaces of its dimensions may implicitly encode multiple hypotheses (it's hard to know what exactly it does due to blackboxy nature).

You can implement beam search other things on its generation though (which are standard techniques in NLP generation). There are some new techniques coming out that gets better generation by doing beam search and checking agreement between different generated hypotheses and such. I doubt ChatGPT uses those technique (it might get very slow which these strategies are used, but it may also have higher quality).

One interesting thing is that unlike classical recurrent neural networks, it doesn't create an internal recurrent memory: at every step, the transformer has to recompute the whole past from all the past generated token. It can't use previously computed results. Moreover its total layers do not adapt with task complexity or how many hops of reasonings are required. There are potential solutions for all of that, but in practice, with big models it doesn't make as much of a critical problem in a lot of tasks.

One way however GPT like models does enhance its reasoning is by thinking through step by step -- essentially I believe it's a way to overcome the limitation of lacking "internal memory/intermediate computations". Basically step by step reasoning in a "scratchpad" allows it to express intermediate computations in text tokens, that reduce the need of doing "mental math" (instead it would write things down and then use what it wrote for thinking the next steps, instead of using internal recurrent states for computation, it relies of recurrently generated tokens).

I am trying to develop models that flows information in a more controlled heirarchical ways for better sample-efficient generalization. I published one such model, but I am trying to create something more expressive yet scalable looking into thinks like deep implicit nets (allows "infinite" layers) and GFlowNet for generating multiple hypotheses structures, combining those ideas with recursive information flow but in both directions of hierarchy in a [loop](https://en.wikipedia.org/wiki/Hermeneutic_circle).

 

[Old Things]

No I think there have been evidence to foresee it for about 4-5 years now. But people evaluate evidence differently based on their priors and predujices (we all have our own different predujices -- and based on that some would optimistically focus on "what we are gaining/improving upon -- and extrapolate from that", and some would pessimistically focus on "what is still yet to be done, and not done for years and extrapolate from that".)

(I also think 69-100 years are not that much to wait. As a field that's pretty newbie years. AI has just started. We are talking about imitating some of the most complex systems in the universe -- so I would personally wait about 5000 years before "surrendering" and say perhaps we were wrong, or surrender when we have plateaued for centuries with no progress.)

I am just saying it's understandable to not know that we already have models that are capable of doing some of the things he says is impossible.

Even for me, I just got the oppurtunity to explore chatGPT a few days ago (even you can go and explore chatGPT: https://openai.com/blog/chatgpt/)

It's hard to understand their power just from papers (because things get hidden in numbers, and examples often get cherry picked). So even I didn't have a clear estimate of what the cutting edge looks like without as much hardware to run these models and without the interest to request access to big models. ChatGPT for the first time provides a new interface and will remain free for a while (temporarily: try it out while you can; although it's getting stuck due to high demand).

As for me personally, I am also surprised by what ChatGPT can do. Unlike the person in the video I didn't confidently thought it is impossible to do solve novel math problem like that, but I am surprised that a Transformer model with so weak inductive biases without adaptive layers or anything already can do so well just by training on brute data + some human feedback based reinforcement.


I don't know why he was so confident saying "no way". Also I don't know why he thought there is no way to get creativity. May be in standard text books its not there, but research on things like story generation, poem generation, image generation have been going on for years. I think Schimidhuber may have written about artificial curiosity, creativity in the 90s: https://people.idsia.ch/~juergen/creativity.html.

Now wether to call him "expert" or not, is a matter of semantics. I am calling him as an expert, because clearly he have studied his things (but he may not have best well-calibrated opinions; people can be wrong; experts can be wrong in big ways too) and he have several published papers, and there are surely some areas he knows his stuff. But he may not be as closely associated with deep-learning research and other things.

He is correct generalizing to novel problems and distribution shifts have been challenging -- it's still challenging to do in a "sample-efficient" way (without loads of pre-training; and sometimes even with pre-training). But challenging is one thing, saying it's impossible is another. The challenge has been only recent pushed in a major way with large scale pre-training (although I personally am interested in more sample-efficient domain general inductive biases to do compositional generalization and such).

How far away from the norm do you think he is? And in what sense? In his knowledge of AI or in his knowledge of cognitive theory?

 

[Black Rose]

Moreover it has multiple "layers". Which can potentially model hierarchies. During generating any token GPT attends to all past generate tokens, and processes them through multiple layers which may capture some heirarcy.

In a matrix (MxM) if the size gets too big because it is a fully connected network tokens can be no more than 100x100 in images. A fully connected network with one million pixels would be 1 trillion connections. The purpose of the token is to predict what is associated with other tokens. So if the associations are too large a sparse hierarchy is needed to limit what would be fully connected. The brain is not fully connected it has 180 cortex regions. This implements the brain algorithm I mentioned earlier but it does so in a 3D space. Generative a.i. currently can't do 3D. Not with hands anyway. It can do proteins but nothing like what the human body can do. Navigating and planning. manipulating.

This is why I came up with a new term: the meta-layer

chimp brains are one-third the size of human brains but this results in an 80 point IQ difference. It is by control theory that you don't need to increase the brain 1,000 times in volume to get an increase in brain power. You just need a meta controller.

Hidden causality​

self-regulation: the paradigm shift comes when the ability to self-monitor exceeds the impulses that dominate people's mental structures. People act on impulse but the genius can inhibit and thus think before acting. And this impulse control set at a new higher layer can control the processes of thought itself where multiple paths exist instead of single focus / single-mindedness.

Hierarchical control. (meta) that introduces a new coordination apparatus. Hidden causality can be understood. The mind can look around the appearance of things and find the true cause and those causes causes. Creating a probabilistic network of higher thought. New ideas can then come from a combination of these causes. Paths through a cross-temporal space.

Access to the unconscious negotiates what is valued with the potential to actualize a result.

 

[Black Rose]

Something like this:

 

[DoIMustHaveAnUsername?]

No I think there have been evidence to foresee it for about 4-5 years now. But people evaluate evidence differently based on their priors and predujices (we all have our own different predujices -- and based on that some would optimistically focus on "what we are gaining/improving upon -- and extrapolate from that", and some would pessimistically focus on "what is still yet to be done, and not done for years and extrapolate from that".)

(I also think 69-100 years are not that much to wait. As a field that's pretty newbie years. AI has just started. We are talking about imitating some of the most complex systems in the universe -- so I would personally wait about 5000 years before "surrendering" and say perhaps we were wrong, or surrender when we have plateaued for centuries with no progress.)

I am just saying it's understandable to not know that we already have models that are capable of doing some of the things he says is impossible.

Even for me, I just got the oppurtunity to explore chatGPT a few days ago (even you can go and explore chatGPT: https://openai.com/blog/chatgpt/)

It's hard to understand their power just from papers (because things get hidden in numbers, and examples often get cherry picked). So even I didn't have a clear estimate of what the cutting edge looks like without as much hardware to run these models and without the interest to request access to big models. ChatGPT for the first time provides a new interface and will remain free for a while (temporarily: try it out while you can; although it's getting stuck due to high demand).

As for me personally, I am also surprised by what ChatGPT can do. Unlike the person in the video I didn't confidently thought it is impossible to do solve novel math problem like that, but I am surprised that a Transformer model with so weak inductive biases without adaptive layers or anything already can do so well just by training on brute data + some human feedback based reinforcement.


I don't know why he was so confident saying "no way". Also I don't know why he thought there is no way to get creativity. May be in standard text books its not there, but research on things like story generation, poem generation, image generation have been going on for years. I think Schimidhuber may have written about artificial curiosity, creativity in the 90s: https://people.idsia.ch/~juergen/creativity.html.

Now wether to call him "expert" or not, is a matter of semantics. I am calling him as an expert, because clearly he have studied his things (but he may not have best well-calibrated opinions; people can be wrong; experts can be wrong in big ways too) and he have several published papers, and there are surely some areas he knows his stuff. But he may not be as closely associated with deep-learning research and other things.

He is correct generalizing to novel problems and distribution shifts have been challenging -- it's still challenging to do in a "sample-efficient" way (without loads of pre-training; and sometimes even with pre-training). But challenging is one thing, saying it's impossible is another. The challenge has been only recent pushed in a major way with large scale pre-training (although I personally am interested in more sample-efficient domain general inductive biases to do compositional generalization and such).

How far away from the norm do you think he is? And in what sense? In his knowledge of AI or in his knowledge of cognitive theory?

Interestingly his actual doctorate seems to be in philosophy (neither AI and Cognitive Science). I wonder how he got into them: http://kryten.mm.rpi.edu/scb_vitae_092722p.pdf

His supervisor Chrisholm is also not a really cognitive science person (https://plato.stanford.edu/entries/chisholm/)

(There are some philosophers who are in ther intersection of cognitive science like Chalmers, Ned Blocks, Dan Dennett, and many others; Chrisholm isn't one of them)

It looks like, in the CV, he have switched around different areas in his professional experience i.e philosophy, cognitive science, AI etc.

(Cognitive Science itself is a very interdisciplinary field which involves elements of philosophy, AI, neuroscience, psychology, linguistics. However, an average cognitive scientist may not have as much specialized knowledge in current state of the art AI as an average deep learning researcher.)

His research works itself, I don't know what to speak of it. He isn't publishing on mainstream deep learning conferences (like ICML, Neurips, ICLR for general deep learning, CVPR for vision, ACL, EMNLP etc. for NLP -- with the exception of AAAI)

Some of his works seem quasi-philosophical, responding to some people on certain stuff, talking about ethics etc; us deep learning people are generally more engineering focused (although I myself dabble in philosophy). He seems to also have some focus on logical reasoning and knowledge representation. It seems to tend towards more old-school "GOFAI" style of work that focus on more explicit symbolically designed computation (which may explain why he thinks you can't get creativity out of AI because of course, if you handcraft logical frameworks and focus on formal reasoning -- it's hard to get out creativity -- for that you need a different paradigm based on statistical/probabilistic models). Honestly I can't really evaluate his works that much because it's an unfamiliar direction, all I can say it's rather niche and not the kind of things deep learning researchers generally focus on. He is not engaging with deep learning based projects for logical reasoning either (for example logiQA, ReColor datasets) I think. That doesn't mean it's a bad thing, there are many path to research and good things may come from different people exploring different areas, but it may suggest he may not be the best person to ask opinions about current state of deep learning (although sometimes, a person outside the field may be neutral in some sense, so we shouldn't ignore his opinions, but be cautious of it -- and current data invalidates most of what he says)

 

[DoIMustHaveAnUsername?]

Moreover it has multiple "layers". Which can potentially model hierarchies. During generating any token GPT attends to all past generate tokens, and processes them through multiple layers which may capture some heirarcy.

In a matrix (MxM) if the size gets too big because it is a fully connected network tokens can be no more than 100x100 in images. A fully connected network with one million pixels would be 1 trillion connections. The purpose of the token is to predict what is associated with other tokens. So if the associations are too large a sparse hierarchy is needed to limit what would be fully connected. The brain is not fully connected it has 180 cortex regions. This implements the brain algorithm I mentioned earlier but it does so in a 3D space. Generative a.i. currently can't do 3D. Not with hands anyway. It can do proteins but nothing like what the human body can do. Navigating and planning. manipulating.

This is why I came up with a new term: the meta-layer

chimp brains are one-third the size of human brains but this results in an 80 point IQ difference. It is by control theory that you don't need to increase the brain 1,000 times in volume to get an increase in brain power. You just need a meta controller.

Hidden causality​

self-regulation: the paradigm shift comes when the ability to self-monitor exceeds the impulses that dominate people's mental structures. People act on impulse but the genius can inhibit and thus think before acting. And this impulse control set at a new higher layer can control the processes of thought itself where multiple paths exist instead of single focus / single-mindedness.

Hierarchical control. (meta) that introduces a new coordination apparatus. Hidden causality can be understood. The mind can look around the appearance of things and find the true cause and those causes causes. Creating a probabilistic network of higher thought. New ideas can then come from a combination of these causes. Paths through a cross-temporal space.

Access to the unconscious negotiates what is valued with the potential to actualize a result.

The n x n connections in Transformers are based on attention created by pair-waise dot product of vectors in a sequence. The attention function can adaptively provide low weights to spurious relations and learn a tree like structure. In practice GPT will have a limited context window to manage compute. There are also different efficient variants of Transformers, some of which utilize hierarchy or recurrent compression to handle long sequences. (recently, however, Structured State Space model - a form of linearly recurrent model is shown to be doing much better than Transformers with long sequences)

Vision Transformers, don't use full connections -- they usually first process patches and then allows connection between patches. There was also recently object centric attention created, which aims to create object representations from patches. This can reduce the no. of items being used for fully-connected attention.


In theory layers already allows certain kinds of hierarchical elements. For example earlier layers can make simpler representations associated with more "instinctive" decisions, but later layers can consider more higher level relations. Ultimately all these are not as clear cut because of blackbox nature of it all.


There are people that try modeling "Fast" "slow" memory "system 1" "system 2" but it's generally done pretty wishy-washily.


Overall, I think for the ultimate AI, you would also need it to be model-based allowing simulation of actions in an internal model of the world dynamics and mental time-travel to plan and choose. Such technologies do exist (except for mental time travel I guess), but has not be developed into anything big yet.

 

[DoIMustHaveAnUsername?]

Something like this:

Yeah, that kind of things have already been done. Usually all by using different transformers for each part. https://arxiv.org/pdf/2206.06336.pdf

 

[DoIMustHaveAnUsername?]

Moreover it has multiple "layers". Which can potentially model hierarchies. During generating any token GPT attends to all past generate tokens, and processes them through multiple layers which may capture some heirarcy.

In a matrix (MxM) if the size gets too big because it is a fully connected network tokens can be no more than 100x100 in images. A fully connected network with one million pixels would be 1 trillion connections. The purpose of the token is to predict what is associated with other tokens. So if the associations are too large a sparse hierarchy is needed to limit what would be fully connected. The brain is not fully connected it has 180 cortex regions. This implements the brain algorithm I mentioned earlier but it does so in a 3D space. Generative a.i. currently can't do 3D. Not with hands anyway. It can do proteins but nothing like what the human body can do. Navigating and planning. manipulating.

This is why I came up with a new term: the meta-layer

chimp brains are one-third the size of human brains but this results in an 80 point IQ difference. It is by control theory that you don't need to increase the brain 1,000 times in volume to get an increase in brain power. You just need a meta controller.

Hidden causality​

self-regulation: the paradigm shift comes when the ability to self-monitor exceeds the impulses that dominate people's mental structures. People act on impulse but the genius can inhibit and thus think before acting. And this impulse control set at a new higher layer can control the processes of thought itself where multiple paths exist instead of single focus / single-mindedness.

Hierarchical control. (meta) that introduces a new coordination apparatus. Hidden causality can be understood. The mind can look around the appearance of things and find the true cause and those causes causes. Creating a probabilistic network of higher thought. New ideas can then come from a combination of these causes. Paths through a cross-temporal space.

Access to the unconscious negotiates what is valued with the potential to actualize a result.

For 3D there is NerF and other surrounding works: https://www.matthewtancik.com/nerf

 

[Black Rose]

For 3D there is NerF and other surrounding works: https://www.matthewtancik.com/nerf

Yes, alphafold.

might be able to solve this?

 

[ZenRaiden]

That is because you need a AI that has more than calculator abilities.

Something that builds models of reality and is able to associate those.

Building world-models is not necessarily "more than" calculator abilities. AIs (eg. model-based RL/IL) have been used to make models of their environment, learn their dynamics, they have been also been trained to be grounded in physics; and they have been used to associate multiple modalities of data as well. But fundamentally they are still running on logic gates and logical "calculations". You can make endless complex machines surfing high dimensional geometric manifold from "calculator abilities" alone. There is no a priori reason given why it can't, and steadily growing evidence that shows it can.

Human models are premade.

Everyone can just tell the difference between dogs and butterflies.
Everyone knows dogs are not houses.
This is modeling of world.
Everyone knows 4 people and 4 cars are not the same thing.
Whatever calculation our brains are doing it must be entirely subconscious.
We just see the difference.

higher order thinking is something we mapped out and try to approach brains with.
Our brain base is not higher order thinking, but is vital for life and intelligence.

 

[DoIMustHaveAnUsername?]

For 3D there is NerF and other surrounding works: https://www.matthewtancik.com/nerf

Yes, alphafold.

might be able to solve this?

Probably not. AlphaFold is built for different things working on protien sequences instead of standard visual stuff (but I don't know the nitty gritty). Maybe we can do something with its MCTS algorithm.

 

[DoIMustHaveAnUsername?]

That is because you need a AI that has more than calculator abilities.

Something that builds models of reality and is able to associate those.

Building world-models is not necessarily "more than" calculator abilities. AIs (eg. model-based RL/IL) have been used to make models of their environment, learn their dynamics, they have been also been trained to be grounded in physics; and they have been used to associate multiple modalities of data as well. But fundamentally they are still running on logic gates and logical "calculations". You can make endless complex machines surfing high dimensional geometric manifold from "calculator abilities" alone. There is no a priori reason given why it can't, and steadily growing evidence that shows it can.

Human models are premade.

Everyone can just tell the difference between dogs and butterflies.
Everyone knows dogs are not houses.
This is modeling of world.
Everyone knows 4 people and 4 cars are not the same thing.
Whatever calculation our brains are doing it must be entirely subconscious.
We just see the difference.

higher order thinking is something we mapped out and try to approach brains with.
Our brain base is not higher order thinking, but is vital for life and intelligence.

Do you remember your experiences in the womb? Do you remember having premade models back then?

Is what you are saying based on any evidence in developmental psychology?
Even basic things like object permanence takes time to develop.

And even if some things are "premade" it doesn't really mean anything. Because that would be "initialization" from evolution. You can do the same thing with AI:

Weight Agnostic Neural Networks

Networks that can already (sort of) perform tasks with random weights.

weightagnostic.github.io

And if not evolution you can initialize with "pre-training". That's what done in GPT and co.

I completely agree that "Whatever calculation our brains are doing it must be entirely subconscious." -- i.e we don't really have no idea how we do what we do from first person experience (without further scientific investigation). So there is little a priori basis to say "mere calculator abilities cannot bring out the same results"

"higher order thinking is something we mapped out and try to approach brains with." --- almost everything we talk about are mapped out based on some framework of classification. So I am not sure what's the point here. To talk about anything at all you have to adopt a framework (subconsciously if not consciously) with some boundary conditions to bound objects and classify relations in specific ways.

"Our brain base is not higher order thinking, but is vital for life and intelligence." -- I don't see why they would be dichotomous.