This post is a reply to the article "AI Is Evolving — And Changing Our Understanding Of Intelligence" by Blaise Agüera Y Arcas and James Manyika in Noema Magazine.
Thanks. I don't get it. That intelligence is symbolic is not a new idea.
Honestly, I think you should read up on the work of others. Historically, breakthroughs in complex matters never come from some genius with his or her own very original idea. It always builds on others. All breakthroughs are by groups or people who are very deeply into the expertise of the field.
Also, if you're using LLMs as an echo chamber, you need to quit that.
I explain what symbols are - ranges of comparable properties. Consider the property "color" and you have "red", "blue", etc. as symbols. They are defined by boundaries. If any shade falls in between, it is viewed as fitting into that symbol. I haven't seen such an explanation before - so how can I base this idea on someone else' ideas?
The number and boundaries of ranges is not and cannot be fixed once and for all. It is purpose-dependent. Kids operate with a limited number. Designers use more. Hospitals operate with one set of temperature ranges, superconductor physics with another.
Mistakes are possible - because of poor visibility, noise, misleading details, etc.
Chapek described robots more than a century ago. Will you say that Boston Dynamics hasn't done anything new?
Of course they have done something new. But they didn't invent robots. And above all, nobody at Boston Dynamics made huge breakthroughs out of nowhere.
I'm not saying your not saying anything new at all. But superficially, at least, it looks to me like you are ignoring the vast work that has been done in the field.
Does your theory make any predictions or explain something that others fail to do? How?
There's an awful lot of insight about colour perception, for example. I don't see why there's any mystery that needs a new theory. Dennett and functionalist neuroscientists and cognitive scientists have figured it out quite well, I think.
We may be wrong but it's not a problem. If two people are equally wrong they will be able to talk about that object. If they have opposite opinions they may decide to use a different property to point at the object or they will ask clarifying questions. Illusion is a temporary problem, just like paradox.
I've just seen too many people, myself included, believing they had it all figured out on their own. The problem is that they waste too much of their life on it, and deceiving themselves that they are geniuses and that the reason it's not widely appreciated is because people don't get it and they are misunderstood geniuses. I hope that doesn't apply to you!
So, with colour illusions like the checkerboard shadow, is the same symbol active in the two boxes, or two different ones?
What determines which symbol? The retina? More downstream?
I can really recommend "consciousness explained" by dennett.
Objects have properties, actions change properties. Therefore, we need to shift focus to and operate with properties.
Phenomena are all unique but some are interchangeable. Point-measurements do not allow for interchangeability, only range measurements do. Ranges of properties are concepts (categories, classes).
A class is defined by differences from other classes not by similarities of its instances. This enables O(logN) complexity of recognition.
Intelligence does not rely on logic or statistics. It selects from available options with respect to available constraints. In 20 Questions, options = known categories, constraints = properties of an object to be recognized.
Now, having this theory, I also don't view intelligence as mysterious. But we do need this theory because I haven't seen a theory that explains how it works. There are theories that talk about measuring intelligent abilities but it's different.
Everything is recognized in comparison. This makes comparison-based selection the essence of intelligence and what stands behind "cognitive computation". It is efficient which you cannot say about exemplar/prototype theory of object recognition. It explains what concepts are and how to use them in language. It demystifies generalization, representations, memory, etc.
It's very much needed theory at this moment when people are crazy about building AGI thinking that prediction or reward is enough. To reach AGI we need to know the exact mechanism of intelligence. I claim that I discovered it.
Thanks. I don't get it. That intelligence is symbolic is not a new idea.
Honestly, I think you should read up on the work of others. Historically, breakthroughs in complex matters never come from some genius with his or her own very original idea. It always builds on others. All breakthroughs are by groups or people who are very deeply into the expertise of the field.
Also, if you're using LLMs as an echo chamber, you need to quit that.
I say this with sincere intentions.
I explain what symbols are - ranges of comparable properties. Consider the property "color" and you have "red", "blue", etc. as symbols. They are defined by boundaries. If any shade falls in between, it is viewed as fitting into that symbol. I haven't seen such an explanation before - so how can I base this idea on someone else' ideas?
I am not using LLMs at all. That's my principle.
I'm glad to hear that you don't use LLMs.
Okay, so you're saying there's a fixed number of color symbols? What about colour illusions? Like the checker board shade illusion.
I agree that everything is symbolic in a sense but I this is not new.
The number and boundaries of ranges is not and cannot be fixed once and for all. It is purpose-dependent. Kids operate with a limited number. Designers use more. Hospitals operate with one set of temperature ranges, superconductor physics with another.
Mistakes are possible - because of poor visibility, noise, misleading details, etc.
Chapek described robots more than a century ago. Will you say that Boston Dynamics hasn't done anything new?
Of course they have done something new. But they didn't invent robots. And above all, nobody at Boston Dynamics made huge breakthroughs out of nowhere.
I'm not saying your not saying anything new at all. But superficially, at least, it looks to me like you are ignoring the vast work that has been done in the field.
Does your theory make any predictions or explain something that others fail to do? How?
There's an awful lot of insight about colour perception, for example. I don't see why there's any mystery that needs a new theory. Dennett and functionalist neuroscientists and cognitive scientists have figured it out quite well, I think.
In colour illusions, are we wrong about which symbol is active?
We may be wrong but it's not a problem. If two people are equally wrong they will be able to talk about that object. If they have opposite opinions they may decide to use a different property to point at the object or they will ask clarifying questions. Illusion is a temporary problem, just like paradox.
That's sound thinking!
I've just seen too many people, myself included, believing they had it all figured out on their own. The problem is that they waste too much of their life on it, and deceiving themselves that they are geniuses and that the reason it's not widely appreciated is because people don't get it and they are misunderstood geniuses. I hope that doesn't apply to you!
So, with colour illusions like the checkerboard shadow, is the same symbol active in the two boxes, or two different ones?
What determines which symbol? The retina? More downstream?
I can really recommend "consciousness explained" by dennett.
Hi. Can you sum up your theory in a few lines?
Hi! Sure!
Objects have properties, actions change properties. Therefore, we need to shift focus to and operate with properties.
Phenomena are all unique but some are interchangeable. Point-measurements do not allow for interchangeability, only range measurements do. Ranges of properties are concepts (categories, classes).
A class is defined by differences from other classes not by similarities of its instances. This enables O(logN) complexity of recognition.
Intelligence does not rely on logic or statistics. It selects from available options with respect to available constraints. In 20 Questions, options = known categories, constraints = properties of an object to be recognized.
Thanks! I'm afraid you lost me at the first sentence. I don't understand anything of that, or why it's a theory of intelligence.
Doesn't your model roughly map on to some other model of the mind or intelligence? Maybe that would be helpful.
Personally, I don't see what's mysterious about intelligence or why we need new theories
Now, having this theory, I also don't view intelligence as mysterious. But we do need this theory because I haven't seen a theory that explains how it works. There are theories that talk about measuring intelligent abilities but it's different.
Everything is recognized in comparison. This makes comparison-based selection the essence of intelligence and what stands behind "cognitive computation". It is efficient which you cannot say about exemplar/prototype theory of object recognition. It explains what concepts are and how to use them in language. It demystifies generalization, representations, memory, etc.
It's very much needed theory at this moment when people are crazy about building AGI thinking that prediction or reward is enough. To reach AGI we need to know the exact mechanism of intelligence. I claim that I discovered it.