I agree that a single deep learning network can only interplay. But a GAN or a self-play network can obviously extrapolate. The question though is where is ingenuity coming from? medium.com/intuitionmachi…
I'm actually very surprised that @ecsquendor who has good videos summarizes the state of understanding in deep learning is fawning over @fchollet ideas. I'm perplexing what Chollet calls extrapolation:
This depiction of accumulation of narrow skills that learning enough skills gets you to a higher level of intelligence is a naive schema. Furthermore, it doesn't help to define the categorization of generalization in this way.
It's as if general intelligence is similar to learning to play the piano or learning how to do math. With enough practice we become proficient in creating our own music or inventing new math. But these two examples assume general intelligence.
Music and math have their own vocabularies. Where do these original vocabularies come from? They come from inventions of a general intelligence. Now ask, where does the vocabulary of general intelligence come from? There is no external designer for this.
To understand general intelligence you have to at least start with at what Marr called the computational level. Sutton is correct in his formulation of RL as a computational level.
There are many alternative computations theories. Any discussion about an approach to intelligence should begin here and not via a premature and arbitrary delineation of categories of generalizations or interpolations. The world generalization is in fact a nebulous concept.
An interesting observation about people who seem to have the highest intelligence also have a peculiar mix of personalities. The big five set which if you tend toward the extreme in all 5 dimensions. Openness Conceintious Extraversion Agreeableness and Calmness.
It's the cognitive inclination of an agent that leads to general intelligence. One might call this more broadly as the intrinsic motivation. This is different from what Sutton proposes as reward is enough. Reward depends on motivation. Not sure how you can avoid this.
The human species has general intelligence as a consequence of their biological capabilities and their unique motivations. Primates are almost biologically like humans, but they don't have the flexible jaws and the dextrous hands of humans.
But what distinguishes humans from primates is the inclination towards shared intentionality. Take this collection of capabilities and a drive towards cooperative behavior and you get general intelligence. Absent these and you don't get there, the reward does not matter!
The biosphere has a multitude of subjective beings. There is an immense benefit for one being to predict the behavior of another subjective being. Herbivores predict plants, carnivores predict herbivores, social carnivores predict other social carnivores, and so on.
In general though, the purpose of prediction is to seek energy. In physics energy is a scalar quantity and have this quality that it's interchangeable (i.e. fungible). This fungibility is not exactly present in complex systems.
This is because what is considered energy differs for every subjective being that consumes energy. Plants can consume sunlight, humans cannot. Humans today need money to get energy. Money is a kind of energy.
It's fascinating that all complex multicellular creatures are composed of eukaryotes. Cells that are symbiosis of a bacteria and an archea (i.e. mitochondria). That mitochondria is the battery of the cell.
The reason why biology is not like physics is because its got these batteries. There are like memories of energy. In the models of how life originates, the focus is on how persistent structures are created out of energy.
What is often overlooked is how biology has created persistent energy stores that are capable of modulating the actions of a subjective agent. Complex behavior implies control of the release of energy. Physics does not cover this control.
The point here is that all intelligence revolves around the acquisition and control of energy. In human civilization this translates to the acquisition and control of money. The game is different for each different species, but it always revolves around some kind of energy.
This is perhaps why Friston is so found with his Free Energy principle as the core of cognition. Homeostatis also revolves around the idea of energy conservation. Rewards are also just a proxy to energy.
What is often missing in this framing is that complex systems have multitudes of complexity and each level of complexity has a different kind of energy that drives it. Competence at each level implies competence in the control and acquisition of this energy.
Where does the competence of these skills come from?
Does it also not feel a bit odd that the intelligence (or competence) to acquire different kinds of energy exist in the multitude of single-celled and multicellular organisms that occupy the biosphere?
There's immense complexity that resides inside all of biology. This is where general intelligence comes from. It's in the wetware. Understanding biology implies understanding cognition. The ironic thing is, computers are not biological.
It is often said that money buys freedom. But it is rarely mentioned that money also buys stability and reduces uncertainty. This other utility is what people seek more than freedom.
A majority of us will sacrifice our freedom and our youth for the stability and certainty of a steady paycheck. They say money buys happiness...
but in reality, it buys predictability. Because as humans, we value competence in this world and you cannot feel competence if you cannot predict the world.
An unexplored scenario for humanity is that as technology becomes more advanced, we become more adept at detecting aliens. But the aliens have no interest in our affairs, so we know of their existence but they never attempt to interact.
The prime directive envisioned in Star Trek may in fact exist and we just happen to be that backward civilization. It's just like how we treat tribes in the Amazon. We allow them to thrive in complete isolation.
Any advanced civilization will likely have a history where members of their race chose to live in isolation. So it should be no surprise that they would leave our civilization alone to grow up for ourselves.
So I'm reading this blog from Walid Saba where he lends 5 examples that are difficult to do in NLP. medium.com/ontologik/sema…
(1) Sara likes to play bridge (2) Sara has a Greek statue in every corner of her apartment (3) Sara loves to eat pizza with her kids (with pineapple) (4) Sara enjoyed the movie (5) The White House criticized the recent remarks made by Beijing
I ran these against GPT-3 to validate its understanding of them. GPT-3 appeared to 'know' that there was more than 1 statue for (2) and wasn't able to resolve what (4) might mean. However, it disambiguated the others correctly.
Artificial General Intelligence is the field with a quest to automate human cognition. AGI is not superintelligence.
Computers and handheld calculators can perform calculations that are beyond the capabilities of any human. We don't call these intelligent things, that's despite 'superhuman' computatoinal capability.
Can AGI lead to superintelligence? The current consensus is yes.
The only thing that remains constant is change. If you think about it, what is constant is relative to time. Furthermore, what is constant is relative to what is moving. For physicists, something that is constant (i.e invariant) describes symmetry.
A symmetry is furthermore defined as a *change* in a reference frame. In otherwords, you can't define a constant unless there is something that changes.
What is constant is what is perceived to not change. It is the very definition of an abstraction. We are able to perceive the world because we abstract the world into things that don't change. Categories are also things that don't change.
The "Reward is Enough" paper offers a piss-poor explanation as to how AGI is achieved. I'm actually more surprised that the @DeepMind wrote such a poorly constructed philosophical paper. sciencedirect.com/science/articl…
The major flaw of the 'Reward is Enough' paper is that the authors don't know how to disentangle the self-referential nature of the problem. So it's written in a way that sounds like a bunch of poorly hidden tautological statements.