Interesting thoughts! Although I'd push back on some of the arguments here.
We can create AGI without first understanding creativity. Understanding the theory is not a prerequisite to creating a practical implementation. Humans can and have created transformational technology without first understanding how it worked. The Wright Brothers created airplanes without understanding the principles of aerodynamics. Fleming created Penicillin, and only decades later did we understand how the underlying mechanics and properties worked.
I'd also argue that the "mix-mashing" of data you see in current models is just an artifact of the models not being large enough to fully represent the subset of reality they're trained on. Fundamentally, LLMs attempt to represent the world in their matrices of numbers. Information Theory shows us how to represent any arbitrary amount of information in bits -- which is exactly what LLMs do.
DALL-E showed us that LLMs are capable of representing Van Gogh's artistic style in their massive matrices. ChatGPT showed us that human language and computer code can be represented accurately enough as vectors for the LLM to produce *and explain* (!!) correct output. In the limit, a large enough model is theoretically capable of representing *and explaining* the nature of reality in its massive arrays of vectors; including language, emotions, mathematics, physics, etc... at which point I'd posit that it's capable of producing novel explanations about reality.
"The Wright Brothers created airplanes without understanding the principles of aerodynamics". Wrong: https://wright.grc.nasa.gov/overview.htm. Also, experimental results are different from creating an explanation.
Oh haha, bad example. But you get the point right? We can create something before having a good explanation for it. We can create AGI without a good explanation for what is creativity or intelligence.
Even the invention of controlled fire required *an explanation*. For ex. after a lightning lit up some forest and everything was cooked people came up with some explanation of what happened to then replicate it in a controlled way. This first theory might have been something like "bright light from the sky when storm creates hot dangerous red stuff that kills animals and makes them taste better" (assuming they didn't even have words for lightning, fire, and cooking).
Or do you have an example of an invention that occurred without a somewhat good explanation?
So if *an explanation* is required, that means even a *bad explanation* is sufficient as a theory to create some technology? The example you provided is not a good explanation by today’s standards, but sufficient as an explanation to create fire.
If so, I think we actually agree here. We may not have the *correct explanation* before developing some technology, we can discover the correct explanation after the technology gets created. And we already have explanations and ongoing debates for why “more compute => AGI.”
"I'd also argue that the 'mix-mashing' of data you see in current models is just an artifact of the models not being large enough to fully represent the subset of reality they're trained on."
My full argument, is that a real world humanoid agent with a large enough neural network and good data engine to train and evolve it—essentially what the Tesla Bot is aiming for—would be able to represent enough in its matrices of numbers about the nature of reality to be able to “intelligently” explain and interact with the world.
I’m not fully convinced of the inductive/abductive reasoning that article is premised on. There’s no law of nature that says human reasoning (abductive reasoning) is the only way to achieve intelligence.
It seems like most arguments for and against AGI run up against the constraint that we don’t fully understand intelligence.
Thats essentially the emergentist's argument. I.e. that general intelligence will somehow emerge from a bigger neural net. There is no valid reason why that should happen. Even human (AGI) didn't suddenly become smart once their brain reached a certain size. It's more plausible that they acquired an algorithm.
There’s also no good argument for why AGI can’t emerge from a large enough NN.
We don’t understand intelligence yet. So I don’t think we have ruled out the possibility of emergent AGI.
Even the Deutschian argument is merely based off of the empirical observation that LLMs don’t *seem* to understand the output they’re producing. How do we know that?
Some will also argue that “it’s just linear algebra, how can that be intelligent?”. But you can also take a reductionist view on how the human brain’s just a bunch of electrical signals in a heap of meat.
Interesting thoughts! Although I'd push back on some of the arguments here.
We can create AGI without first understanding creativity. Understanding the theory is not a prerequisite to creating a practical implementation. Humans can and have created transformational technology without first understanding how it worked. The Wright Brothers created airplanes without understanding the principles of aerodynamics. Fleming created Penicillin, and only decades later did we understand how the underlying mechanics and properties worked.
I'd also argue that the "mix-mashing" of data you see in current models is just an artifact of the models not being large enough to fully represent the subset of reality they're trained on. Fundamentally, LLMs attempt to represent the world in their matrices of numbers. Information Theory shows us how to represent any arbitrary amount of information in bits -- which is exactly what LLMs do.
DALL-E showed us that LLMs are capable of representing Van Gogh's artistic style in their massive matrices. ChatGPT showed us that human language and computer code can be represented accurately enough as vectors for the LLM to produce *and explain* (!!) correct output. In the limit, a large enough model is theoretically capable of representing *and explaining* the nature of reality in its massive arrays of vectors; including language, emotions, mathematics, physics, etc... at which point I'd posit that it's capable of producing novel explanations about reality.
"The Wright Brothers created airplanes without understanding the principles of aerodynamics". Wrong: https://wright.grc.nasa.gov/overview.htm. Also, experimental results are different from creating an explanation.
Oh haha, bad example. But you get the point right? We can create something before having a good explanation for it. We can create AGI without a good explanation for what is creativity or intelligence.
No.
Even the invention of controlled fire required *an explanation*. For ex. after a lightning lit up some forest and everything was cooked people came up with some explanation of what happened to then replicate it in a controlled way. This first theory might have been something like "bright light from the sky when storm creates hot dangerous red stuff that kills animals and makes them taste better" (assuming they didn't even have words for lightning, fire, and cooking).
Or do you have an example of an invention that occurred without a somewhat good explanation?
So if *an explanation* is required, that means even a *bad explanation* is sufficient as a theory to create some technology? The example you provided is not a good explanation by today’s standards, but sufficient as an explanation to create fire.
If so, I think we actually agree here. We may not have the *correct explanation* before developing some technology, we can discover the correct explanation after the technology gets created. And we already have explanations and ongoing debates for why “more compute => AGI.”
"I'd also argue that the 'mix-mashing' of data you see in current models is just an artifact of the models not being large enough to fully represent the subset of reality they're trained on."
So you're saying "if the LLM model is large enough general intelligence abilities will ~emerge~"? I like Erik J. Larson's critique of this "emergentist" argument in The Myth of Artificial Intelligence (see here for a summary https://medium.com/codex/can-computers-think-like-humans-reviewing-erik-larsons-the-myth-of-artificial-intelligence-4dc282847318).
My full argument, is that a real world humanoid agent with a large enough neural network and good data engine to train and evolve it—essentially what the Tesla Bot is aiming for—would be able to represent enough in its matrices of numbers about the nature of reality to be able to “intelligently” explain and interact with the world.
I’m not fully convinced of the inductive/abductive reasoning that article is premised on. There’s no law of nature that says human reasoning (abductive reasoning) is the only way to achieve intelligence.
It seems like most arguments for and against AGI run up against the constraint that we don’t fully understand intelligence.
Thats essentially the emergentist's argument. I.e. that general intelligence will somehow emerge from a bigger neural net. There is no valid reason why that should happen. Even human (AGI) didn't suddenly become smart once their brain reached a certain size. It's more plausible that they acquired an algorithm.
There’s also no good argument for why AGI can’t emerge from a large enough NN.
We don’t understand intelligence yet. So I don’t think we have ruled out the possibility of emergent AGI.
Even the Deutschian argument is merely based off of the empirical observation that LLMs don’t *seem* to understand the output they’re producing. How do we know that?
Some will also argue that “it’s just linear algebra, how can that be intelligent?”. But you can also take a reductionist view on how the human brain’s just a bunch of electrical signals in a heap of meat.
> There’s also no good argument for why AGI can’t emerge from a large enough NN.
So what. This is the same as saying there is no good argument to prove that god does not exist.
Yes correct. We don’t know if god exists or not.
But — tying this back — the original post took a definitive stance *against* emergent AGI.
> AGI won’t emerge as a product of this ever-increasing complexity.
It’s the decisiveness that I think is flawed.