4 ms·

“Soon it will reach levels of complexity vastly exceeding any human brain.” I doubt the “soon” part. Artificial neurons are vast simplifications of real neuron

“Soon it will reach levels of complexity vastly exceeding any human brain.” I doubt the “soon” part. Artificial neurons are vast simplifications of real neurons, and even complex networks like GPT don’t come close the complexity of biological networks, both in network structure and in terms of what goes on between the neurons (eg chemical processes, as opposed to just the activity of the neurons themselves). Individual neurons have been shown to have capability to both process and store information, something which artificial ones don’t do. Besides, we are only scratching the surface in our understanding of biological neurons and brains. How can we say “this thing we built that is a vast simplification of this thing we barely understand will soon exceed the complexity of the thing we don’t yet understand”?

p1esk 2y ago
Because it’s already able to do almost everything a brain can? State of the art AI models can already learn, reason, communicate, and even create - better than most humans. Using a lot less neurons, and much simpler neuronal structures. All trends indicate we’re only a couple years away from an AI superintelligence. No additional understanding of biological brains is required to get there.
- altruios 2y ago
  I remember a similar argument made since the 1980's... 40 years later and a lot of stall-outs... It could be hubris to assume we know enough yet to replicate our 'kind' of intelligence. Just as it might be hubris to assert AI doesn't have some 'kind' of experience. A minimum (which will be raised once we have a better understanding) is still: change of state, to have an experience - we simply don't know what else exactly is required, so that threshold minimum remains to be raised.
  - p1esk 2y ago
    This argument could not be made before March 14 2023, when the first actual AI was released (GPT-4). I remember that day very well because I was extensively testing every GPT model before that (part of my job as an AI researcher). The entire history of human civilization can be separated into before and after that milestone. We do not need to "replicate" human intelligence. It's enough to "simulate" it. The coming AI models will be entirely "alien" types of intelligence to us, and that's OK as long as they are useful. Most likely these AI models will be able to finally explain to us how our own brains work.
    - dkersten 2y ago
      The GPT4 you’re using must be a different one form what I’m using, because for me GPT4 was only an incremental improvement over GPT3.5 and while both are a step up from everything that came before, it’s output is still very mechanical to the point where it often seems like little more than a very elaborate and sophisticated expert system with probabilistic response choices and a large knowledge base. It has far too many shortcomings to make the claims you seem to be making and the worst part is that it’s entirely unaware of its own shortcomings and will confidently and incorrectly assert things without any self awareness or understanding and I’ve yet to see a single model that can determine when it doesn’t know something. There’s still a long way to go before human level intelligence. Also, think of the amount of energy that is required to run GPT4 compared to a typical human brain. That’s not to say that it’s not useful in its current form, because it is, but that’s not what we are arguing here. You asserted that “Soon it will reach levels of complexity vastly exceeding any human brain.” while I’m asserting that the human brain is far, far more complex than you give it credit for and that the latest AI models are not even close to that yet. I am doubting that LLM’s and similar machine learning will exceed human brain complexity, and certainly not VASTLY, in any timeframe that I would call “soon”, unless there are fundamental discoveries made. It’s not even clear the current models can scale any further due to the lack of any new non-LLM-generated training data to train bigger models on, so the GPT3.5 to 4 shift is unlikely to reoccur.

p1esk 2y ago

Because it’s already able to do almost everything a brain can? State of the art AI models can already learn, reason, communicate, and even create - better than most humans. Using a lot less neurons, and much simpler neuronal structures. All trends indicate we’re only a couple years away from an AI superintelligence. No additional understanding of biological brains is required to get there.

altruios 2y ago
I remember a similar argument made since the 1980's... 40 years later and a lot of stall-outs... It could be hubris to assume we know enough yet to replicate our 'kind' of intelligence. Just as it might be hubris to assert AI doesn't have some 'kind' of experience. A minimum (which will be raised once we have a better understanding) is still: change of state, to have an experience - we simply don't know what else exactly is required, so that threshold minimum remains to be raised.
- p1esk 2y ago
  This argument could not be made before March 14 2023, when the first actual AI was released (GPT-4). I remember that day very well because I was extensively testing every GPT model before that (part of my job as an AI researcher). The entire history of human civilization can be separated into before and after that milestone. We do not need to "replicate" human intelligence. It's enough to "simulate" it. The coming AI models will be entirely "alien" types of intelligence to us, and that's OK as long as they are useful. Most likely these AI models will be able to finally explain to us how our own brains work.
  - dkersten 2y ago
    The GPT4 you’re using must be a different one form what I’m using, because for me GPT4 was only an incremental improvement over GPT3.5 and while both are a step up from everything that came before, it’s output is still very mechanical to the point where it often seems like little more than a very elaborate and sophisticated expert system with probabilistic response choices and a large knowledge base. It has far too many shortcomings to make the claims you seem to be making and the worst part is that it’s entirely unaware of its own shortcomings and will confidently and incorrectly assert things without any self awareness or understanding and I’ve yet to see a single model that can determine when it doesn’t know something. There’s still a long way to go before human level intelligence. Also, think of the amount of energy that is required to run GPT4 compared to a typical human brain. That’s not to say that it’s not useful in its current form, because it is, but that’s not what we are arguing here. You asserted that “Soon it will reach levels of complexity vastly exceeding any human brain.” while I’m asserting that the human brain is far, far more complex than you give it credit for and that the latest AI models are not even close to that yet. I am doubting that LLM’s and similar machine learning will exceed human brain complexity, and certainly not VASTLY, in any timeframe that I would call “soon”, unless there are fundamental discoveries made. It’s not even clear the current models can scale any further due to the lack of any new non-LLM-generated training data to train bigger models on, so the GPT3.5 to 4 shift is unlikely to reoccur.