3 ms·

> What seems to make AIs more intelligent is simply their scale. More neurons, more connections (referred to as “parameters” in the community) means better, and

> What seems to make AIs more intelligent is simply their scale. More neurons, more connections (referred to as “parameters” in the community) means better, and techniques that don’t work at one scale may jump to human-level at a larger one. I think for this reason human artists won’t be replaced anytime soon. The human brain is insanely efficient. I have a theory that humans are smart not because of some magic algorithm in the brain, but more because we process so much input we can recognize so many patterns at once. IIRC we‘ve build a super-computer with as much “processing power” as a human brain (although we also don’t know exactly how neurons work). But it’s not just that. We have to build that computer, run it for decades on all kinds of input, and then we have something as smart and creative as one human. Not necessarily even an artistic one either. Meanwhile there are over 7 billion humans, plenty of whom can make art for cheap. And that’s why we’re far from companies using AI to make their graphics. I’d love to see AI stories and art, but until we get much more processing power i doubt we’ll get AI making novel creations. These abstract art pieces? Sure. Transferring styles (e.g. Van Gogh, water color, polygonal) between paintings? Sure. In fact I bet most artists will be using AI-augmented tools. But for a long time AI will still involve human work and/or be relegated to different styles.

6177c40f 5y ago
I've been thinking something along these lines recently too. A single neuron seems like an absurdly complex device. I doubt they are really similar at all to "artificial neurons" except in some vague abstract sense. So the brain is essentially a network of billions of powerful, hyper-efficient computers (not to mention all the other nerves in the body), whose function, communication protocols, and network topology have been developed for millions of years in such a way that has lead to human intelligence. And yet it still takes decades to develop a human mind. Even just to become competent with a single language takes years and it requires much more than just learning that language. If we could develop an AI system that hypothetically could achieve human intelligence, could we even train it to that level faster than it takes a human mind to mature? Is there some kind of maximum rate at which the process by which "intelligence"- whatever that may mean exactly- can be reliably run?
- Jensson 5y ago
  The main advantage of AI is that we can just copy it. So if we take 10 years to train an Einstein level AI in a field, we can copy that a million times and have a million Einsteins. Of course a big part in creating progress is learning new things, so this approach might not create scientific progress faster than current humans, but at least it will be way more efficient for run of the mill knowledge work.

6177c40f 5y ago

I've been thinking something along these lines recently too. A single neuron seems like an absurdly complex device. I doubt they are really similar at all to "artificial neurons" except in some vague abstract sense. So the brain is essentially a network of billions of powerful, hyper-efficient computers (not to mention all the other nerves in the body), whose function, communication protocols, and network topology have been developed for millions of years in such a way that has lead to human intelligence. And yet it still takes decades to develop a human mind. Even just to become competent with a single language takes years and it requires much more than just learning that language. If we could develop an AI system that hypothetically could achieve human intelligence, could we even train it to that level faster than it takes a human mind to mature? Is there some kind of maximum rate at which the process by which "intelligence"- whatever that may mean exactly- can be reliably run?

Jensson 5y ago
The main advantage of AI is that we can just copy it. So if we take 10 years to train an Einstein level AI in a field, we can copy that a million times and have a million Einsteins. Of course a big part in creating progress is learning new things, so this approach might not create scientific progress faster than current humans, but at least it will be way more efficient for run of the mill knowledge work.