I think any prediction based on a ‘singularity’ neglects to consider the physical limitations, and just how long the journey towards significant amounts of AGI would be.
The human brain has an estimated 100 trillion neuronal connections - so probably a good order of magnitude estimation for the parameter count of an AGI model.
If we consider a current GPU, e.g. the 12 GB GFX 3060, it can hold about 24 billion parameters at 4 bit quantisation (in reality a fair few less), and uses 180 W of power. So that means an AGI might use 750 kW of power to operate. A super-intelligent machine might use more. That is a farm of 2500 300W solar panels, while the sun is shining, just for the equivalent of one person.
Now to pose a real threat against the billions of humans, you’d need more than one person’s worth of intelligence. Maybe an army equivalent to 1,000 people, powered by 8,333,333 GPUs and 2,500,000 solar panels.
That is not going to materialise out of the air too quickly.
In practice, as we get closer to an AGI or ASI, there will be multiple separate deployments of similar sizes (within an order of magnitude), and they won’t be aligned to each other - some systems will be adversaries of any system executing a plan to destroy humanity, and will be aligned to protect against harm (AI technologies are already widely used for threat analysis). So you’d have a bunch of malicious systems, and a bunch of defender systems, going head to head.
The real AI risks, which I think many of the people ranting about singularities want to obscure, are:
An oligopoly of companies get dominance over the AI space, and perpetuates a ‘rich get richer’ cycle, accumulating wealth and power to the detriment of society. OpenAI, Microsoft, Google and AWS are probably all battling for that. Open models is the way to battle that.
People can no longer trust their eyes when it comes to media; existing problems of fake news, deepfakes, and so on become so severe that they undermine any sense of truth. That might fundamentally shift society, but I think we’ll adjust.
Doing bad stuff becomes easier. That might be scamming, but at the more extreme end it might be designing weapons of mass destruction. On the positive side, AI can help defenders too.
Poor quality AI might be relied on to make decisions that affect people’s lives. Best handled through the same regulatory approaches that prevent companies and governments doing the same with simple flow charts / scripts.
I think you’re right on the money when it comes to the real dangers, especially your first bullet point. I don’t necessarily agree with your napkin maths. If the virtual neurons are used in a more efficient way, that could make up for a lot versus human neuron count.
The human brain has an estimated 100 trillion neuronal connections - so probably a good order of magnitude estimation for the parameter count of an AGI model.
Yeah, but a lot of those do things unrelated to higher reasoning. A small monkey is smarter than a moose, despite the moose obviously having way more synapses.
I don’t think you can rely on this kind of argument so heavily. A brain isn’t a muscle.
This is something I think needs to be interrogated. None of these models, even the supposedly open ones are actually “open” or even currently “openable”. We can know the exact weights for every single parameter, the code used to construct it, and the data used to train it, and that information gives us basically no insight into its behavior. We simply don’t have the tools to actually “read” a machine learning model in the way you would an open source program, the tech produces black boxes as a consequence of its structure. We can learn about how they work, for sure, but the corps making these things aren’t that far ahead of the public when it comes to understanding what they’re doing or how to change their behavior.
I think any prediction based on a ‘singularity’ neglects to consider the physical limitations, and just how long the journey towards significant amounts of AGI would be.
The human brain has an estimated 100 trillion neuronal connections - so probably a good order of magnitude estimation for the parameter count of an AGI model.
If we consider a current GPU, e.g. the 12 GB GFX 3060, it can hold about 24 billion parameters at 4 bit quantisation (in reality a fair few less), and uses 180 W of power. So that means an AGI might use 750 kW of power to operate. A super-intelligent machine might use more. That is a farm of 2500 300W solar panels, while the sun is shining, just for the equivalent of one person.
Now to pose a real threat against the billions of humans, you’d need more than one person’s worth of intelligence. Maybe an army equivalent to 1,000 people, powered by 8,333,333 GPUs and 2,500,000 solar panels.
That is not going to materialise out of the air too quickly.
In practice, as we get closer to an AGI or ASI, there will be multiple separate deployments of similar sizes (within an order of magnitude), and they won’t be aligned to each other - some systems will be adversaries of any system executing a plan to destroy humanity, and will be aligned to protect against harm (AI technologies are already widely used for threat analysis). So you’d have a bunch of malicious systems, and a bunch of defender systems, going head to head.
The real AI risks, which I think many of the people ranting about singularities want to obscure, are:
I think you’re right on the money when it comes to the real dangers, especially your first bullet point. I don’t necessarily agree with your napkin maths. If the virtual neurons are used in a more efficient way, that could make up for a lot versus human neuron count.
Yeah, but a lot of those do things unrelated to higher reasoning. A small monkey is smarter than a moose, despite the moose obviously having way more synapses.
I don’t think you can rely on this kind of argument so heavily. A brain isn’t a muscle.
This is something I think needs to be interrogated. None of these models, even the supposedly open ones are actually “open” or even currently “openable”. We can know the exact weights for every single parameter, the code used to construct it, and the data used to train it, and that information gives us basically no insight into its behavior. We simply don’t have the tools to actually “read” a machine learning model in the way you would an open source program, the tech produces black boxes as a consequence of its structure. We can learn about how they work, for sure, but the corps making these things aren’t that far ahead of the public when it comes to understanding what they’re doing or how to change their behavior.