Similarly, coordinating thousands or millions of AGI instances into "group agents" or virtual agent economies could produce collective superintelligence. What is unclear today is how AI group intelligence could scale as we add more agents and compute, depending on how agents are organized.
We may not need "collective superintelligence". Modest AI intelligence plus ordinary organization may be enough to do the job a company's management does. An org chart where most of the boxes are AIs is a likely near future.
An organization of AIs has some advantages over a human organization, even if the AIs are not super-bright. They have far more communications bandwidth. They can ingest the info on a new subject very fast, as we see when Claude reads a Git repository. Onboarding a new AI is very fast. Layoffs are not a problem. The human resource problems go away along with the humans. So there's a lot more organizational flexibility. Starting up a project means initializing some agent instances.
There's a whole theory and practice of AI organizational behavior waiting to be developed.
Whether it will be developed by humans or by AIs remains to be seen. It's going to happen.
We are treating AGI like is something that we can measure. We don't even know what intelligence really is let alone measure it.
Maybe we can get something that resembles intelligence but in reality is another thing. Maybe intelligence is emergent but not the only thing that can emerge, maybe there are things that share some characteristics with the intelligence we know and aren't intelligence.
They get asked about it pretty often (frontier lab ceo) but ofc none of them will say yeah we'll fire everyone the moment we can. Here is a Sam Altman blog on the topic: https://blog.samaltman.com/three-observations
The math they propose to justify their predictions is so problematic that I struggle taking the more abstract parts seriously. You can't just extrapolate the percentage change in AI investment over the last 5 years and expect investors to keep upping the ante at the same rate. There's only so much GDP to sink into it - and we're already pretty capped out.
> Interestingly, we have a good mathematical understanding of the very upper end of the intelligence continuum - the theoretical limits of general intelligence.
Feels strange to use our current understanding of mathematics and physics to determine the limits of something that would obviously have a much better understanding of mathematics and physics.
we need a less subjective and arbitrary definition for "AGI" so the current crop of LLMs can stop pretending to have achieved what most people know in their gut is more than sampling from a probability distribution of learned behavior.
> Now that AGI is coming within reach, we can no longer put off seriously considering what comes next. While far from certain, AI could continue to advance at its current pace - and possibly even accelerate - towards artificial general superintelligence, or ASI.
These people are so full of shit it is amusing to read this on a regular basis. Each time I'm looking forward to how they will spin it.
If the current LLM providers were near AGI they would be using it to run businesses instead of selling shovels. There doesn't seem to be an existing LLM that can be told "Go make money" and have it do so. If they get to that point they are not going to sell it, because why would they?
If almost everyone loses their job at once, why won't all those people respond by voting for socialist candidates promising radical economic redistribution?
Socialism seems to already be growing in popularity in the US (traditionally the most anti-socialist Western nation) even without AGI firing everybody.
> Now that AGI is coming within reach, we can no longer put off seriously considering what comes next. While far from certain, AI could continue to advance at its current pace - and possibly even accelerate - towards artificial general superintelligence, or ASI
I wonder why they are so uncertain about it. It seems obvious to me that AGI will certainly lead to ASI. Unless human-level intelligence is something like the speed of light?
I don't even believe there is a clear line between AGI and ASI.
> Unless human-level intelligence is something like the speed of light?
I'm sceptical we will reach fully human-level intelligence simply by scaling current AI approaches, since they are so extremely inefficient compared to human brains, and have other severe limitations such as hard boundaries between training and runtime inference.
Sure, people are working on new architectures that aim to overcome that, but we don't know yet whether those proposals will actually succeed in practice.
Yeah, the machines will be connected, communicating at lightning speed, and working 24/7. Seems like AGI would make superhuman intelligence a forgone conclusion.
I feel like claiming that “today’s models can literally memorize large parts of the internet” as a positive attribute when the defense to most AI copyright lawsuits is “no they can’t, and if they can they shouldn’t” may be ill-advised.
I’m curious: has the definition of AGI ever been agreed upon? Honestly, what people call ASI now seems to be what AGI used to mean. I’m not sure. People say AGI, but what is the reference point for AGI and what is the social consensus around it, and who set that standard?
In fact, working in IT, what I feel is that there are too many terms that are not defined and that each person uses differently. So it seems like rather than AI having advanced, someone arbitrarily set a finish line called AGI and then said we have reached it. Actually, in the books of singularity advocates, what they called AGI back then seems to be what we now call ASI.
Sometimes I think about this. How on earth can we measure an abstract concept like intelligence? The concept of intelligence we approach now seems to be about whether something can create problems and solve those problems. If so, what kind of intelligence does current AI have? Since we cannot directly face the abstract entity called intelligence, it seems like we are trying to guess its shape by looking at shadows. But whether current AI even has intelligence feels a bit difficult to say.
Even if I use the best frontier AI model accessible on the market, not the hidden latest models inside frontier labs, it still often makes wrong mistakes that humans would not make unless it works alongside humans. In programming, there are cases where AI tries to explicitly handle things that humans would implicitly pass over, and that actually makes things go wrong.
I’m not sure. Whether we should call this intelligence.
We may not need "collective superintelligence". Modest AI intelligence plus ordinary organization may be enough to do the job a company's management does. An org chart where most of the boxes are AIs is a likely near future.
An organization of AIs has some advantages over a human organization, even if the AIs are not super-bright. They have far more communications bandwidth. They can ingest the info on a new subject very fast, as we see when Claude reads a Git repository. Onboarding a new AI is very fast. Layoffs are not a problem. The human resource problems go away along with the humans. So there's a lot more organizational flexibility. Starting up a project means initializing some agent instances.
There's a whole theory and practice of AI organizational behavior waiting to be developed. Whether it will be developed by humans or by AIs remains to be seen. It's going to happen.
Maybe we can get something that resembles intelligence but in reality is another thing. Maybe intelligence is emergent but not the only thing that can emerge, maybe there are things that share some characteristics with the intelligence we know and aren't intelligence.
I'm not being snarky, that seems like a natural question for anyone interviewing a frontier lab CEO, but I haven't seen it asked.
I am sure Dario has a bigger more cringe one
Feels strange to use our current understanding of mathematics and physics to determine the limits of something that would obviously have a much better understanding of mathematics and physics.
These people are so full of shit it is amusing to read this on a regular basis. Each time I'm looking forward to how they will spin it.
* Citation needed
If the current LLM providers were near AGI they would be using it to run businesses instead of selling shovels. There doesn't seem to be an existing LLM that can be told "Go make money" and have it do so. If they get to that point they are not going to sell it, because why would they?
Socialism seems to already be growing in popularity in the US (traditionally the most anti-socialist Western nation) even without AGI firing everybody.
I wonder why they are so uncertain about it. It seems obvious to me that AGI will certainly lead to ASI. Unless human-level intelligence is something like the speed of light?
I don't even believe there is a clear line between AGI and ASI.
I'm sceptical we will reach fully human-level intelligence simply by scaling current AI approaches, since they are so extremely inefficient compared to human brains, and have other severe limitations such as hard boundaries between training and runtime inference.
Sure, people are working on new architectures that aim to overcome that, but we don't know yet whether those proposals will actually succeed in practice.
In fact, working in IT, what I feel is that there are too many terms that are not defined and that each person uses differently. So it seems like rather than AI having advanced, someone arbitrarily set a finish line called AGI and then said we have reached it. Actually, in the books of singularity advocates, what they called AGI back then seems to be what we now call ASI.
Sometimes I think about this. How on earth can we measure an abstract concept like intelligence? The concept of intelligence we approach now seems to be about whether something can create problems and solve those problems. If so, what kind of intelligence does current AI have? Since we cannot directly face the abstract entity called intelligence, it seems like we are trying to guess its shape by looking at shadows. But whether current AI even has intelligence feels a bit difficult to say.
Even if I use the best frontier AI model accessible on the market, not the hidden latest models inside frontier labs, it still often makes wrong mistakes that humans would not make unless it works alongside humans. In programming, there are cases where AI tries to explicitly handle things that humans would implicitly pass over, and that actually makes things go wrong.
I’m not sure. Whether we should call this intelligence.