In the past (say, before 2015), there had been multi-decade delays in adoption even of already known innovations which turned to be very fruitful solely because of insufficient manpower combined with tendencies of people to be a bit too conservative (ReLU, residual streams in feedforward machines, etc, etc). Even “attention” and Transformers seem to have come much later in the game than one would naturally expect looking back. “Attention” is such a simple thing, it’s difficult to understand why it has not been well grasped before 2014.
In the present, the main push for alternative architectures and alternative optimization algorithms comes from various underdogs (the leaders seem to be more conservative with the focus of their innovation, because this relatively conservative focus seems to pay well; the net result is that exploration of radically new architectures and radically new algorithms is still quite a bit underpowered, even with a lot of people working in the field).
I can agree that qualitatively there is a lot left to do. Quantitatively, though, I am still not quite seeing the smoking gun that human level AI will be able to smash through 15 OOM like this. But, happy to change my mind. I’ll check out the anthropic link! Cheers.
Oh, when one is trying to talk about that many orders of magnitude, they are just doing “vibe marketing” :-) In reality, we just can’t extrapolate this far. It’s quite possible, but we can’t really know...
But no, it’s not the human level AI, the AI capability is what is changing the fastest in this scenario, the actual reason why it might go that far (and even further) is that a human level AI is supposed to rapidly become superhuman (if it stays at human level then what is all this extra AI research even doing?), and then even more superhuman, and then even more superhuman, and so on, and if there is some saturation at some point it is usually assumed to be very far above the human level.
If one has a lot of AI research done by artificial AI researchers, one would have to impose some very strong artificial constraints to prevent that research from improving the strength of artificial AI researchers. The classical self-improvement scenario is that artificial AI researchers making much better and much stronger artificial AI researchers is the key focus of AI research, and that this “artificial AI researchers making much better and much stronger artificial AI researchers” step iterates again and again.
Logically, I agree. Intuitively, I feel suspect that it just won’t happen. But, intuition on such alien things should not be a guide, so I fully support some attempt to slow down the takeoff.
We at least know two things.
In the past (say, before 2015), there had been multi-decade delays in adoption even of already known innovations which turned to be very fruitful solely because of insufficient manpower combined with tendencies of people to be a bit too conservative (ReLU, residual streams in feedforward machines, etc, etc). Even “attention” and Transformers seem to have come much later in the game than one would naturally expect looking back. “Attention” is such a simple thing, it’s difficult to understand why it has not been well grasped before 2014.
In the present, the main push for alternative architectures and alternative optimization algorithms comes from various underdogs (the leaders seem to be more conservative with the focus of their innovation, because this relatively conservative focus seems to pay well; the net result is that exploration of radically new architectures and radically new algorithms is still quite a bit underpowered, even with a lot of people working in the field).
So at least the model in https://www.lesswrong.com/posts/Nsmabb9fhpLuLdtLE/takeoff-speeds-presentation-at-anthropic does not seem to be exaggerated. There are so many things which people would like to try and can’t find time or personal energy, and so many more further things they would want to try, if they had better grasp of the vast research literature…
I can agree that qualitatively there is a lot left to do. Quantitatively, though, I am still not quite seeing the smoking gun that human level AI will be able to smash through 15 OOM like this. But, happy to change my mind. I’ll check out the anthropic link! Cheers.
Oh, when one is trying to talk about that many orders of magnitude, they are just doing “vibe marketing” :-) In reality, we just can’t extrapolate this far. It’s quite possible, but we can’t really know...
But no, it’s not the human level AI, the AI capability is what is changing the fastest in this scenario, the actual reason why it might go that far (and even further) is that a human level AI is supposed to rapidly become superhuman (if it stays at human level then what is all this extra AI research even doing?), and then even more superhuman, and then even more superhuman, and so on, and if there is some saturation at some point it is usually assumed to be very far above the human level.
If one has a lot of AI research done by artificial AI researchers, one would have to impose some very strong artificial constraints to prevent that research from improving the strength of artificial AI researchers. The classical self-improvement scenario is that artificial AI researchers making much better and much stronger artificial AI researchers is the key focus of AI research, and that this “artificial AI researchers making much better and much stronger artificial AI researchers” step iterates again and again.
Logically, I agree. Intuitively, I feel suspect that it just won’t happen. But, intuition on such alien things should not be a guide, so I fully support some attempt to slow down the takeoff.