No doubt from me that it’s a sigmoid, but how high is the plateau? That’s also hard to know from early in the process, but it would be surprising if there’s not a fair bit of progress left to go.
Human brains seem like an existence proof for what’s possible, but it would be surprising if humans also represent the farthest physical limits of what’s technologically possible without the constraints of biology (hip size, energy budget etc).
Biological muscles are proof that you can make incredibly small and forceful actuators. But the state of robotics is nowhere near them, because the fundamental construction of every robotic actuator is completely different.
We’ve been building actuators for 100s of years and we still haven’t got anything comparable to a muscle. And even if you build a better hydraulic ram or brushless motor driven linear actuator you will still never achieve the same kind of behaviour, because the technologies are fundamentally different.
I don’t know where the ceiling of LLM performance will be, but as the building blocks are fundamentally different to those of biological computers, it seems unlikely that the limits will be in any way linked to those of the human brain. In much the same way the best hydraulic ram has completely different qualities to a human arm. In some dimensions it’s many orders of magnitudes better, but in others it’s much much worse.
I don’t think it’s hard to know. We’re already seeing several signs of being near the plateau in terms of capabilities. Most big breakthrough these days seems to be in areas where we haven’t spent the effort in training and model engineering. Like recent improvements in video generation. So of course we could get improvements in areas where we haven’t tried to use ML yet.
For text generation, it seems like the fast progress was mainly due to feeding the models exponentially more data and exponentially more compute power. But we know that the growth in data is over. The growth in compute has a shifted from a steep curve (just buy more chips) to a slow curve (have to make exponentially more factories if we want exponentially more chips)
Im sure we will have big improvements in efficiency. Im sure nearly everyone will use good LLMs to support them in their work, and they may even be able to do all they need to do on-device. But that doesn’t make the models significantly smarter.
The wonderful thing about a sigmoid is that, just as it seems like it's going exponential, it goes back to linear. So I'd guess we're not going to see 1000x from here - I could be wrong, but I think the low hanging fruit has been picked. I would be surprised in 10 years if AI were 100x better than it is now (per watt, maybe, since energy devoted to computing is essentially the limiting factor)
The thing about the latter 1/3rd of a sigmoid curve is, you're still making good progress, it's just not easy any more. The returns have begun to diminish, and I do think you could argue that's already happening for LLMs.
Progress so far has been half and half technique and brute force. Overall technique has now settled for a few years, so that's mostly in the tweaking phase. Brute force doesn't scale by itself and semiconductors have been running into a wall for the last few years. Those (plus stagnating outcomes) seem decent reasons to suspect the plateau is neigh.
with autonomous vehicles, the narrative of imperceptibly slow incremental change about chasing 9's is still the zeitgeist despite an actual 10x improvement in homicidality compared to humans already existing.
There is a lag in how humans are reacting to AI which is probably a reflexive aspect of human nature. There are so many strategies being employed to minimize progress in a technology which 3 years ago did not exist and now represents a frontier of countless individual disciplines.
This is my favorite thing to point out from the day we started talking about autonomous vehicles on tech sites.
If you took a Tesla or a Waymo and dropped into into a tier 2 city in India, it will stop moving.
Driving data is cultural data, not data about pure physics.
You will never get to full self driving, even with more processing power, because the underlying assumptions are incorrect. Doing more of the same thing, will not achieve the stated goal of full self driving.
You would need to have something like networked driving, or government supported networks of driving information, to deal with the cultural factor.
Same with GenAI - the tooling factor will not magically solve the people, process, power and economic factors.
> You would need to have something like networked driving, or government supported networks of driving information, to deal with the cultural factor.
Or actual intelligence. That observes its surroundings and learns what's going on. That can solve generic problems. Which is the definition of intelligence. One of the obvious proofs that what everybody is calling "AI" is fundamentally not intelligent, so it's a blatant misnomer.
One of my favorite things to question about autonomous driving is the goalposts. What do you mean the “stated goal of full self driving”, which is unachievable? Any vehicle, anywhere in the world, in any conditions? That seems an absurd goal that ignores the very real value in having vehicles that do not require drivers and are safer than humans but are limited to certain regions.
Absolutely driving is cultural (all things people do are cultural) but given 10’s of millions of miles driven by Waymo, clearly it has managed the cultural factor in the places they have been deployed. Modern autonomous driving is about how people drive far more than the rules of the road, even on the highly regulated streets of western countries. Absolutely the constraints of driving in Chennai are different, but what is fundamentally different? What leads to an impossible leap in processing power to operate there?
Why couldn’t an autonomous vehicle adapt to different cultures? American driving culture has specific qualities and elements to learn, same with India or any other country.
Do you really think Waymos in SF operate solely on physics? There are volumes of data on driver behavior, when to pass, change lanes, react to aggressive drivers, etc.
All these folks are once again seeing the first 1/4 of a sigmoid curve and extrapolating to infinity.
No doubt from me that it’s a sigmoid, but how high is the plateau? That’s also hard to know from early in the process, but it would be surprising if there’s not a fair bit of progress left to go.
Human brains seem like an existence proof for what’s possible, but it would be surprising if humans also represent the farthest physical limits of what’s technologically possible without the constraints of biology (hip size, energy budget etc).
Biological muscles are proof that you can make incredibly small and forceful actuators. But the state of robotics is nowhere near them, because the fundamental construction of every robotic actuator is completely different.
We’ve been building actuators for 100s of years and we still haven’t got anything comparable to a muscle. And even if you build a better hydraulic ram or brushless motor driven linear actuator you will still never achieve the same kind of behaviour, because the technologies are fundamentally different.
I don’t know where the ceiling of LLM performance will be, but as the building blocks are fundamentally different to those of biological computers, it seems unlikely that the limits will be in any way linked to those of the human brain. In much the same way the best hydraulic ram has completely different qualities to a human arm. In some dimensions it’s many orders of magnitudes better, but in others it’s much much worse.
1 reply →
I don’t think it’s hard to know. We’re already seeing several signs of being near the plateau in terms of capabilities. Most big breakthrough these days seems to be in areas where we haven’t spent the effort in training and model engineering. Like recent improvements in video generation. So of course we could get improvements in areas where we haven’t tried to use ML yet.
For text generation, it seems like the fast progress was mainly due to feeding the models exponentially more data and exponentially more compute power. But we know that the growth in data is over. The growth in compute has a shifted from a steep curve (just buy more chips) to a slow curve (have to make exponentially more factories if we want exponentially more chips)
Im sure we will have big improvements in efficiency. Im sure nearly everyone will use good LLMs to support them in their work, and they may even be able to do all they need to do on-device. But that doesn’t make the models significantly smarter.
The wonderful thing about a sigmoid is that, just as it seems like it's going exponential, it goes back to linear. So I'd guess we're not going to see 1000x from here - I could be wrong, but I think the low hanging fruit has been picked. I would be surprised in 10 years if AI were 100x better than it is now (per watt, maybe, since energy devoted to computing is essentially the limiting factor)
The thing about the latter 1/3rd of a sigmoid curve is, you're still making good progress, it's just not easy any more. The returns have begun to diminish, and I do think you could argue that's already happening for LLMs.
Progress so far has been half and half technique and brute force. Overall technique has now settled for a few years, so that's mostly in the tweaking phase. Brute force doesn't scale by itself and semiconductors have been running into a wall for the last few years. Those (plus stagnating outcomes) seem decent reasons to suspect the plateau is neigh.
Human brains are easy to do, just run evolution for neural networks.
with autonomous vehicles, the narrative of imperceptibly slow incremental change about chasing 9's is still the zeitgeist despite an actual 10x improvement in homicidality compared to humans already existing.
There is a lag in how humans are reacting to AI which is probably a reflexive aspect of human nature. There are so many strategies being employed to minimize progress in a technology which 3 years ago did not exist and now represents a frontier of countless individual disciplines.
This is my favorite thing to point out from the day we started talking about autonomous vehicles on tech sites.
If you took a Tesla or a Waymo and dropped into into a tier 2 city in India, it will stop moving.
Driving data is cultural data, not data about pure physics.
You will never get to full self driving, even with more processing power, because the underlying assumptions are incorrect. Doing more of the same thing, will not achieve the stated goal of full self driving.
You would need to have something like networked driving, or government supported networks of driving information, to deal with the cultural factor.
Same with GenAI - the tooling factor will not magically solve the people, process, power and economic factors.
> You would need to have something like networked driving, or government supported networks of driving information, to deal with the cultural factor.
Or actual intelligence. That observes its surroundings and learns what's going on. That can solve generic problems. Which is the definition of intelligence. One of the obvious proofs that what everybody is calling "AI" is fundamentally not intelligent, so it's a blatant misnomer.
One of my favorite things to question about autonomous driving is the goalposts. What do you mean the “stated goal of full self driving”, which is unachievable? Any vehicle, anywhere in the world, in any conditions? That seems an absurd goal that ignores the very real value in having vehicles that do not require drivers and are safer than humans but are limited to certain regions.
Absolutely driving is cultural (all things people do are cultural) but given 10’s of millions of miles driven by Waymo, clearly it has managed the cultural factor in the places they have been deployed. Modern autonomous driving is about how people drive far more than the rules of the road, even on the highly regulated streets of western countries. Absolutely the constraints of driving in Chennai are different, but what is fundamentally different? What leads to an impossible leap in processing power to operate there?
4 replies →
Why couldn’t an autonomous vehicle adapt to different cultures? American driving culture has specific qualities and elements to learn, same with India or any other country.
Do you really think Waymos in SF operate solely on physics? There are volumes of data on driver behavior, when to pass, change lanes, react to aggressive drivers, etc.
4 replies →
"If you took a Tesla or a Waymo and dropped into into a tier 2 city in India, it will stop moving."
Lol. If you dropped the average westerner into Chennai, they would either: a) stop moving b) kill someone
> a technology which 3 years ago did not exist
Decades of machine learning research would like to have a word.