A physics-based ocean AI might be difficult to develop, and even if we managed it, it might not know where to start. We might largely avoid the GO problem of pruning heuristics if we break the mCDR problem down into many steps, following an initial hypothesis. The hypothesis might be that growing more phytoplankton should improve mCDR. Then, What most limits such growth?, growth is limited mainly by nutrient deficiencies. Which nutrients? : FePSi ones as diazotrophs can made nitrogenous ones. Where?: near the ocean surface. How?: use buoyancy. At what release rate?: ultra slow. et cetera until we end up with my Buoyant Flakes method to trial and optimise.
Alpha go was in the middle of a game so of course it’s not going to reveal it strategy, but wouldn’t it have been possible to have it cary out the rest of the game so we can understand why it made that move. That’s assuming we don’t care about beating it and we just wanna know why it made the move.
So if a AI comes up with something drastically new on climate change, why wouldn’t it be able to explain to us how it would work?
I think you are right- if you see the absolute bonkers responses to people who think “Bill Gates is CoNtRoLiNg tHe WeAtHeR” because he gave a grant to a guy who wanted to put a couple of kgs of sulfur into the stratosphere, I’m not sure “we just do what the AI box tells us to do” really has much legs.
It’s fundamentally a challenge of shots on goal- we only have one earth to run experiments on, so things will go pretty slowly unless you are just modeling the system, but then you are just running Monte Carlo simulations which is kind of standard fare.
Thank you for this welcome respite from my chronic AI doomerism. (Of course, the doomer voice in my head chimes in, “wouldn’t it be a delicious irony if AI solved climate change AND eliminated the human race!)
First, let me say that I admire your work and am very much on board with the idea of carbon removal as an essential approach. It appeared likely, and is now amply demonstrated, that we'll continue to pump CO2 into the atmosphere as long as it is the cheapest way to generate energy. We need a way to get it out cost-effectively.
However, it would surprise me if we were at a stage where optimization is a more important problem than understanding. AlphaZero depends on a formal model of the system being searched: you have to know the rules of Go or Chess, exactly, to evaluate the expected value of different search paths. Is the physics of oceanography really perfectly (or even adequately) modeled?
Some of AZ's successors (e.g. MuZero) do build a model from observation -- but the methodology depends on vast numbers of samples unlikely to be available in real world, physical research.
I had the same thought. If you teach the AI model with simulated data, then you need to be very sure that your models are correct.
Another thought: Why would the best approach in carbon removal necessarily be 'move 37 type'? AlphaGo has made many many moves that met human expectations, before it made one unexpected 'move 37'. This would suggest that at least for Go, the need for type 37 moves is rare and that 'unsurprising' moves provide the best answer most of the time. Maybe it is different for carbon removal, but nothing in the article explains why.
It's a good question. I think the difference people have been studying Go and playing Go and obsessing about Go for thousands of years. But they've been thinking about mCDR for about 10 minutes.
Imagine if AlphaGo had been available to the first two generations of Go players. That's where we're going to be for AI for mCDR.
But then, I'm just working out my views on this stuff...
This seemed worth noting:
https://oceanvisions.org/phytoplankton-report/
New Ocean Visions Report: Can Phytoplankton Help Close the Carbon Dioxide Removal Gap?
Couldn't agree more. 'Move 37' changing the game - totaly mind-blowing!
A physics-based ocean AI might be difficult to develop, and even if we managed it, it might not know where to start. We might largely avoid the GO problem of pruning heuristics if we break the mCDR problem down into many steps, following an initial hypothesis. The hypothesis might be that growing more phytoplankton should improve mCDR. Then, What most limits such growth?, growth is limited mainly by nutrient deficiencies. Which nutrients? : FePSi ones as diazotrophs can made nitrogenous ones. Where?: near the ocean surface. How?: use buoyancy. At what release rate?: ultra slow. et cetera until we end up with my Buoyant Flakes method to trial and optimise.
Alpha go was in the middle of a game so of course it’s not going to reveal it strategy, but wouldn’t it have been possible to have it cary out the rest of the game so we can understand why it made that move. That’s assuming we don’t care about beating it and we just wanna know why it made the move.
So if a AI comes up with something drastically new on climate change, why wouldn’t it be able to explain to us how it would work?
Definitely.
I think you are right- if you see the absolute bonkers responses to people who think “Bill Gates is CoNtRoLiNg tHe WeAtHeR” because he gave a grant to a guy who wanted to put a couple of kgs of sulfur into the stratosphere, I’m not sure “we just do what the AI box tells us to do” really has much legs.
It’s fundamentally a challenge of shots on goal- we only have one earth to run experiments on, so things will go pretty slowly unless you are just modeling the system, but then you are just running Monte Carlo simulations which is kind of standard fare.
Thank you for this welcome respite from my chronic AI doomerism. (Of course, the doomer voice in my head chimes in, “wouldn’t it be a delicious irony if AI solved climate change AND eliminated the human race!)
First, let me say that I admire your work and am very much on board with the idea of carbon removal as an essential approach. It appeared likely, and is now amply demonstrated, that we'll continue to pump CO2 into the atmosphere as long as it is the cheapest way to generate energy. We need a way to get it out cost-effectively.
However, it would surprise me if we were at a stage where optimization is a more important problem than understanding. AlphaZero depends on a formal model of the system being searched: you have to know the rules of Go or Chess, exactly, to evaluate the expected value of different search paths. Is the physics of oceanography really perfectly (or even adequately) modeled?
Some of AZ's successors (e.g. MuZero) do build a model from observation -- but the methodology depends on vast numbers of samples unlikely to be available in real world, physical research.
As it happens, I'm working on a piece that addresses most of this — out later this week.
I had the same thought. If you teach the AI model with simulated data, then you need to be very sure that your models are correct.
Another thought: Why would the best approach in carbon removal necessarily be 'move 37 type'? AlphaGo has made many many moves that met human expectations, before it made one unexpected 'move 37'. This would suggest that at least for Go, the need for type 37 moves is rare and that 'unsurprising' moves provide the best answer most of the time. Maybe it is different for carbon removal, but nothing in the article explains why.
It's a good question. I think the difference people have been studying Go and playing Go and obsessing about Go for thousands of years. But they've been thinking about mCDR for about 10 minutes.
Imagine if AlphaGo had been available to the first two generations of Go players. That's where we're going to be for AI for mCDR.
But then, I'm just working out my views on this stuff...