Are artificial intelligence (AI) & machine learning (ML) helping or hurting the climate fight? With MIT's Priya Donti, head of the nonprofit Climate Change AI, I discuss some of their current applications, their systemic effects, and how policy can channel them toward pro-social uses.
1) Very surprised to your guest say "AI won't find a pattern unless one is there." THat's because we humans are absolutely brilliant at finding patterns that are not "there"
And since we already know that LM and AI systems are terrific at picking up on and replicating suboptimal thinking and behaviorial traits we are not conscious of (or like to admit) -- like racism keeps coming up bigly in systems trained on internet data -- it would be an astonishment of the first order if LM/AI systems did not replicate our preternatural ability to "see" patterns that are not "there."
(Scare quotes because, by definition, if a pattern is "seen" it is there, even though it may dissolve immediately upon different thought . . . ).
2) Stray thought: Cryptocurrency (which I tend to call "Griftocurrency" since its main use seems to be separating rubes from their money) is a colossal problem for the environment. Those of us concerned with climate need to start pushing now for a new tax regime that strongly discourages the kind of computing masturbation used to "mine" bitcoin and similar follies (I would set the tax at 75 - 100% for all existing cryptocurrencies that were created by having them solve worthless and increasingly more laborious computational exercises) and then require that all crytocurrencies created after the new tax regime is passed be "mined" by working on actually useful problems of the sort you discussed with your guest or else they too are subject to the tax.
That is, we basically create a federal project where the National Academies of Science and National Institute of Health and NOAA and all our other starved basic science research public research agencies get together and create a huge atlas of problems that need massive computational power (protein folding, chemical toxicity calculations for interactions, grid optimization, climate modeling, etc. etc.) and they figure out a way to weight each project so that anyone who wants to "mine" crypto currency through computers sitting around burning through calculations is actually working on a useful problem, and the resulting cryptocoin gets born with a tax stamp that exempts it from the stiff tax applied to all existing cryptocurrency created before this new tax regime.
Basically, if you want to pollute and destabilize the climate by having a computer sit around and spend its days turning useful energy into waste heat and cryptocoins, expect to have the government tax the living shit out of that. But if you want to apply all that power to solving the world's real problems, the coin you make gets a steep discount on the tax or exempted entirely, so long as the energy spent went to solving a real problem.
Just found a great passage in a book called “19 ways of looking at consciousness“ by PATRICK house. This PhD neuroscientist rights “the brain abhors a story vacuum and because the mammalian brain is a pattern-recognizing monster, a briny sac full of trillions of coincidence detectors that are only useful if there are connections between things. Even a wrong pattern, a guess, is at least a pattern to learn against.”
The discussion on the impact of AI on climate change was good to bring up. Like you, I'm skeptical this is going to be a big issue. As your guest says, right now this is only 2% of emissions. While computing is going up for sure, 2 aspects mitigate I think -
a) as companies move from data centers to the cloud (i.e. someone else's data center) there is more concentration of computing power, which can then be addressed by both optimization (which your guest talks about) as well as by converting those centers to 100% renewable power which you allude to.
b) the other side of the equation needs to be considered too - if the marginal benefit of adding a unit of AI capacity results in a positive carbon equation (say 2-1 or more Kg of carbon reduced per Kg increased due to AI) I would think we would do this trade any day (at least at this point, and for several years, by which time we hopefully will have a much cleaner grid anyway).
This situation is quite different from crypto - where in some schemes at least the stake any player has in the system is directly related to the amount of "work" put in (i.e. energy expenditure). In AI, there is no incentive to spend on power - it's only a by-product of the current computational hardware.
The point about embodied carbon in hardware is well-made, though - hadn't thought about that. I also particularly liked the framework Priya Donti laid out, with 4 thematic uses of AI.
1) Very surprised to your guest say "AI won't find a pattern unless one is there." THat's because we humans are absolutely brilliant at finding patterns that are not "there"
https://en.wikipedia.org/wiki/Apophenia
And since we already know that LM and AI systems are terrific at picking up on and replicating suboptimal thinking and behaviorial traits we are not conscious of (or like to admit) -- like racism keeps coming up bigly in systems trained on internet data -- it would be an astonishment of the first order if LM/AI systems did not replicate our preternatural ability to "see" patterns that are not "there."
(Scare quotes because, by definition, if a pattern is "seen" it is there, even though it may dissolve immediately upon different thought . . . ).
2) Stray thought: Cryptocurrency (which I tend to call "Griftocurrency" since its main use seems to be separating rubes from their money) is a colossal problem for the environment. Those of us concerned with climate need to start pushing now for a new tax regime that strongly discourages the kind of computing masturbation used to "mine" bitcoin and similar follies (I would set the tax at 75 - 100% for all existing cryptocurrencies that were created by having them solve worthless and increasingly more laborious computational exercises) and then require that all crytocurrencies created after the new tax regime is passed be "mined" by working on actually useful problems of the sort you discussed with your guest or else they too are subject to the tax.
That is, we basically create a federal project where the National Academies of Science and National Institute of Health and NOAA and all our other starved basic science research public research agencies get together and create a huge atlas of problems that need massive computational power (protein folding, chemical toxicity calculations for interactions, grid optimization, climate modeling, etc. etc.) and they figure out a way to weight each project so that anyone who wants to "mine" crypto currency through computers sitting around burning through calculations is actually working on a useful problem, and the resulting cryptocoin gets born with a tax stamp that exempts it from the stiff tax applied to all existing cryptocurrency created before this new tax regime.
Basically, if you want to pollute and destabilize the climate by having a computer sit around and spend its days turning useful energy into waste heat and cryptocoins, expect to have the government tax the living shit out of that. But if you want to apply all that power to solving the world's real problems, the coin you make gets a steep discount on the tax or exempted entirely, so long as the energy spent went to solving a real problem.
Thoughts?
Just found a great passage in a book called “19 ways of looking at consciousness“ by PATRICK house. This PhD neuroscientist rights “the brain abhors a story vacuum and because the mammalian brain is a pattern-recognizing monster, a briny sac full of trillions of coincidence detectors that are only useful if there are connections between things. Even a wrong pattern, a guess, is at least a pattern to learn against.”
I generally prefer reading transcripts to listening. When will the Priya Donti interview transcript be available?
Buzz Blackett
Shouldn't be more than 2-3 days.
Really interesting podcast David.
The discussion on the impact of AI on climate change was good to bring up. Like you, I'm skeptical this is going to be a big issue. As your guest says, right now this is only 2% of emissions. While computing is going up for sure, 2 aspects mitigate I think -
a) as companies move from data centers to the cloud (i.e. someone else's data center) there is more concentration of computing power, which can then be addressed by both optimization (which your guest talks about) as well as by converting those centers to 100% renewable power which you allude to.
b) the other side of the equation needs to be considered too - if the marginal benefit of adding a unit of AI capacity results in a positive carbon equation (say 2-1 or more Kg of carbon reduced per Kg increased due to AI) I would think we would do this trade any day (at least at this point, and for several years, by which time we hopefully will have a much cleaner grid anyway).
This situation is quite different from crypto - where in some schemes at least the stake any player has in the system is directly related to the amount of "work" put in (i.e. energy expenditure). In AI, there is no incentive to spend on power - it's only a by-product of the current computational hardware.
The point about embodied carbon in hardware is well-made, though - hadn't thought about that. I also particularly liked the framework Priya Donti laid out, with 4 thematic uses of AI.