Remarks on AI to Support Climate Mitigation & Adaptation by Professor Ronald L. Rivest for a panel (one of three) Held at CSAIL, MIT on January 21, 2022 Panelists are Una-May O'Reilly, Alan Edelman, and Ronald L. Rivest I would like to begin by thanking Daniela Rus, Elsa Olivetti, and the organizers for pulling this event together, and for inviting me to participate. Climate change is an immensely important topic, and it is good to see us getting together to share ideas on how best to contribute to its solution. I'd like to make four points. The first two are more negative in character, the second two more positive. * (1) (Core) Computer Science and Climate Science are like oil and water. They don't mix well. Computer Science (at least, "core computer science") is about an abstraction -- a world where everything is represented by 0s and 1s. Climate Science, on the other hand, is about the Real World, made up of atoms (carbon atoms, oxygen atoms, etc.) Computer Science would be very much the same in a universe with different physical laws, or an a planet where computers were made of zirconium and powered by methane. If you can represent 0s and 1s, and if you can represent computation on those bits, you have Computer Science. By contrast, Climate Science depends crucially on the physical laws and the physical properties of atoms and molecules in this world, on this planet. Thus: (1) Computer Science and Climate Science are like oil and water. So, what can we Computer Scientists do to help with Climate Change? * (2) Computer Scientists can try to make computation more efficient, but that may backfire. We have, over the past several decades, seen enormous efficiency gains in both the hardware for computation, and the software (that is, the algorithms) for computation. Some problems, such as solving large linear-programming instances, have seen efficiency improvements of over six orders of magnitude over several decades, with approximately three orders of magnitude coming from hardware improvements, and three orders of magnitude coming from algorithmic improvements. [Bixby, 2004] One might think that such efficiency improvements could help with climate change. Yet we would be wise not to forget the "*Jevons Paradox*". In 1865, economist William Stanley Jevons, when studying coal use in Britain, said, "It is a confusion of ideas to suppose that the economical use of fuel is equivalent to decreased consumption. The very contrary is the truth." Jevon's Paradox suggests that as computation becomes more energy-efficient, we will use *more* of it, not less, negating the supposed benefits improved efficiency and resulting in greater energy use. At the moment, data centers use about 3 percent of the global electricity supply and are responsible for about 2 percent of total greenhouse gas emissions. Will this go up, or down? I don't know. The human brain is a "proof" that enormous computations can be done with little energy use. [ https://www.datacenterknowledge.com/energy/study-data-centers-responsible-1-percent-all-electricity-consumed-worldwide ] Even worse, there are computations, such as the notorious "proof of work" at the heart of cryptocurrencies like Bitcoin, that are *designed* to be resistant to efficiency improvements. We need to replace these with more efficient "proof of stake" approaches. Russia, China, and the EU are in the process of banning such "bitcoin mining"; we should do the same here in the US. Thus: (2) Computer Scientists can try to make computation more efficient, but that may backfire. But there are nonetheless ways we can help! * (3) Computer Scientists can help climate scientists do better research, by helping to build better climate models, and facilitating communication between climate scientists. Professor Alan Edelman can speak to the role the programming language Julia (which he helped design) can play in building large efficient modular climate models. (Alan and I taught a seminar on climate change, with Professor John Fernandez, in Fall 2019. I'd like to thank again the many guest speakers we had from this community.) So, first, AI and machine learning can help improve the approximations inherent in any such model, allowing for more accurate and longer-term predictions. Second, Facebook, arXiv, and Google Groups help climatologists communicate. Large, cheap storage helps organize the large datasets needed for machine learning. Nonetheless, I think one should not overestimate the benefits of improved modeling, compared to those fundamental scientific, engineering, and economic innovations that help reduce emissions. And the impact of facilitating communication between climate scientists may be as important as improved modeling. Still, (3) Computer Scientists can help climate scientists do better research, by helping to build better climate models, and facilitating communication between climate scientists. Finally: * (4) Computer scientists can help innovate in the area of "mechanism design", providing innovative ways to incentivize emissions reduction and sustainable behavior. As an example, let me strongly recommend the new novel by science-fiction writer Kim Stanley Robinson: *Ministry for the Future*. [E.g. https://ew.com/books/the-ministry-for-the-future-climate-change/ ] In it, Robinson envisions a "carbon coin" -- “a digital currency, disbursed on proof of carbon sequestration to provide carrot as well as stick, thus enticing loose global capital into virtuous actions on carbon burn reduction.” Read the novel for more details! This is close to, but not quite the same as, "carbon pricing", which would be mandatory. See the recent paper by Abrell et al. for a discussion as to how machine learning can help evaluate the impact of carbon pricing. Of course, computer scientists have no monopoly on mechanism design; this is realm of economists as well. Thus: (4) Computer scientists can help innovate in the area of "mechanism design", providing innovative ways to incentivize emissions reduction and sustainable behavior. I propose such "mechanism design" as a "challenge problem" for this workshop. Challenge Problem: *How do we best incentivize ourselves to take the actions our models* *tell us we should be taking?* This ends my remarks. I'd like to again thank the organizers for putting this panel together.