Daniel Erasmus on ClimateGPT, AI for climate decisions, social intelligence solutions, and surfacing hidden connections (AC Ep51)

“The promises are tremendous and the peril is climate, not AI. “

– Daniel Erasmus

What you will learn

• Discussing the real existential threat of climate change
• Exploring AI’s role in addressing climate challenges
• Daniel Erasmus’s background in foresight and scenario planning
• The development and impact of ClimateGPT
• The importance of Human-AI collaboration
• Equitable access to AI technologies for climate solutions
• Innovative climate resilience strategies and examples

Episode Resources

• Artificial intelligence (AI)
• The Promise and the Perils of AI
• Rotterdam climate initiative
• BloombergGPT
• ChatGPT
• World Economic Forum
• ClimateGPT
• Singapore Sea Lion
• European Central Bank
• Systemic Risk Board
• FSB (Financial Stability Board)
• NOAA (National Oceanic and Atmospheric Administration)
• Sea Ban (European legislation for a carbon border tax)
• SDGs (Sustainable Development Goals)
• TCFD (Task Force on Climate-related Financial Disclosures)
• TNFD (Taskforce on Nature-related Financial Disclosures)
• chess.com (freestyle chess competition)

Transcript

Ross Dawson: Daniel, it is awesome to have you on the show.

Daniel Erasmus: It’s great to see you again, mate. It’s been far too long. And it was wonderful seeing you in San Francisco last year. I mean, it was a fascinating event, the title of the event was The “Promise and the Perils of AI”. In the audience, we had Rusty, they’re working on meteorites and a whole set of sort of existential issues facing humanity. And, and the point that I made there is, that people tend to place AI as an existential threat within these, and instead of sort of challenges for human supremacy, but the real threat, the peril is somewhere else. And the peril is not AI, it’s climate change. Climate change is a structural threat that will face humanity, at the scale of the UN estimates 200 million climate refugees by 2050, maybe half a billion, that’s 26 years from now, half a billion a decade later. 

Now, the European project barely survived one and a half million Syrian refugees. So the kind of things that we are talking about here, we’re going to have to get really, really good at not just anticipating what’s happening but acting on that early preparing for that with the least amount of human and of course, planetary suffering. 

And so that’s the promise of AI. And I think it’s far more interesting to look at AI with those terms. How can it help us? And how can we, together with AI come to very, very different solutions than we have in the past for the real existential threat, which is climate change?

Ross: Yep, absolutely. The challenges we face are unprecedented in complexity and scale. So we hopefully have some tools which can assist us in that. But I think that goes a little bit to your background and where we’ve crossed over in the past is understanding complex systems. So it’d be great just to hear a little bit about your background and how you’ve come to this point from your work in foresight over the years

Daniel: I’m South African, my origin, and I witnessed the transition of South Africa from an oppressive racist regime to a democracy, which was perhaps one of the most exciting things to happen in my youthful life, but it was a youthful life. But within that one gets the bones of looking ahead, scenarios transformation, and that the same people in the room, looking at the thing very differently, can come with very, very different conclusions. And then spent almost an hour actually, but over a quarter of a century, running scenarios and foresight processes, largely for multinational companies. So the Fortune 50 type of things, countries, cities, and doing a set of transformation projects around this, of which, there’s certainly some clerk climate work that came out of that Rotterdam climate initiative to half CO2 levels from their 9090 level, which was launched before. Al Gore’s film even came out in 2005, anticipating the global financial crisis for a bank, which led to them having their most profitable year in 150-year history in 2008. Running the first central bank digital currencies for central banks, anticipating the oil price collapse for an oilfield, so several multibillion-dollar exercises for clients, but at one point, one takes a step back and says these are legacy and foresight, which talks to the practice of foresight and bringing people together. But the challenges that we find look ahead and take foresight seriously, the challenges are going to require far more than changing single organizations. It’s going to require changing entire ecosystems of organizations’ industrial ecosystems, the way that cities are coupled with each other, the way the country’s bill is coupled with each other, and the policy issues around that. And, so the focus then became, how does one build the tooling to facilitate that type of decision-making on this? And I will say it’s not an AI solution that we built by the social intelligence solution.

Ross: So, what have you built when we spoke last year, you said what are you doing, focusing entirely on AI and climate change. So what does that look like? What have you done? What are you and your colleagues doing? 

Daniel: We’re in a very fortunate situation but over 15 years ago, we started collecting large amounts of money and we built a professional search engine that complemented the scenario practice. So one could get a better assessment of that. And it looks a lot like some of the diagrams behind you at this point. And one can get an interactive view. But we were in a very fortunate position because we were doing web-scale crawls and processing, just over a quarter of a billion URLs every day.

We had the kind of collections with which one could build tasks and specific domain-specific models. And, then decided to build something akin to chatGPT is about, you know, half the size of tokens. So similar, say to Bloomberg GPT, has about the same amount of tokens. So it’s a task-specific model. And we built a family of models, so a foundational model. So that’s trained completely from scratch and fine-tuned models. 7 billion 70 billion. So there’s a family of models, 13 billion parameter models, and it was a unique and exciting challenge, you know, it required us to look for green power — one of the things we put a, you know, stake in the ground and said we wanted to build a model on green power. And experts far smarter than ourselves said, ‘No, this is not possible in this day and age.’ And now it’s there. And it’s the first to the best of my knowledge, an AI paper with a foundational model that indicates its sustainability, balanced scorecard, and says exactly how much CO2 went in it. And even today, we run it with green inference. So the whole model runs on green inference. And these things are eminently possible to be done.

The model, again, benchmarks there was, you know, deciding lack of benchmarks, so we had to add to the benchmarks so that one can test the model on its performance on general reasoning tasks, so we had about seven of those. And then climate-specific benchmarks which we were very happy to partner with Exeter that had a collection of climate myths and disinformation. So when contested because they looked upstream and actually found that missing disinformation came from four places. So four very discreet institutes create a lot of climate myths and disinformation, which makes it rather easy to track. And we wanted the model to help. And not hindering doesn’t mean models don’t hallucinate, they don’t make mistakes, etc, etc, I think we’re all pretty much aware of the limitations of models and that it should be the best effort possible. It was a consortium of equity Labs, which focused on responsible AI and model provenance. So you can see where the models are an uptake, which is the focus development house with deep skills in translation and NLP. So the three of us together, and we were supported by a grant from the UAE sovereign fund so that we could launch at coop, and then model one point X was launched at the World Economic Forum earlier this year. 

Ross: Fantastic. So well, perhaps, let’s come back to some of the model foundations, but how is this used? So we’ve got climateGPT? So what is the value of having this model? How is this used, who will use it, and what will be the outcomes of it existing? 

Daniel: So, technologies, other than a very, very few exceptions tend to diverge going forward. So the car became the motorbike, the SUV, the sports car, the city car, the little smart electric cars, so they tend to diverge for specific use cases, as they are. There are very, very few exceptions, where that doesn’t happen. And we expect the same to be happening in terms of large language models that we’re going to get a lot more domain-specific models. Some things like someone like Bloomberg GPT, or even company models, suddenly starting to get country models, the Singapore Sea Lion has an incredible effort of addressing Singapore and its environs languages and having that represented. And so we wanted to carve out something for climate because this is the most interesting and exciting task of our generation.

The scope of the problem is we’re going to have to reallocate about $100 trillion of assets, which is more or less equal to all of the stock. It depends on stock or bonds. But for all of the property on Earth, it’s more or less, we’re going to have to reallocate that, which also indicates, this is a spend of one to $5 trillion. And every time we do these numbers, this is from nature. Every time we do these numbers, we end up with a very different calculation. So, earlier action matters, and better action matters here. And so how could we make the decision-making threshold as low as possible for climate information, and really build a feedback loop there that enables this? So you can go to climategpt.ai Just apply for access, and normally within 24 hours, when is provisioned for that? It’s tied to this as model 1.3. Now, we’re well on our way to model 2. And so which we launched a couple of weeks ago, 1.3. So the next steps in this, and it gives a multi-perspective answer. 

So what we see people using it for is everything from ‘I’m the climate policy, I’m in charge of climate policy for Madagascar, and I work to prepare our island for this, or ‘we’re 120 climate scientists in the Caribbean. And we want to look at strategies for climate resilience for the Caribbean, or from the head of NOAA, too. We’re looking at climate resilience within Turkey with an Argentina within this, we’re Levi Strauss company, and we’d like to develop a net-zero plan. So, it encompasses the broad south swath of public policy changes that are to be done and help support them. And I’ll touch base on how that is done.

Organizational changes, which is a set of regulatory challenges, but also decision-making support that are really needed, and then city level down to individual level. And at this point, what we see from the utility is it kind of encompasses the gambit of those things. 

Ross: So well, well, you mentioned a lot of policy applications there. You also mentioned 100 trillion in reallocation of assets. So there are some, obviously some institutional investment or other essentially those who are holders of capital who are making some decisions there as is it other ways, how might they use these tools as well?

Daniel: The European Central Bank, as well as the Systemic Risk Board, as well as the FSB, various institutions, most central banks on Earth, have said, you know, we have a massive carbon asset bubble. And so that carbon assets are structurally priced in a way that doesn’t measure the long-term liabilities, nor whether that carbon will be taken out of the ground. So it’s a structural, structural, I mean, financial system, encompassing structural risks that are sort of at the scale of the GFC, those kinds of risks, all the way down to two other sets of risks. 

So, from an institutional perspective, most organizations are now doing ESG reports. There are about 3000 organizations doing ESG reports. And we just make it one to make it a little bit easier and a little bit more disciplined in making those reports. So as an organization, you can upload your documents, it can reference your documents, it can tie that to the newest research that is there. And we’ve tied it to just under 2 million climate facts. And we will tie that to more climate facts. And in the near future, we’ll add a couple of 100 million general knowledge facts to that as well. So, using retrieval augmented generation, it gives the references there, but then it can complement it with your institution’s references as well. And you can start to see whitespaces, the delta between the general discussion say about climate change when your organization’s discussion about that, or the SDGs, or whichever lens one picks on so the first is sort of ESG there’s a set of regulatory things. So TCFD, the Task Force for Climate-related Financial Disclosure, TNFD nature-related disclosure, as well as the Sea Ban, which is the European legislation for a carbon border tax, you know, that’s 500 million consumers that starts next year, and companies need to start to answer for their carbon budget and so on report on that. 

So what we’re doing is just facilitating that process, and just making it a little bit easier, and a little bit more systematic, using these AI tools, because it’s a tremendous amount of work done. That’s simply the regulatory side. In the models, we’ve had clients who say, ‘write a climate change scenario for a 1.5 degree as for Sub-Saharan Africa, giving specific emphasis on unseen risks and cascade events of how these things can interconnect from you know, 2025 to 2045 and draw specific systemic conclusions for that and you can say, given that, what should a bank do about this, what should a bank prepare right a plan of action..’

So the normal way when it interacts in the model, um, seems to perform rather well at those tasks. And one can cascade that from different regions and different issues and different scenarios on that. So basically decision support and regulatory requirements are coming in to facilitate that. Of course, linking parties to each other you might not be aware of, but those people are also working on a similar problem, perhaps you should be talking to each other and learning from each other. 

Ross: So I understand that the model is being provided for free?

Daniel: Yes, yes. I’m here discussing with various partners to continue to enable this, we’ve paid for it. So access is free. We don’t think particularly given the deep political nature of the climate, that somebody in Ghana, you know, that’s contributed almost nothing to global warming, is left with a great deal of liability. So to charge that person $10 a month, or that team or to charge that simply just begs reason. So we will continue to fund public access, but we do fund a smaller model. And we will grow that to larger models, as we get more funds available, but access is free, you simply need to log in. And we’ve got these three layers to that. So the one layer is free public access to the model. Two is to work together with sets of universities, of which we’re in discussion with most of the Ivy League and an example process there is with Georgetown, they measured the model in a taxonomy of 20 climate issues. And then measured model performance. And in 80% of them it gave correct, factual answers, but also sufficiently explained in certain detail, et cetera, et cetera. And I can come back to hydrogen is kind of an interesting example there.

It performed best in sort of the geopolitics and the ethical issues around geoengineering and albedo management. So 95 and above, and the worst in climate change, impact in Rwanda. So there’s clearly where our work is cut out to be done. But that’s the second tranche working together with universities, and profiling university research within that, but also creating climate mentors and partners to that. And the last one is all institutions, of which they’ll pay a fee for that. And hopefully, that can fund the general public access to this.

Ross: Which takes me to this point around the humans in the loop. So since it’s not 100%, and I suppose no model ever will be, we still need the humans in the loop, which we always do anyway. So what does this look like as in human experts or human policymakers, assisted by these tools? So, what are some of the ways this can work? Is it to ask the question, and then get it verified by an expert? Or how is the best support for human experts and policymakers? 

Daniel: So one has a few layers to that and some of those things we’ve done and some of those things we’re heading into. So I’m happy to talk about this. I’m an old-school engineer. Although the background here is in foresight, I prefer to talk about work we’ve done rather than you know, vaporware, so I tend to favor the prior rather than the latter. But back in 1997, I argued, when Kasparov made the argument that he was defending the dignity of humanity. I was writing a column at the time, and I thought it was a very limited view. And I’ve lost many, many chess games against computers, and I’m certain I will continue to do so. And I hope my dignity is more or less intact here. 

And so the more interesting thing is, how do we, together with computers, play chess, and the early lessons came from a competition which was held by chess.com. It’s called freestyle chess, in which they had various chess players play, either with machines or teams of chess players or some versions thereof. And they had Grandmaster-level machines play and then Grandmaster-level chess players playing but the competition was won by Zack and Steve. And Zack and Steve, are they prodigies? No, not really. Do they have a mother or something that works and has a supercomputer in the basement? No, no, not really. It was three PCs Dell and two Pentium machines and Zack Steve and three PCs could win against Grandmaster level. Chess computers or chess players, and had figured out this augmentation relationship better. So the human brain does about one move a second. And that’s what makes your podcast so fascinating. If you’re talking about this, I think the most interesting issue of our day in age is how to create this augmentation piece between humans and computers.

And this is an early indication that despite the sort of press of the day about the fantastic aspects of AI, and how to replace it, I think that’s, again, a very limited vision and I hate to criticize Kasparov. But it is a limited vision. The more interesting vision is how we do this together.

Ross: So Kasparov did realize that soon after. 

Daniel: Yeah, he did. He absolutely did any profiled freestyle test. And I think without his championship there, you know, what had been theoretical and that really bordered on an angle bar arguments earlier on, I’m in no way unique for mentioning that, there’s a long history of making these arguments there. But they certainly carried some aspects of consciousness at the time. 

And so the first thing is basic feedback systems. So climateGPT, the public interface, you can select something and you can immediately give feedback and within the next round of that, of the model, and of course, we update all the time. So that’s there. Secondly, as you provide tooling for people so they can turn retrieval or intergeneration. So rag with climate references on or that can tie to the general press. But certainly, for scenario tasks, where you want the models to create something that doesn’t yet exist, you don’t want to tie it to a rag, you actually want to have it be a bit more creative. And, we’re getting better at this. So you provide to people, you know, not all of the buttons, but some of the buttons so that there’s some level of utility because we can develop some level of understanding of how to use these systems, rather than simple black boxes and keeping it by keeping it to a limited set of critical choices on that. 

Lastly is there’s behind the, there’s a number of actions on this, but there’s sort of behind the scenes. There are model competitions, which one runs with model arenas where we test various versions of the models against each other. You test larger models, and you incorporate some of the insights from larger models into the smaller models. And then when complimenting that constantly, we added that one of the training sets of this model was breakthroughs, Global game-changing breakthroughs, and extreme weather through the last decade. So 10 years of extreme weather events at the scale of the planet, there’s a public open interface where we can look at this. And again, one can give feedback on any of those things, this is millions of articles, that one can give feedback on any of those things, I’m misassigned on that.

Ross: So I want to take it into the foresight space now. So this is an extraordinary time. Last year was the hottest year on record every month since then, and has been the hottest month on record. So we now have more tools to be able to think as you just talked about resilience is an important part of being able to adjust. So there are many, many factors here. One is ways of mitigating carbon emissions and others, potentials for geoengineering or another sort of climb, you know, planetary level impacts, and other it’s at national level in terms of resilient, others institutional level organizational level. 

So how do you see the response, assisted by AI now…What is the role of AI and being able to address these incredibly complex, incredibly difficult, incredibly impactful challenges? 

Daniel: So let me answer that in two parts. The first part is looking at the extreme weather data set. There’s a lot of work being done on metrology. There’s Earth 2, there are various things that posit extreme weather, and they model that in the global climate system, but it kind of negates the human activity system. So for many people, it doesn’t matter that much, that it rains a little bit more in the Sahara, which is very important from a methodological point of view, but not from a human activity system. So by using news articles, as well as scientific articles, as well as the body of research around that, one can get better insight into climate cascade events. And one can get better insight into what leads to what in a very unexpected way. 

So I’ll give an example. Last year was not reported, barely reported in early November, there was significant flooding in Kenya, and in East Africa. And so you’ll see in the news articles from this, the flooding happens, significantly. And then very close to that the Kenyan government is looking for 15 million mosquito nets, which in hindsight seems really obvious. And everything and foresight seem really obvious in hindsight, but planning for that is a very different thing. And so planning for mosquito nets in the wake of flooding, etc, etc, at the scale of the planet, and being able to move those resources in ways that mitigate the unclean water, etc, etc. 

So the first thing that we noticed is that there start to be the beginnings of utility. Have some predictions in those articles, and the human activities level, obviously, you now have articles at the level of that city. So you can zoom in, and you can see that city and those places, so the loss and damage discussion can go down to that city, that newspaper in Somalia that talked about that, which may not be around in 10 years time, but is a cryptographically signed event. And we expect the loss and damage discussion to be around a trillion dollars. So how do we know where those losses actually occurred? So building a collection like that matters. So the first thing is, you’re building standard collections. We’ve had the first climate litigators start to approach us and say, Can they have access to this, the donor community saying they can target it very differently because they can now target at this level of a city, not necessarily the sovereign, which enables much more targeted donations. So from prediction from just knowing what happened, so archiving, and then lastly, from anticipation, those collections itself, and AI underpins those things to build those collections. So we tend to think of AI as chunks now, but it’s not, it’s underpinning those delivering the lift planetary scale collections. 

The second level in terms of the chat is a given example. So somebody asked which livestock are best suited for climate change in the Sahel. And it’s not the first thing on my mind, but it is an important question, somebody that was very large and then comes an answer, saying camels and goats are far more suited than cattle. Because they’re more resilient against drought, they eat a broader variety of plants, and they’re capable of dealing with a broader variety, and some other utilities, in terms of the sort of tertiary products, the hair, etc, etc. And so, there’s a set of reference articles on measuring climate resilience in the Sahel. And one can go through hundreds of those types of questions. So it helps reveal hidden connections, which would otherwise not have been seen, it’s those little hidden things of you don’t know. But putting mangroves around ocean Islands has a massive difference in terms of their climate resilience. Now, it’s not going to help against a three, four-meter ocean level rise, but in the meantime, it can have a significant utility, and again, climateGPT comes with an answer saying this is expected by 2050 to have saved $150 billion. And this is the cost-benefit ratio of those various things. So it helps to reveal hidden connections that one would otherwise not have known. 

Ross: Sorry, just want to make a point. I mean, the broader point here is that even if we go to…Well, we’ve already had significant climate change. And even if we got to net-zero this year, which we’re not, then we will continue to have climate change. So we’re moving to a different planet. So I suggest that a lot of the applications that you’re talking about for climateGPT are in terms of understanding that the world is changing in terms of its climate, which includes agriculture in terms of land use, in terms of impact on humans, and we will need to transition from the world we were to the world we will be. And so what you’re really giving the world now is this tool to assist us in managing the trends necessary to transition to a world that is already different and will become more different whatever we are able to do in terms of mitigating carbon emissions. 

Daniel: It is exactly right. And there is a set of decisions here in which better decisions matter and earlier decisions matter and so by 5x. And so, one wants to create an interface that gives multiple perspectives, it gives an economic perspective, it can give a risk perspective, it gives a scientific perspective, a general perspective gives a biological perspective and one can add perspectives. 

So often one can see those different viewpoints on things on that, and they are from a policy perspective, but also from an individual organization. Multiple choices for example, right here, Old Faithful, there’s an Excel file found in an old faithful, that converts nutrient-rich waters to proteins. And so rather than cutting down swaths of forests to grow soybeans on them as done in many parts of the world, one can grow these proteins within this mineral-rich and use that as feedstock for pigs, which a lot of the soybeans in the world are being used, that radically alters our water calculation that radically alters of things that radically alters the requirements for specific system level changes. And one can go through example, after example, example after that, of things that aren’t yet known. Sometimes they’re not at the system-scale, or they’re early stages or they’re proven technologies. South Australia is growing tomatoes in seawater, and doing it exceedingly well supplying millions of tons of tomatoes, growing just seawater and sun. So most people when one mentions this, it seems science fiction, but these things are running at scale. And they’re and so you’re starting to create those and weaving those connections to accelerate change, because one of the primary ways in which one does this and say, Look, over there, they’ve done it, it’s not that risky, we can change and talk to them, learn from them, and then scale that up.

Ross: Fantastic. So the climategpt.ai, I believe. So is there anything else that anyone asks or any pointers for people who want to find out more?

Daniel: The organization behind it is erasmus.ai, you can see some of the extreme weather examples on there. And then obviously linked through to climategpt.ai. If there are parties that would like to support us, either through datasets or through grants, to enable open equitable access to these technologies or want to work together with us on solving particular narrow problems, there’s a whole set of this kind of narrow problems that we’re dealing with on this, or from a university point of view, and then use it and organizations to use it and facilitate because it’ll be both private and public monies going into this changes and businesses, and often are in much more progressive in their viewpoints of understanding the challenges that they have for system wide change. Many, many businesses have signed Net-Zero targets for 2030, 203,5 2040. And simply, you know, when you speak with the CEO, they have no idea how to reach that. And this is the task of our age. And we should be using AI in this regard because the promises are tremendous and the peril is climate, not AI.

Ross: Fantastic. Thank you so much for your insights and all of your work on such an important problem, Daniel.

Daniel: Great seeing you again, mate. And thank you for your work. It’s wonderful, addressing this challenge of how to get people and computers and intelligence to work together in fundamentally different ways and lending some critical thinking to that and profiling work in this regard, I think is a tremendous benefit to us all.