KI und Nachhaltigkeit – Doc J Snyder & Marcus Voss // #SFTF

Okay let's let's do this.

So maybe everybody takes the headset otherwise it will be too silent I fear I guess.

First of all one question before we start this session.

Is there anybody who doesn't understand fluently German?

Like showing her hand or his hand.

Okay so there's one person so we will do this in in English and sorry before because you know I am speaking rather globish but I hope you will understand and the reading will be in English and afterwards the talk everything okay.

So welcome to the session I'm really happy that you're here and I'm really happy to be here in this festival because I really like that festivals nowadays are getting more hedonistic and more like mixed with laughter and music and culture.

I'm Dr. Julia Schneider aka Dr. J. Schneider.

I'm a comic essayist and former labor market and innovation researcher and this is my partner in crime for this comic Marcus Voss and he and his AI firm Birds on Mars supported with their knowledge on AI and sustainability this comic.

Important also is the participation of Pauline Kremer our famous and really talented illustrator who now is in Jerusalem right now and I want to say even if we talk here right now about AI and sustainability they are really important analog topics in this region for example and our hearts go out there to all people who suffer there okay just this is not everything that happens in the world of course but now we will talk on the rather digital topics AI and on the rather real topic sustainability and in the form of a comic essay yeah and let's start.

Hello our planet has finite resources artificial intelligence is a tool that could both help and hinder us from living more sustainably.

AI refers to the development of algorithms and computers capable of handling complex tasks that require human-like intelligence.

AI has many applications such as prediction optimization and classification but AI is not limited to flashy things like text or image generation it's not only JGPT it can also be used in research or engineering.

Sustainability is a loaded concept and has many definitions there are three perspectives social economic and environmental.

For reasons of space this comic focuses on the environmental perspective although all three are related.

One concept of sustainability the planetary boundaries model defines a safe operating range for the human species for us.

Today many of those boundaries have already been crossed.

We as a collective have been struggling to live sustainably for centuries meanwhile AI is here to stay and advancing faster than ever.

AI can outperform us in some tasks saving time and resources and creating new opportunities for sustainability but also risks.

Let's look at some examples.

Where we have problems because we are overwhelmed bored or tired AI can automatically learn and analyze connections between lots of data and specific results enabling us to make better analysis and predictions in particular in collaboration with classic methods from science and statistics.

On a global scale AI is helping to preserve forests oceans the atmosphere and other ecosystems by identifying damage and facilitating preventive action.

For example through remote sensing that analyzes masses of data coming from satellites drones sensors and cameras.

To detect and measure damage already done and support decisions for quick and accurate responses.

With AI we can more accurately simulate really complex phenomena like no it's not trivial weather and climate or optimize crops and fertilizer use.

This is especially valuable in light of climate challenges such as intensifying droughts.

On an industrial level AI can optimize the use of energy resources and materials while minimizing waste and production processes.

It can help us reduce emissions and support effective waste management practices or improves supply chain visibility to track sustainability efforts.

And AI's potential for optimization helps ensure the sustainability of other complex systems such as transportation.

Here AI can increase the use of renewable energy improve traffic flow and provide more sustainable transportation options.

And predicting damage before it happens can decrease costs making transportation cheaper too.

AI's ability to analyze data and predict trends also makes it a valuable tool for in urban planning.

Using sensor satellite drone and weather data AI can efficiently manage trees and vegetation optimizing water and resource usage for better maintenance.

In our buildings AI can facilitate energy savings by managing heating systems and optimizing energy usage.

On a personal level AI could steer us towards sustainable online shopping or travel by making eco-friendly recommendations.

AI is also at work in faster development and testing of new ideas materials and technologies.

Whether it's recommendation or translation algorithms or indeed generating new ideas through language or image models.

Or sifting through countless research articles and then hypothesizing.

Developing hydrogen technology for example becomes easier this way as research at large.

Based on the information I found in an incredible myriad of research articles I proposed the following hypothesis.

And let's not underestimate another potential benefit of AI.

The ability to very accurately simulate in advance what could become dangerous in real life.

In other words AI has shown potential to contribute to a sustainable future.

Yeah nice.

So everybody sit back play bubble shooter and let the AI nerds do whatever they do.

Well surprise no.

AI is not just our cloud-based friend.

Before we can kick back we need to solve some serious AI problems.

First training running storing and cooling AI systems and the infrastructure still uses up made massive amounts of energy resources that are desperately needed like drinking water.

Other resources are mined under terrible conditions like rare earths.

AI produces significant amounts of co2 plus e-waste.

Currently the fastest growing waste stream in the world.

Also environmental technology technological advances like AI can backfire.

More efficient, more convenient, more consuming, more polluting than before.

And for the sake of space we will only briefly mention the dangers of overly complex and intransparent models, false or biased data, hallucinations, surveillance, control and manipulation by AI.

The elimination of certain job tasks without providing alternatives for those who perform them or technical and economic dependence on large AI players.

Or simply that we might lose our key skills and knowledge by blindly trusting the accuracy of data and algorithmic results.

AI isn't.

So what now?

The bottom line is that we need to find a way to live more sustainably and soon.

Fortunately our planet offers multiple intelligences in the form of problem-solving skills to meet this challenge.

Trees, jellyfish, our diverse human intelligences.

Let's combine them with artificial intelligence in ways that help our planet stay within its boundaries and even regenerate.

We've only just begun to see AI at scale.

Whether you believe in AI or not, now is the perfect moment to get involved in sustainability.

Debates about opportunities and risks are just now moving out of AI labs into policy, civil society and companies.

Let's strive for AI that matters, in a good way.

To be continued.

For the choreography, Markus will now head on to some use cases and on some deeper insights into the topic of AI and sustainability and show you the website and stuff.

And after that we have time for exchange.

Because for me you don't need to be fans of AI or so.

Just, you know, it matters.

The exchange matters.

So yeah.

Okay.

So I don't know if you've wandered around here.

You've already probably also seen the comic up there.

So we have this here.

It's hanging over there, I think.

So it's not a very long comic, right?

So it's a lot of information that needed to be compressed into a very short form.

So this is actually how the project started.

That when Julia and I started to talk, that I sent her like 10, 20 pages, I don't know, 15 pages of, this is like, links and ideas of things that could be included.

But it was like 8 point, a real 8 point, I would say, 15 pages.

With references, studies and all those things.

And because I was working on this area of AI and sustainability, AI and climate for several years now, as a researcher at TU Berlin.

Then I was engaged in a NGO that's called Climate Change AI, where I was also a member of the board for a while.

And I was organizing summer schools and workshops in the area of climate change and AI.

And so I sent all that to Julia and Julia said, okay, this is a lot.

And then did something which I found quite incredibly, because I can't really do this, like put this into this very, very short form.

And then I was like, kind of like, oh, but couldn't I put just footnotes everywhere?

And she was like, no, this doesn't work in a comic.

So I don't want to put footnotes everywhere.

And so what we did now is a compromise that we said, okay, we have this webpage, there you can actually download this comic also freely.

So it's a non-commercial license, open source, so you can use it for all non-commercial purposes.

If you want to use it for commercial purposes, you can of course also contact us.

And we said, okay, we put some sources there.

So you can actually click there also and find some more of those readings that I collected and actually read a little bit more.

And in this talk, I actually want to show you some examples that have inspired parts of this comic essay.

Actually, it's also in German there now, thanks to TU Munich, where we did a similar event like this.

We could use the funding to also translate it and have Pauline create it also in German form.

So the pigeon says hallo now as well, not just hello.

So if you start, for example, at this very first panel, which starts rather broadly, you could say that some of this has been inspired by a very big research paper that was actually... so I already mentioned Climate Change AI and some of the founding members of Climate Change AI, like David, Priya and Lynn, now all professors by now, back then I think they were all PhD students, were writing like this very big report tackling climate change with machine learning.

I think officially published it was like in just recently, but initially it was written in I think 2019, 2020.

And it's a long paper, has like a hundred of pages, hundreds of resources in there, and it's a really great overview paper if you really want from like more academic perspective get into this topic.

So it starts by giving an overview of all the areas where climate, where AI can be used in the area of climate.

So there's of course the areas you'd expect, so you can look at electricity systems, building, land-use industry, transportation, so all these areas where CO2 emissions are happening.

And within each of these you can then zoom into like a lot of example applications that have been shown.

And I mean we've by now done these workshops, where now hundreds of researchers have submitted every year to those workshops more and more hundreds of more typical applications.

Also an adaptation, and what I found something that I didn't have in mind so much at the beginning when when I started jumping into this topic is also in climate science itself.

So AI and machine learning is used a lot now by climate researchers as we would see also in a one or two examples.

I mean this is a very of course very broad and going into all of this is a lot, but you kind of see like also in those papers there's some pattern emerging and one of those is for example what happens all the time is something that we also included in the comic like the remote sensing like analyzing tons of satellite data for example something that nowadays is almost not possible without some computer vision.

Because I mean those satellites are flying all over earth like there's now thousands of satellites collecting a lot of data and I mean no human could look at all of this and for example see how the earth is changing like land use for example over time couldn't track all what's happening in the energy systems couldn't for example detect methane lakes leaks and all the different count ships that are going over the ocean and all these different applications that that I've been shown in research.

I actually want to give one one concrete example I mean it's not climate but weather so so but you see kind of similar things also in the climate area so for me one of the favorite AI models in the last year and and one very interesting trick of research is that now machine learning models are on a similar capabilities to those classic weather models.

I don't know if you what you know about weather forecasting but but in weather forecasting what's done right now is that every three hours every six hours like huge supercomputers are calculating and simulating the physics of the earth and calculating this and this has some limitations because it's very compute intensive and you only can do this in certain resolutions for example in cells of 10 kilometers by 10 kilometers for the whole earth and I mean the physics is known you have all the equations but you just don't have enough time to compute this every day because if it would take a day to calculate then the weather is already gone for the next day and this is where machine learning can now come in and machine learning like like neural networks can learn basically the physics just by looking at a lot of data of the past and can now make also predictions for the next days and this can help to make calculations for example in higher resolution timely also locally and this is now not just something that here for example DeepMind says the model is capable of but there's now actually also the European Center for medium-term weather forecasts which is really writing a lot about this like readly and also offering a lot of like opportunities to benchmark models with their big models that they have for Europe for example and show that I mean sometimes for certain parameters those traditional models are still a little bit better but for some more parameters now machine learning models are better but now it's a really interesting time in this area and this is something which which I find quite interesting and follow a little bit and by now yeah it's just a matter of time that these models are really I am they are almost used in production right now I mean that you can get those weather forecasts also if you want to have them but I think they still their main models are still the traditional ones but this is really changing right now so this is one concrete example in this case from the ECMF and DeepMind but I also wanted to include some of the examples that I've personally been involved with so I work for a small AI consultancy birds on Mars and maybe some projects in the area of sustainability that we recently started as for example with the WWF and here you see the kickoff of the project in Stralsund and that's about mapping seagrass with computer vision so there the idea is to use really cheap hardware so instead of big sonar we are using camera systems there the idea is to put those two Fisher boats and then use computer vision to analyze and map the seagrass and I think there have been talks already yesterday about or some startups have pitched about seagrass also what you can do with it if you know where it is so there's different applications for that and really interesting project is also with the Museum für Naturkunde in Berlin they have like a large I don't know if you've been there but I mean there's a lot of things that you see but there's so much more things which you actually don't see which are hidden in their closets and everywhere because I mean they've been collecting things that animals all over the earth for centuries now almost I mean at least decades but since the 1800s and each of these special specimen they have like like labels and they're not digitized right now and some of them are handwritten some of them are written with the typewriter some of them are written by computer and so this is an interesting computer vision task also now they are collecting all these labels and we are helping them to see how well you can actually digitize that because I mean this is not something like that you're out of the box OCR model can do because like like I said I mean some of the handwritten of the 1800s for example are really challenging in this area and another project is with B&D where we support them using it's also OCR but a little bit simpler with like if you're in your drugstore for example and you want to take images of your creams and to detect what chemicals are in there so that they have an app that's called ToxFox they can show you okay what's the ingredients in there right now they do this with a barcode scanner but it's really difficult to keep this on track because companies are changing all the times what they're putting in there and so we are now trying to see if we can read it directly from the labels and then they can give you some recommendations just some recent examples maybe for this transportation panel there's also a project that we person mastered with Deutsche Bahn so that inspired this predicting damage before it happened so what we do there is that the Deutsche Bahn has a project where they put cameras basically just iPhones and in the front of the cockpits and collected data for all 30,000 kilometers of tracks twice a year and then they can detect different things and you're kind you would be kind of surprised how much they don't know what things are there and aren't there anymore because I mean young people just go there sometimes and take signs maybe also old people but I assume from my childhood that it's possibly young people stealing stuff I don't know they just didn't measure things they sometimes don't know how long this is and one sustainability related thing that we did is what they did up to two years ago is they put out glyphosate everywhere so they just sprayed and sprayed glyphosate so they didn't have so much vegetation now they didn't want to do this anymore because I mean this is meant to let stuff through and so it of course goes into the environment and so they need to do other measures now and there are certain measures that they can choose from there's there's some other chemicals which are not as aggressive that they can use there's hot water they can send out people with torches or apparently there's also some yeah with fire but also some laser of torches or something to get rid of the weeds and each of them are different in cost and and so they need to decide that and to collect information we use computer vision so you can actually see here now that we use machine learning to detect all the vegetation there you could say okay do you really need machine learning I can't you just count green pixels we tried that and it works to some extent but then in summer for example vegetation is brown sometimes it's blue like if you if you're in the shades so it doesn't work so well and machine learning actually beat counting pixels quite a bit yeah and so this is a something that has been in use now at Deutsche Bahn and is now extended to other for example detecting damages on train tracks another project that inspired this panel up here is a project that we did with the city of Berlin and Technologie Stiftung Berlin this is called quantified trees it was funded by the Ministry of Ministerium and there the idea is to use a machine learning model to integrate several data around the trees so Berlin has a very good open data about the trees so they know for each so there's around 800,000 trees that they have data on 400,000 the street trees 400,000 are some in the parks and they know what type of tree it is how old it is what's the size and they need to water those trees because in the changing climate some of these trees are not especially the older ones are not adopting very well and so they water the young trees of course but they now also sometimes need to water older trees and for that we use the data we have of the trees some of the trees in Berlin Mitte and Neukölln have sensors that we had access to and we could of course have access to some of the watering data that's something in the project it's Excel table so it's not a very good interface right now so that's a little bit sad this is also why the project is currently not in use because this connection was never really implemented we but what we did also we calculated the shade for every tree in Berlin so we know how shady for example this tree is more in the shade than this tree and this is also very relevant information as we found when talking to the expert and then yeah of course the weather data and then we put this into a simple machine learning model and there of course you don't need some complex chat GPT something something you can use something that's called random forests which I kind of like also because it's trees making predictions about trees so it's for computer scientists I think a quite nice joke yeah and then you can make forecasts for for the next weeks for example or you can just also predict for all the trees where you don't have sensors what's the most likely sensor value based on on all the other data that you have so that's a pretty cool project and we provided this in the dashboard for the greenery department so they could actually look at the data and use this when they optimize their routes we also have this for for people basically to provide more transparency it's called Baumblick which like I said we would like to launch more openly but the project is over and we are now looking for a little bit of funding by the city to actually launch this and and and actually yeah have people actually use this app for four more days you can actually vote for this project if you like this project just saying maybe go to bla bla compass if you like to and I want to support this project just just say yeah maybe to kind of go towards wrapping this up oh yeah and of course when when we started the comic I think it was I mean chat GPT was there but this big Jenny I hype was starting ish and so we kind of included this of course also meant back then for example I mean one was like simulating different alternative futures and for me one of the very first things I did with when that Lee had this in painting feature was like take a picture of my street in Berlin Schöneberg and and see okay for example how would it look like if I flooded it back then it looked okay ish I think if you did this with mid journey or something now it would be much much much cooler but I kind of like that because I mean it was just a little bit of prompting to have an alternative version of my street and I think this is actually pretty good like up to some years ago there wasn't similar tool that was much more difficult to develop burning the trees was kind of difficult then on the right I also used it for some more positive future so like my street is usually full parked with full streets and cars and cars are driving there so I kind of asked it to imagine like a more yeah walkable city neighborhood and I kind of like this to just prototype a new my neighborhood and paint a positive image and I mean I should have made this a product I guess because now there's Dutch cycling lifecycle.com and you can actually go there to every area in Google yeah like where we have these Street View images and make it more Dutch so for example I don't know if you if you're from here if you recognize the discussions about Berlin Friedrichstraße so this is where the local government opened and closed opened and closed open and closed Friedrichstraße for cars and it was always complained about how ugly it is that they put just some green plants there but if they instead did something like this and made it a little bit more Dutch with some generative AI it would have maybe inspired people a little bit more I think I mean it looks a little bit too Dutch almost but I mean with a little bit of prompting you could you could possibly also do this differently.

Another chance which which is also something that we see with our enterprise customers now a lot is yeah the access to knowledge I think this is a very interesting chance of generative AI of course there have been a lot of problems at the beginning with the hallucinations and so on but I mean some of this you can really work against now and for example applications like this which is called climateqa.com is an app that works quite nicely to give you access to all the knowledge in the IPCC reports right so you can actually I don't know the IPCC reports are the big reports that are the climate scientists in the world write about and it's thousands of pages thousands of pages and even the management summaries have hundreds of pages like I don't know if you looked at this it's really a lot of to read and this gives you access to that so and actually it makes it in a way that I mean of course hallucinations still happen every once in a while but it shows you the sources shows you where it's got it from so it provides some more trust and shows you some more yes that this actually helps you answer your questions that you may have and if you think about generative AI in this area it doesn't always have to be fancy with images and text something that I've done some work some years ago with a student was working in energy systems and their generative AI can also be used to generate other things in this case it was time series and I was working on energy forecasting so predicting realistic curves for for smart meter data for example and this could be used in simulations as a forecast and we use something that's called normalizing flows and what you see there's now like instead of images or text like realistic okay this could be people are cooking this is people watching TV so realistic looking time series that can be used for simulations so like like you also said in the comic AI is not just such a pity and not just Ali and you go back for a second I want to because they used Bernstein polio normal normalizing flows just saying yeah so yeah but but of course when we talk about these big these big very large models like it's still of course that we need to talk about the footprint of AI so that's also what we included in the comic and yeah this is this panel right so this is something a discussion that has been started like five years ago with a paper by Emma Strube where she first looked at those language models which have been actually much smaller back then and but it's good it's complicated right so it really depends on what you're talking about and this is a little bit also a message that I want to give here and I gave some examples of some simpler models like those random forests for example this is something we have this discussion doesn't matter so much right I mean training a random forest can be done in in seconds or one minute on my laptop here so this is like like giving this talk produces more energy than training this model but if you're talking of those very very large language models then of course the training is one aspect that is relevant but now it's actually more also the inference right I mean talking to chat GPT I think there's numbers of like like it's several liters of water like it for for several like a like a longer conversation that you have with chat GPT so it's if you're just looking up for information that also Google can give you then Google is like I think 20 times more efficient than just getting the same information out of chat GPT for example but I mean this is very intransparent still and very difficult so here's for example from a paper from Lynn Kark that one of the CCI found us that try to give put this into a framework so you have in the middle or these compute related impacts that are from training from for making predictions and then of course it's but but then it's more important what you do with it right so if you do it for example use those models to then save energy in buildings and transport and industry then it can of course be net positive in a way but then it gets more complicated even like if you know some economics and then there's things like a rebound effect so just because you can do things more efficiently as you suddenly do things more from one example could be taking a car with the uber right I mean making a route more efficient can save some energy but if it's getting too comfortable then you suddenly use uber small or like like an example from another areas light bulbs right I mean light bulbs have been much more efficient now than several years ago but suddenly they are part of interior design and you just put up a lot more light bulbs everywhere and so the light the amount of energy that light is using electric lighting is kind of similar to several decades ago even though every light bulb is much more efficient and so it's very complicated it's very intransparent I think this is the main message I can give you here to this topic but we can of course discuss this also at lunch a little bit more I mean there's some studies now but for example open AI is very very intransparent about this yeah so let's care about the AI footprint and let's talk about this keep talking about let's watch it let's if you're doing projects let's see yeah let's create some more transparency but let's also yeah shape the AI handprint so what you're actually using it for so I think there are a lot of talks about sustainability and have a startups working in this area like doing something like useful with AI and deciding also maybe to not use AI for certain use cases for example I mean if you're working on detecting the best area to drill for oil for then maybe this is not not a very good handprint then of course there's a lot of areas like I think some of these examples that I showed you where the model is reasonably small but it's just as small as it needs to be and can be used in several applications that I would say are reasonable and then there's of course this very large area in the middle where I don't know you get your news recommended you get new videos to recommend it to watch where it's kind of a debate we have to have like like is this a use case you want that we want that we as a society one and so I think this is the very interesting part where we should talk about and if you're for example looking at graphcast that I mentioned to you this is actually like I mentioned already right I mean weather forecasting has been done on on supercomputers now for for the past decades I get centuries decades not centuries decades and now you can actually run a state-of-the-art weather forecast for the next days for the whole world on a desktop computer so you can basically run this now on on on a MacBook and that's that's kind of kind of cool I mean I actually wanted to try this always haven't done it so far but yeah I just believe this nature I think I mean it's nature so it should be reasonable yeah so to wrap this up and leave some some time for questions yeah currently everyone has the feeling that yeah I is just chat GPT and subs at least some some people have that but this is also one part of generative AI this is even just one part of AI so there's all those random forests all those other methods that are much simpler much smaller and can dedicate a computer vision models which we at Burson must still use like if you don't need a like a large vision model if you can use a much more simpler model and this is only one part of digitalization and this is even just a smaller part of efficient solution for increasing sustainability so in a lot of areas I mean you see great examples here you don't need AI you don't need digitization you can come up with really cool ideas some good products talk about the zero race and so on and you don't need AI for that so this is also just something to keep in mind yeah and I want to end this year and if you have questions we have the microphone there now here's the microphone thank you for the talk I have a question how do we know the benefits of using AI for sustainability projects or AI in general in our lives will outweighs the drawbacks as you mentioned in your talk like the more amount of energy we are using and all the resources so when do we need to stop or do we need to stop I mean on a global scale it's hard to say right I mean the benefits will not our way outweigh the the footprint because I mean like I already mentioned I mean there's some applications which definitely will harm the climate and the world I mean like oil companies oil and gas is investing in AI okay let's not even talk about that then there's I would say there's a very small amount of projects like if like let's say if it's just 5% or 10% which are doing things like you also see here for example like mapping seagrass starting again increasing efficiency in transport and building so so this will be some amount of projects but the large amount of applications of AI are now in business context in person life right I mean Apple puts it in every hardware device and and those areas will not have a net positive effect like I mean they're not sick I mean they're making your life more comfortable maybe and they're making companies more efficient but this is where those sort of use cases where we have to have the discussion if we need like a fridge that can talk to you like I personally don't but then I'm lazy to type on my phone so I'm happy if Apple puts like a better Siri in there that it can actually understand me and type my messages for me because I hate typing and I also hate text messages but this is my personal choice and this is something that I would just gives me a little bit more comfort and I like to have recommendations of newspaper articles that are relevant to me but then I don't need tick-tock recommendation but a lot of people need tick-tock recommendation so I don't know it's it's really the sort of debate that we need to do but it will not be a net positive definitely but but for certain cases I think it's possible to do so if you do projects of course you need to at the very beginning talk about impact measurement and like like for example quantify trees as an example where at the beginning we didn't talk so much about it actually like of course like saving water would be one of the measures that we could do but it's really hard to do we never get to that in the project at the end so we can't answer those questions how much water can be then they save now like in Berlin and this is definitely something if you're starting a project that you should then do and you can actually track of machine learning models if you're training them yourself there's ways to track the co2 emissions like if you're running the Python code for example there's libraries that you can use and you put in what GPU you have and then you can calculate the footprint but if you're using open AI it's it's not transparent they don't tell you that and maybe to add I think this is a really difficult question that you pose and I don't think that anyone could answer this but I have a more pragmatic view I think AI is not going away not at all never because it's too comfortable for us you know and for rich people also too comfortable so it won't go away and this is why I think okay so what now what's do about it apart from that I'm not a pessimist what to determine concerning technology but I think okay I is here to stay and we are not good with sustainability so please let's combine these things that so that it matters in a good way because it will matter anyhow so any more questions or comments or we can just bring coffee like if you don't don't agree with what we said it's also fine it's fine you're totally cool with that I am so yeah then thanks a lot for being here and come talk to us if you want to or have a drink with us yeah and yeah we're happy that you have been here and we are honored that you listen to us thank