- il y a 3 mois
The Butterfly Effect: Rethinking AI’s Impact on the Planet and Society
Catégorie
🤖
TechnologieTranscription
00:00Hi everyone, welcome. Welcome to my keynote. In case you don't follow me on social media, I have a real
00:08thing for butterflies.
00:09So today I'm going to talk to you about the butterfly effects on AI on the planet and society.
00:15So often you might hear, and actually we're going to talk about this later on in our panel, that AI
00:20is either good or bad for the planet.
00:22And so you hear things like training an AI model emits as much carbon as five cars in their lifetimes.
00:27You have tech companies using AI to get more oil and gas. You have, AI is going to solve the
00:34climate crisis.
00:35And so you have these positive and these negative promises and coverage, but in reality it's a lot more complex
00:41than that.
00:42There are definitely positive effects, and we're going to talk about this right after my keynote in the panel.
00:49And so AI has a big role to play in discovering new materials for solar panels and batteries.
00:54It can help us understand our world better. For example, there's a startup that uses AI to listen to the
01:01rainforest and detect illegal logging and deforestation using AI.
01:06And this is something that can be done on a large scale to an extent that human beings can't patrol
01:10that much territory.
01:12You've got satellite imagery that is collected 24-7 and that can be used to detect methane leaks to monitor
01:19coral reefs.
01:20So there is true potential in AI and the fight against climate change.
01:25But it comes at a cost, and this is something that I've been working on for the recent years.
01:29You've got the water cost. Data centers use a lot of water for cooling.
01:32You've got power demand. Energy is becoming harder and harder to get to the data centers that need huge amounts
01:40of it.
01:41You have tech companies who are essentially losing track of their own net zero commitments because of AI.
01:47And it's estimated that the cloud, so everything we use for Internet, AI, Netflix, etc., has a greater carbon footprint
01:54than the airline industry.
01:57There's also really interesting geopolitical impacts.
02:00If you think about it, you have now tech companies that are privatizing nuclear power plants,
02:05which essentially brings the tech mentality of move fast and break things,
02:09and the nuclear mentality of move slowly so you don't blow things up.
02:12You have countries like Taiwan who have to choose between giving water to farmers for their crops
02:19or giving water to chip makers to make hardware for training AI models and computers.
02:25You've got models like DeepSeq that have profoundly overturned the stock market overnight
02:30because everyone was freaking out about the fact that you could actually train AI for less money.
02:35So you have these really tangible impacts of AI on geopolitics.
02:39But what I've come to understand in the recent years is that this is not the whole story.
02:45And this is why I like to think of the butterfly effect of AI.
02:48So butterfly effects are really small changes that have huge impacts that are hard to predict, right?
02:56Like they say that the flap of a butterfly's wings can cause a hurricane on the other side of the
03:00world.
03:01And I think this is a great metaphor for AI because the small AI systems that we use that are,
03:06you know,
03:06on our phones can actually have very, very profound impacts.
03:10So we need to think beyond just the environment and start thinking about AI's impacts on society.
03:16So let's start with material effects.
03:19I don't know how often you think about how AI impacts our material world, the world that we live in.
03:25So for example, AI is replacing objects.
03:28Who owns a map anymore?
03:30Who buys dictionaries?
03:31Who even buys recipe books nowadays because, you know, you can use AI to generate recipes?
03:36And that's, it can seem sustainable because the objects don't exist anymore in the real world.
03:41But on the other hand, we have to be building more and more data centers.
03:45And we're also outsourcing our human knowledge to the AI machine, right?
03:50Because we don't have to think about things anymore.
03:52We can just use ChatGPT to generate a recipe for us.
03:55So there are profound impacts on also our brains and our societies and the objects that we interact with in
04:03our daily lives.
04:04There's also the effects of scale.
04:06And it's really interesting because we are used to large-scale production being cheaper and less intensive.
04:12So for example, IKEA is cheaper furniture because they make so much of it that you can buy a bed
04:18for 200 euros and not 500 euros.
04:20And because there's such a huge scale.
04:22And actually, AI has largely followed this scaling laws and bigger is better AI models.
04:29But the impacts are actually growing.
04:31Contrary to IKEA, the impacts of AI are actually growing.
04:35So data center energy use is growing.
04:38Now data centers use more energy than countries, countries like Argentina and Egypt.
04:45So, you know, we use more and more AI.
04:47The models are getting bigger and bigger, but the cost isn't going down.
04:50So, you know, logically speaking, there's something wrong with this approach.
04:55There's also economic effects.
04:57There's something called the Jevons paradox.
05:00That's a really interesting concept in economics.
05:02That says that when you get better and more efficient at using a resource.
05:06So, for example, coal.
05:08At the end of the 19th century, we started getting really, really good at using coal.
05:14And Jevons, who was an economist, actually observed that instead of using less coal, we were using more coal.
05:21Because now we could do more things with it.
05:23And an example of this is cars, right?
05:25Our cars have gotten so fuel efficient, we can go further and further for the same amount of fuel.
05:30But yet we're driving further and further because nowadays we can spend 100 euros and go, I don't know, 300
05:36kilometers or 400 kilometers away.
05:38And we can travel more and we don't need to use trains.
05:41For example, in North America, it's a real problem that people just drive everywhere because it's cheap.
05:46And so, essentially, there's this hypothesis that we're seeing a Jevons paradox in AI.
05:52And actually, the CEO of Microsoft tweeted about this earlier this year saying that, well, actually, AI is becoming more
05:58efficient.
05:59Everyone's using it.
06:00It's going to become the next, like, oil, essentially, right?
06:03The next coal.
06:04But in parallel to this, the cost of training a single AI model of the state of the art, the
06:10ones that are used by everyone, is growing a couple of years ago.
06:15So when I started out in AI research, you could train a model on your laptop.
06:19A student in a university could train a model and deploy it and it could be used.
06:25And nowadays, models cost tens of millions of dollars to train.
06:28So, actually, despite the fact that we're getting more efficient at using AI, we're still using more AI and the
06:35costs are still rising.
06:39And also, what's interesting about this is that it also precludes people from entering the field of AI.
06:45Because, as I said, when I was a PhD student, I could just code something up on my laptop and
06:49submitted it to a top AI conference.
06:51But now, what grad student has 10 million dollars in order to train an AI model?
06:56Very few grad students have that money.
06:58And so, researchers are more and more dependent upon companies to give them access to computes, to give them funding.
07:05And this creates, of course, a conflict of interest.
07:09There's more and more corporate influence of big tech companies in AI research.
07:14And it really makes it difficult for new people outside of the existing community to join.
07:20Because if you're a grad student, how are you going to compete with, you know, people from Microsoft or Facebook
07:25or Google unless you have this money?
07:27And so, we're really having this concentration of power.
07:31In terms of indirect economic effects, it's really interesting because we're all using AI every day.
07:37We're more and more dependent on AI.
07:39We're relying on AI to navigate, to make our grocery lists, to even, you know, play songs for our children.
07:46But we're also buying more and more gadgets for this.
07:49If you think about it, nowadays, sure, you can have all sorts of cool stuff on your phone.
07:53But you also have a smartwatch, you have a smart speaker.
07:56Some people now have smart microwaves and smart ovens because apparently that's a thing.
08:00And actually, the e-waste problem of AI is rising.
08:03And it's electronic waste or e-waste is becoming the most, the biggest growing, fastest growing type of waste globally.
08:12Because people are updating their gadgets so often.
08:15And because we're so bad at recycling electronic gadgets.
08:18It's because it costs so much money in order to get all the precious metals that are in your phone
08:23once you throw it away.
08:24So, this costs so much money.
08:25And this is really adding up as you want to upgrade your phone in order to have the latest Apple
08:30intelligence,
08:30to have the latest, I don't know, chat GPT access.
08:33You want to have a faster phone.
08:34And so, you throw away your phones more and more rapidly.
08:37And so, this is really becoming a huge issue that we don't see because these millions of tons of waste
08:43are far away from us.
08:46And the third type of effect that I'd like to invite you to think about are societal effects.
08:51And I think that this is something that, once again, we don't think about enough.
08:54Because AI is something that's so ingrained in our societies.
08:58And it's so immaterial and invisible that we have a hard time understanding how it impacts our societies at large.
09:05And so, for example, AI innovations are changing how we use time.
09:10And so, Google did a study a couple of years ago saying people who use Google Maps save time.
09:15They spend less time in traffic.
09:17They spend less time in their cars.
09:19And so, this is a net benefit, right?
09:21We're gaining time.
09:23Time is magically appearing.
09:24There was another study that was done a couple of years later that looked at how we spend this newfound
09:30time.
09:31And they actually found that, yes, automation, so robots that help you around your house or, you know, smart speakers,
09:38et cetera, can save us time in the direct task that we're trying to do.
09:41But actually, the environmental impacts are higher because we're going to buy more things.
09:46We're going to travel more often because now we have more free time.
09:50And so, it's interesting because if you only look at time as if it's the only thing you're measuring, you
09:55can think that, yes, AI is saving us time.
09:57We spend less time in traffic, et cetera, et cetera, but the environmental impacts and the societal impacts can actually
10:02grow completely disproportionately to this time.
10:06And something that people ask me a lot is can we compare the environmental impacts of people and AI?
10:13How do we understand whether AI, as opposed to human beings, is good or bad for the planet?
10:18And there was a study done last year where they did exactly this.
10:22They looked at human writers and human graphic designers and they took their yearly carbon footprint in the United States
10:30and in India and they divided it by roughly the amount of time that it takes to write a page
10:36of text or design an image.
10:39And then they compared it to AI models.
10:41And they concluded that AI writing produces 130 to 1500 tons less CO2 per page than a human author.
10:47And so, of course, the conclusion to this would be we should just replace all human authors with AI models
10:52and it's going to be better for the climate.
10:54And this study was re-shared and tweeted by Yann LeCun.
10:58It became a thing on Twitter.
10:59Everyone was like, yes, Sasha, your work is terrible because we should just be replacing all human beings with AI
11:05models for text and for image.
11:07But then if you read the study in detail, it says that the emissions analyses that they did do not
11:12account for social impacts like professional displacement, legality, and rebound effects.
11:17Because, of course, you can't just focus on the time that it took a human being to write a page
11:23and compare it to an AI model that generated a page of content because that AI model was trained on
11:28the output of the human writer in the first place.
11:30And also, you can't replace human beings with machines in general because there's more to human beings than just the
11:37work we do.
11:38So these kinds of analyses are completely incomplete and flawed because they don't look at rebound effects.
11:45They don't look at life cycle analysis.
11:46They just focus on how much time does it take a human being to write a page of content and
11:50what's our carbon footprint.
11:51And so nowadays, when people ask me this question, can we compare the environmental impacts of people in AI, I
11:57just say no, because you can't compare humans and machines.
12:01You can't compare human intelligence and machine intelligence.
12:04They're two different things because human beings are more than just the tasks that we do.
12:10So, as a conclusion, and hopefully we'll continue this conversation as part of the panel later on, the climate crisis
12:17is real.
12:18I think we all agree with this despite temporary setbacks on a geopolitical level.
12:23The climate crisis is real, and we have to be honest and proactive about how AI can stay part of
12:29the problem or become part of the solution.
12:32And I think that this is really a reflection we should be having and thinking above and beyond positive and
12:36negative impacts, but we should be thinking about things like rebound effects, things like how do we compare, how do
12:41we make decisions, and how do we make it work as a society.
12:46And I think we should be looking beyond environmental impacts.
12:50We should be looking at social and economic impacts because at the end of the day, any one of the
12:57three categories that I showed you is only one of the categories, but there's more to it than that.
13:01If we're looking at sustainability, we have to be thinking of economics.
13:04If you're looking at economics, you have to be looking at societal impacts.
13:07If you're saying AI is going to replace writers or accountants or human designers or whatever, you have to look
13:14at what are the repercussions on society, what are the repercussions on labor markets.
13:19You can't just say, oh, well, you know, human writers don't need to exist anymore, let's move on.
13:24You have to go deeper than that.
13:27And I think that as a community, as people who use AI, this is not limited to only the people
13:31who make AI, but also the people who use AI, we can reflect and reimagine the relationship between AI technologies,
13:39business objectives, because, yes, business objectives, and ecological imperatives.
13:43And usually when we talk about sustainability, it's limited to environmental sustainability, but in the 1980s, when sustainable development became
13:53a field, there was a proposition that sustainable development has to take into account the environment, the economy, and society,
14:00without which it's not truly sustainable, because we need these three pillars, otherwise we're only looking at part of the
14:07picture.
14:08So, thank you very much, and I'll be back in a second for the panel.
Commentaires