On Silicon Dreams and Carbon Nightmares
If you follow AI news you will have come across discussions around the environmental effects of AI. These conversations typically branch into three themes: energy, material footprint and the polluting impacts, such as carbon emissions. I provide a brief overview of this discussion, but mostly want to point you to a very insightful interview titled ‘Silicon Dreams and Carbon Nightmares: The Wide Boundary Impacts of AI’. It’s a lengthy conversation of almost two hours, but I believe it would be worth your time.
The guest, Daniel Schmachtenberger, is one of the most brilliant thinkers of our time, and you will rarely come across anyone with a similar breadth and depth of knowledge. Daniel’s ability to to contextualize complex issues is extraordinary. There are a two important concepts to be familiar with going in (Jevon’s Paradox and the Multi-Polar Trap) and I introduce below. I also clarify the difference between Artificial Intelligence (AI) and Artificial General Intelligence (AGI), which is beneficial as you listen to the conversation.
This is not an easy topic and you might ask: This already seems somewhat dystopian and if AI is so bad for the planet, why are you promoting it? It’s a fair question. I believe AI could afford humanity with an enormous opportunity, but we must evaluate and apply it with open eyes, understanding what we are doing.
The Three Mainstream Ecology Aspects of AI
Energy and Infrastructure
The energy discussion mostly breaks down into two branches: Energy consumption by data centres, where the chip clusters are housed and infrastructure needs.
a) In terms of pure energy consumption, it is fair to say that it is enormous. As Bloomberg reports at the end of June 2024, altogether data centres in Q1 2024 use more electricity than most countries, with only 16 nations (including the US and China) consuming more.
Nota bene: just for sake of clarity, I want to underline that the demand for more data centres is not exclusively driven by AI, but other data intense industrial and consumer applications as well.
b) On the infrastructure side it is becoming clear, that the energy grid is going to be a major constraint in scaling AI. In many parts of the world electricity demand from data centres is outstripping the available power supply. This leads to concerns of brown-outs and price increases in regions with the highest density of data centres. Goldman Sachs shared an eye-opening interview in this regard in its recent June 2024 Top of Mind Newsletter.
Material Footprint
The second part of the discussion is more familiar and analogous to the debate about the electric vehicles. Here, critics have pointed out that worldwide Lithium reserves are insufficient to produce the batteries we would require to fully electrify transportation.
Similar concerns regarding the material footprint of AI, cover anything from the mining of metals required to manufacture GPUs and other silicon chips, the mining of other materials required to build out the power grid, new power plants and the data centres itself, to the evaporation of fresh water for cooling of data centres.
Carbon Footprint
The third prong of the debate is possibly most familiar and mainly concerns the emissions from operating AI infrastructure. According to the World Economic Forum already in 2022 data centres generated 2.5% of the global CO2 emissions, more than the aviation industry with 2.1% - though this could have been skewed by COVID19 pandemic induced reduction in travel. Goldman Sachs estimates that data centre related CO2 emissions could double by 2030.
Recent reports from Google and Microsoft seem to substantiate these projections: In its Sustainability Report 2024 Google disclosed that despite its goals to become carbon net-zero by 2030, since 2019 its greenhouse gas emissions rose by 48%. It attributed this surge to its data centre energy consumption and supply chain emissions.
Similarly, Microsoft, in its sustainability report released in May 2024, said that its emissions grew by 29% since 2020, which they attribute to the investment “in the infrastructure needed to advance new technologies.”
Like in other areas of society, within this topic increasing attention is paid to the localized costs of AI and the consequential uneven distribution of the impacts, or ‘environmental equity’.
Silicon Dreams and Carbon Nightmares
In his interview Daniel goes quite a bit beyond the above scope and takes a close look at whether AI can lead to the breakthroughs in science and technology that many are predicting. This is not a technical analysis of whether or why AI can or cannot aid in these endeavours, but rather a stocktake of the setup and the motivations driving the development of AI technology. Daniel articulates considerable doubts. The tenets of his argument are rooted in two key concepts: Jevon’s Paradox and the Multi-polar Trap. In my experience, people are often unfamiliar with both, so here is a brief introduction:
Jevons Paradox
Jevons Paradox, named after the British economist William Stanley Jevons, describes the counterintuitive situation that improvements in energy efficiency lead to an overall increase in energy consumption, rather than a decrease. Jevons observed this phenomenon already 1865: as coal-powered engines became more efficient, coal consumption increased because the lower cost of energy spurred economic growth and higher energy use. Essentially, greater efficiency (whether on the consumption side by replacing lightbulbs with LEDs, or on the production side by cheaper oil exploration or solar panels) leads to lower relative costs. This makes certain applications and products suddenly economically viable, that otherwise would not have been profitable to purse. Energy efficiency gains therefore tend to result in a net increase in energy demand (not, as one would expect a decrease) and thus more than offset any energy reduction induced by efficiency gains.
The Multi-Polar Trap
The multi-polar trap refers to a situation in game theory and social dynamics where multiple parties, acting independently and rationally to maximize their own benefits, inadvertently create a collective outcome that is worse for everyone involved. This often occurs in competitive environments where individuals or entities cannot trust others to cooperate, leading to mutually destructive behaviours.
On occasion, this is also referred to as a Moloch Trap, a metaphorical term, drawing inspiration from the ancient Canaanite deity Moloch. Moloch represents the concept of a powerful, insatiable force driving these dynamics, symbolizing the destructive competition and systemic failures that arise from individuals’ rational actions leading to irrational collective outcomes. Moloch exemplifies the power dynamics that lead to competitive pressures forcing individuals or entities to act against their collective best interests, but perpetuating cycles of competition and conflict.
The ‘tragedy of the commons’ intuitively illustrates the multi-polar trap: In a village numerous craftsmen rely on a nearby forest for their wood. Each craftsman seeks to maximize their profit by cutting down trees to produce goods. As more craftsmen harvest the wood the forest becomes increasingly depleted and unable to regenerate. Ultimately, the collective interest in preserving a healthy forest (which offers protection, food, construction materials for housing) is compromised due to each craftsman’s pursuit of personal gain. But the craftsmen are each in a bind, because they cannot reduce their wood consumption, as their competitors would drive them out of the market. The tragedy of the commons is the powerful force that drives the craftsmen to exploit the shared resource, leading to its degradation and ultimately harming the entire community.
Like the tragedy of the commons the rapid development of AI poses a multi-polar trap. The competition between corporations and countries to gain AI dominance is in full swing and pursuing AI superiority drives different actors to prioritize progress over safety, potentially leading to unintended consequences or the misuse of advanced AI systems. As fear of falling behind in the AI race intensifies, the pressure mounts to push boundaries without fully considering the long-term consequences and risks. Overcoming this grid-lock on the other hand would require international cooperation or other collaborative efforts, for which proper institutions and mandates do not currently exist. The leaders in this race have little incentive to ‘slow down’, not only for fear of falling behing, but also driven by the conviction (or claim) that them to win the race is the best outcome, because they are the ‘good guys’. Winners write history, right?
History will be kind to me for I intend to write it.
Winston Churchill
AI vs AGI
At the beginning of the conversation Daniel talks about the race to Artificial General Intelligence (AGI) and the concerns around this, most passionately argued by Eliezer Yudkowsky, an American computer scientist and researcher. Yudkowsky believes that AGI poses an existential risk to humanity, as it could act in ways beyond human control, leading to potentially catastrophic outcomes (the ‘alignment problem’ or the ‘paperclip maximiser’).
Let me be clear: I have great sympathy for the argument and humanity would be well advised to figure this out, before AGI arrives. I admit that sometimes we need to sound an alarm to wake people up, but there are currently no indications that we are anywhere near creating AGI or even only concepts on how we might get there. In this sense nuclear fusion energy is potentially more realistic than AGI at this point, it might even require it.
To make this point properly I need to delineate between AI and AGI: In a rough approximation AI encompasses a broad spectrum of technologies and applications (such as LLMs), whereas AGI specifically refers to the pursuit of creating machines with general cognitive capabilities comparable to those of humans.
Why is this distinction important? While it would be silly to deny that AI has made previously unimaginable strides in the last 18 months, despite any impressions you might have from the news, AGI would require, for example, the machines’ ability to reason and plan. This doesn’t sound like such a big deal, but it actually is!
One of the most prominent AI researchers, Meta’s Yann LeCun, gives a very intuitive example: Let’s say you want to go to Hong Kong from where-ever you are reading this right now. Seems easy, right? Book a flight, go to the airport, and boom, some hours later, here we are. Well, actually, not that easy! Think about it: You need to book a flight, and maybe a machine can handle this bit. But – and let’s leave robotics aside for the moment – what about the next step? You must get to the airport. How do you do that? You must get up and go to the street to catch a cab. How do you get there? You need to go to the hallway and call the lift to go down, etc. Would you rely on ChatGPT, probably the most advanced AI system currently available, to guide you through this?
Well, you say: Hold on a second! Maybe GPT4 is not yet capable of this, but GPT5 or GPT6 surely will be! Very unlikely, because leading AI researchers, who are honest in their assessment (and maybe not driven by the need to sell a product) say that we have absolutely no idea on how to achieve this kind of planning ability. Other opinions, of course, exist; most prominently Geoffrey Hinton, the inventor of backpropagation (the technique that allows neural networks to learn) who resigned from Google in 2023 over safety concerns.
However, if you pay close attention, you may have noticed that Mina Murati, the OpenAI CTO in an interview with her alma mater Dartmouth Engineering from June 2024 shared that GPT5 will not come out until late 2025 or early 2026, and that the system’s PhD level capabilities only apply to some tasks. No mention, of course, of planning capabilities.