I’ve been thinking about the idea that the massive data center expansion needed to drive the current AI boom is increasing energy demand, and in the near to medium term may be limited by energy supply. It is also apparently competing for the energy supply we have currently, driving up prices for consumers and businesses across the board, along with toxic air pollution and greenhouse gas pollution. The big technology companies are trying to tell a story that this short term pain will go away when AI “solves climate change” with its superior intelligence. That seems like a longer-term thing that will not solve our short term issues, although the tech-bro-cracy will tell us we are just not smart enough to think exponentially like them, and what we think is long term is in fact just around the corner.

I’ve been thinking about the Kardashev Scale. This is the idea that the level of sophistication of a civilization can be linked to the amount of energy it uses – just the energy striking the surface of the planet, all the energy of its star, all the energy of all the stars in its galaxy, etc. (we are just off the left end, trying to get onto this scale). This might seem not to square with the idea that more energy use means more pollution, negative health and environmental impacts for us at our current level of technology. The way you square that logically is to say that we have to get over the hump from dirty to clean energy technology before we can take the next step to a higher order of civilization. If AI can increase our rate of scientific and technology advancement, maybe it can help us get there sooner. Currently we are at the stage where we are turning decommissioned coal plants back on to fuel the AI, without seeing any measurable benefits so far. Realists/cynics (don’t cynics always consider themselves realists?) might also point out that there is a lot of friction built into human institutions and sociopolitical systems, and that this tends to set a speed limit on the rate we can progress, even if the state of science and technology could allow us to progress faster.

One thing I always wondered was what came first – the coal chicken or the steam engine egg? It turns out, this is a true chicken and egg story that went something like this (citation: Gemini for the background, not fact checked – the words are from my human brain):

• The ancient Romans, Chinese, and other cultures were aware of coal and burned it to heat buildings. There isn’t really a record of when it was “discovered”, you can imagine deposits of it were first found just lying around on or very near the surface, and people found it burned hotter and longer than wood.
• By the 1700-1800s, London grew to the point that England started to run out of accessible coal to heat all the buildings. They were digging deeper and deeper to get at the coal, and starting to hit the water table. So the first steam engine was invented in 1712 to power pumps to lower the groundwater table in coal mines. This was a dirty, inefficient machine but they didn’t care because it was literally at the mine.
• From there, the steam engine technology was refined (famously by James Watt in the 1760s) and gave rise to railroads and steam ships (both of which could move coal), and coal-powered factories. From there, the feedback loop of economic growth and energy demand (and localized toxic air pollution and global greenhouse gas emissions, not to mention labor exploitation and industrial war) took off.

So an analogous process today, we might want to think, is that the sharp increase in energy demand for the data centers will incentivize innovation around clean energy, or at least cheaper energy than what fossil fuels are currently providing. Fossil fuels seem to be pretty close to the limit of how much power they can reasonably supply our civilization, both economically and in terms of pollution we are willing to tolerate. It is almost an iron law of economics that this will fuel innovation. So AI is on the energy demand side of the equation – there is no requirement that it also has to be on the innovation side of the equation because humans are doing pretty well advancing clean energy technology (contrary to the fossil fuel industry-funded propaganda we are being force-fed in the U.S.). But humans will figure out ways to use the AI technology to innovate, and at some point maybe AI will be able to do some innovation on its own with high-level human oversight. It will make a big difference whether this process takes years or decades to unfold (I am not buying the “months” time frame being pushed by the tech-bro-cracy). If it’s decades, the ecological impacts and human suffering are going to be enormous.

What kinds of technologies could possibly be boosted by AI? This is just me brain storming:

• Somewhat obviously, we can continue refining/scaling up the solar, wind, and battery technology we have today. Electrification of buildings, industry, and transportation are important here too, and a smarter/more resilient electric grid is also a big part of this equation. Scaling these involves the high-friction human institution-mediated processes where AI can probably point out how we can do things more rationally and efficiently, but our human inefficiency and irrationality are going to limit the pace for the foreseeable future. AI can probably design much better institutions for us, but we would have to come to some consensus and then successfully implement them. This may happen very slowly – but I see this taking decades for sure, and not much AI will be able to do about it.
• Fusion – Let’s try to pick up the pace of basic research leading to safe commercialization.
• Fission – I don’t think we should give up on it. The small modular reactors seem to hold out some promise even if there is also a certain amount of industry exaggeration. There is also a big push-pull between industry, government regulation, and public acceptance, and I am not sure this is something we want to speed up. It may need to just take its course to find solutions we agree are both effective and safe enough – again, decades. If we actually get fusion figured out, maybe this technology will be obsolete before it can be scaled.
• Space-based solar technology – The energy from just our one Sun is unlimited in any practical near-term sense. There must be lots of technical problems to be worked out here – straightforward mechanical/electrical/chemical engineering stuff. Let’s get to it! The private sector may actually take care of this one without much prompting. No, I don’t want to make Elon Musk even richer, but he is a leader on this so how about some real competition. Even if a few people get unreasonably rich, there are plenty of policy options to share the wealth. AI can easily point these out for us, but again this is one of those human socio-political institutional friction areas where any progress at all is going to be painfully slow, even if we can reverse current trends and get it moving back in the right direction.
• Applied research on energy storage, materials science, etc – there are plenty of ideas for storing energy beyond our current battery technologies. You can simply pump water up a hill while the sun is out and let it flow back down, turning a turbine when the sun goes in. You can take advantage of pressure/temperature/salinity gradients in the ocean, or electrolyze water to charge hydrogen fuel cells, just to name a few. Intelligent AIs should be able to work on problems involving materials and chemicals, first using numerical (differential equation based) simulations, and later bringing robots in real laboratories into the mix to test out the most promising ideas.
• Fundamental research to unlock the things we can’t even think of yet – Physicists are forever searching for the holy grail of understanding gravity, relativity, quantum physics, dark matter, and probably other things I have not heard of. I have heard some of them say that new branches of mathematics may need to be invented (discovered? a philosophical question) before the right questions can even be asked. Then new instruments to collect new forms of data may need to be invented before experiments can even be designed to begin answering the new questions. This seems like an area where if AIs are eventually as capable as a leading human physicist, you can have a team of a million or billion or trillion leading physicists working on a problem 24 hours a day, without ever getting tired. One tiny little detail – we need ethical frameworks and nation-state regulations and enforceable international agreements to control the dangerous stuff that could result.