1000x energy consumption in 10-20 years is a really wild prediction, I would give it a <0.1% probability. It’s several orders of magnitude faster than any previous multiple, and requires large amounts of physical infrastructure that takes a long time to construct.
1000x is a really, really big number.
Baseline
2022 figures, total worldwide consumption was 180 PWh/year[1]
Of that:
Oil: 53 PWh
Coal: 45 PWh
Gas: 40 PWh
Hydro: 11 PWh
Nuclear: 7 PWh
Modern renewable: 13 PWh
Traditional: 11 PWh
(2 sig fig because we’re talking about OOM here)
There has only been a x10 multiple in the last 100 years—humanity consumed approx. 18 PWh/year around 1920 or so (details are sketchy for obvious reasons).
So historically, the fastest rate of doubling has been 20 years.
Build it anyway
It takes 5-10 years for humans to build a medium to large size power plant, assuming no legal constraints. AGI is very unlikely to be able to build an individual plant much faster, although it could build more at once.
Let’s ignore that and assume AGI can build instantly.
What’s in the power plant
At current consumption, known oil, gas and coal reserves are roughly 250 years in total. Thus at 1000x consumption they are consumed in less than three months.
Nuclear fuel reserves are a similar size − 250 years of uranium, so assuming reprocessing etc, let’s say 1000-2000 years at 2022 consumption.
So the AGI has less than 3 years of known fuel reserves at 1000x current consumption.
However, “reserves” means we know where it is and how much could be economically extracted. Exploration will find more, and of course there are many other, more esoteric methods of electricity generation known or believed to be possible but currently uneconomic or unknown how to build.
How about Space?
Solar irradiance is roughly 1380 W/m2 at Earth’s orbital distance. Call it 12 MWh/year/m2, or 12 TWh/year/km2
We’re looking for 180,000,000 TWh/year, so we need a solar panel area of around 20,000,000 km2 at >50% efficiency. That’s a circle >2500km radius—much bigger than the Moon!
Fusion
The hidden assumption is that AGI not only figures out large-scale fusion in 4 years, but rolls it out immediately.
I strongly disagree. The underlying reason is that an actual singularity seems reasonably likely.
This involves super-exponential growth driven by vastly superhuman intelligence.
Large scale fusion or literal dyson spheres are both quite plausible relatively soon (<5 years) after AGI if growth isn’t restricted by policy or coordination.
I think you aren’t engaging with the reasons why smart people think that 1000x energy consumption could happen soon. It’s all about the growth rates. Obviously anything that looks basically like human industrial society won’t be getting to 1000x in the next 20 years; the concern is that a million superintelligences commanding an obedient human nation-state might be able to design a significantly faster-growing economy. For an example of how I’m thinking about this, see this comment.
I was surprised by this number (I would have guessed total power consumption was a much lower fraction of total solar energy), so I just ran some quick numbers and it basically checks out.
This document claims that “Averaged over an entire year, approximately 342 watts of solar energy fall upon every square meter of Earth. This is a tremendous amount of energy—44 quadrillion (4.4 x 10^16) watts of power to be exact.”
Our World in Data says total energy consumption in 2022 was 179,000 terawatt-hours
Plugging this in and doing some dimensional analysis, it looks like the earth uses about 2000x the current energy consumption, which is the same OOM.
173,000 terawatts of solar energy strikes the Earth continuously. That’s more than 10,000 times the world’s total energy use.
But plugging this number in with the OWiD value for 2022 gives about 8500x multiplier (I think the “more than 10000x” claim was true at the time it was made though). So maybe it’s an OOM off, but for a loose claim using round numbers it seems close enough for me.
[edit: Just realized that Richard121 quotes some of the same figures above for total energy use and solar irradiance—embarrassingly, I hadn’t read his comment before posting this, I just saw kave’s claim while scrolling and wanted to check it out. Good that we seem to have the same numbers though!]
Drosophila biomass doubles every 2 days. Small things can assemble into large things. If AGI uses serial speed advantage to quickly design superintelligence, which then masters macroscopic biotech (or bootstraps nanotech if that’s feasible) and of course fusion, that gives unprecedented physical infrastructure scaling speed. Given capability/hardware shape of AlphaFold 2, GPT-4, and AlphaZero, there is plausibly enough algorithmic improvement overhang to get there without needing to manufacture more compute hardware first, just a few model training runs.
Here’s something that I suspect is getting missed: AI… Even a hyperintelligent AI, doesn’t have hands. Even a post singularity AI can’t build a power plant or a transmission line. Furthermore, the map is not the territory, a hyperintelligent AI might be able to solve the string theory equations, and design a futuristic power plant, like Tony Stark’s arc reactor, but it’s <1% likely to be something we can build with present day tooling that we have at scale. We’re almost certainly going to have to (assuming that it gives us the designs) build new factories to build the tooling, to build the factories to build the tooling, to build the factory to build the arc reactors Likely (90%+). And that’s AFTER we build the exoscale LHC to validate the theoretical predictions (because any number of theories with valid math have been proven wrong). Who builds that? All of that is going to have to be built by humans, because we don’t have the robots, or the factories to build the robots, or… And during this time, civilization can’t “stop”, those humans still need to be fed, need entertainment, and sleep, and downtime. Sure, there may be more labor available since pretty much all office jobs will be gone, but training those humans to be useful an a construction site, or factory floor? Not so trivial.
And no guiding intelligence, regardless of how smart, can escape those constraints. It may be able to tell us exactly where to go to find the rare earths, but it can’t move the dirt, build the refinery, or fabricate the finished products.
Dyson spheres? Where is the rocketry? Orsay it does better than that, say, repulsive lift off of earth’s magnetic field, where are the miles of superconductors going to come from? New superconductive materials? Where’s the factory to build them? I doubt that you’re going to just “retool” a present wire drawing factory to draw unobtanium dioxide wire. Where’s the asteroid mining infrastructure going to come from in the 5 year timescale?
IF the Manhattan project had been furnished ab initio with a complete blueprint of an A bomb, complete with BOM, where to source the required materials, it would STILL have taken likely 4 years to build the first one. Oak ridge wasn’t built in a day. And that’s to build a single 5 ton device, now thousands of power plants, thousands of mines, refineries, fabs, thousands of miles of transmission lines, etcetera. the scale of 1000x ing power generation is similarly 1000X the scale of the manhattan project or the moonshot.
A superintellignece can accelerate all of this, most definitely, but only SO much. you can’t build an Iphone with a rock. Even with a detailed schematic.
1000x energy consumption in 10-20 years is a really wild prediction, I would give it a <0.1% probability.
It’s several orders of magnitude faster than any previous multiple, and requires large amounts of physical infrastructure that takes a long time to construct.
1000x is a really, really big number.
Baseline
2022 figures, total worldwide consumption was 180 PWh/year[1]
Of that:
Oil: 53 PWh
Coal: 45 PWh
Gas: 40 PWh
Hydro: 11 PWh
Nuclear: 7 PWh
Modern renewable: 13 PWh
Traditional: 11 PWh
(2 sig fig because we’re talking about OOM here)
There has only been a x10 multiple in the last 100 years—humanity consumed approx. 18 PWh/year around 1920 or so (details are sketchy for obvious reasons).
Looking at doubling time, we have:
1800 (5653 TWh)
1890 (10684 TWh) − 90 years
1940 (22869 TWh) − 50
1960 (41814 TWh) − 20
1978 (85869 TWh) − 18
2018 (172514 TWh) − 40
So historically, the fastest rate of doubling has been 20 years.
Build it anyway
It takes 5-10 years for humans to build a medium to large size power plant, assuming no legal constraints.
AGI is very unlikely to be able to build an individual plant much faster, although it could build more at once.
Let’s ignore that and assume AGI can build instantly.
What’s in the power plant
At current consumption, known oil, gas and coal reserves are roughly 250 years in total.
Thus at 1000x consumption they are consumed in less than three months.
Nuclear fuel reserves are a similar size − 250 years of uranium, so assuming reprocessing etc, let’s say 1000-2000 years at 2022 consumption.
So the AGI has less than 3 years of known fuel reserves at 1000x current consumption.
However, “reserves” means we know where it is and how much could be economically extracted.
Exploration will find more, and of course there are many other, more esoteric methods of electricity generation known or believed to be possible but currently uneconomic or unknown how to build.
How about Space?
Solar irradiance is roughly 1380 W/m2 at Earth’s orbital distance. Call it 12 MWh/year/m2, or 12 TWh/year/km2
We’re looking for 180,000,000 TWh/year, so we need a solar panel area of around 20,000,000 km2 at >50% efficiency.
That’s a circle >2500km radius—much bigger than the Moon!
Fusion
The hidden assumption is that AGI not only figures out large-scale fusion in 4 years, but rolls it out immediately.
Hannah Ritchie, Pablo Rosado and Max Roser (2020) - “Energy Production and Consumption”
https://ourworldindata.org/energy-production-consumption
I strongly disagree. The underlying reason is that an actual singularity seems reasonably likely.
This involves super-exponential growth driven by vastly superhuman intelligence.
Large scale fusion or literal dyson spheres are both quite plausible relatively soon (<5 years) after AGI if growth isn’t restricted by policy or coordination.
I think you aren’t engaging with the reasons why smart people think that 1000x energy consumption could happen soon. It’s all about the growth rates. Obviously anything that looks basically like human industrial society won’t be getting to 1000x in the next 20 years; the concern is that a million superintelligences commanding an obedient human nation-state might be able to design a significantly faster-growing economy. For an example of how I’m thinking about this, see this comment.
IIUC, 1000x was chosen to be on-the-order-of the solar energy reaching the earth
I was surprised by this number (I would have guessed total power consumption was a much lower fraction of total solar energy), so I just ran some quick numbers and it basically checks out.
This document claims that “Averaged over an entire year, approximately 342 watts of solar energy fall upon every square meter of Earth. This is a tremendous amount of energy—44 quadrillion (4.4 x 10^16) watts of power to be exact.”
Our World in Data says total energy consumption in 2022 was 179,000 terawatt-hours
Plugging this in and doing some dimensional analysis, it looks like the earth uses about 2000x the current energy consumption, which is the same OOM.
A NOAA site claims it’s more like 10,000x:
But plugging this number in with the OWiD value for 2022 gives about 8500x multiplier (I think the “more than 10000x” claim was true at the time it was made though). So maybe it’s an OOM off, but for a loose claim using round numbers it seems close enough for me.
[edit: Just realized that Richard121 quotes some of the same figures above for total energy use and solar irradiance—embarrassingly, I hadn’t read his comment before posting this, I just saw kave’s claim while scrolling and wanted to check it out. Good that we seem to have the same numbers though!]
Drosophila biomass doubles every 2 days. Small things can assemble into large things. If AGI uses serial speed advantage to quickly design superintelligence, which then masters macroscopic biotech (or bootstraps nanotech if that’s feasible) and of course fusion, that gives unprecedented physical infrastructure scaling speed. Given capability/hardware shape of AlphaFold 2, GPT-4, and AlphaZero, there is plausibly enough algorithmic improvement overhang to get there without needing to manufacture more compute hardware first, just a few model training runs.
Thank you!
Here’s something that I suspect is getting missed: AI… Even a hyperintelligent AI, doesn’t have hands. Even a post singularity AI can’t build a power plant or a transmission line. Furthermore, the map is not the territory, a hyperintelligent AI might be able to solve the string theory equations, and design a futuristic power plant, like Tony Stark’s arc reactor, but it’s <1% likely to be something we can build with present day tooling that we have at scale. We’re almost certainly going to have to (assuming that it gives us the designs) build new factories to build the tooling, to build the factories to build the tooling, to build the factory to build the arc reactors Likely (90%+). And that’s AFTER we build the exoscale LHC to validate the theoretical predictions (because any number of theories with valid math have been proven wrong). Who builds that? All of that is going to have to be built by humans, because we don’t have the robots, or the factories to build the robots, or… And during this time, civilization can’t “stop”, those humans still need to be fed, need entertainment, and sleep, and downtime. Sure, there may be more labor available since pretty much all office jobs will be gone, but training those humans to be useful an a construction site, or factory floor? Not so trivial.
And no guiding intelligence, regardless of how smart, can escape those constraints. It may be able to tell us exactly where to go to find the rare earths, but it can’t move the dirt, build the refinery, or fabricate the finished products.
Dyson spheres? Where is the rocketry? Orsay it does better than that, say, repulsive lift off of earth’s magnetic field, where are the miles of superconductors going to come from? New superconductive materials? Where’s the factory to build them? I doubt that you’re going to just “retool” a present wire drawing factory to draw unobtanium dioxide wire. Where’s the asteroid mining infrastructure going to come from in the 5 year timescale?
IF the Manhattan project had been furnished ab initio with a complete blueprint of an A bomb, complete with BOM, where to source the required materials, it would STILL have taken likely 4 years to build the first one. Oak ridge wasn’t built in a day. And that’s to build a single 5 ton device, now thousands of power plants, thousands of mines, refineries, fabs, thousands of miles of transmission lines, etcetera. the scale of 1000x ing power generation is similarly 1000X the scale of the manhattan project or the moonshot.
A superintellignece can accelerate all of this, most definitely, but only SO much. you can’t build an Iphone with a rock. Even with a detailed schematic.