No, we don't have enough energy to reach AGI

I've seen several times the notion in this sub that we have more than enough computing power to reach AGI. We simply need better algos like transformers that will somehow magically lead us to AGI.<br><br>I haven't seen anyone quantifying how much energy we would need, so I've done it myself.<br><br># Estimates of AGI energy needs<br><br>We can estimate how much energy AGI has to use by studying the brain, which is an AGI. The human brain is roughly 2% of the body's mass, yet it uses 20% of its energy. After adjusting for body mass differences, we can estimate the brain uses the same amount of energy as the entire human body when in a resting state. By my estimates, we need roughly 100 kW of power to run AGI. That's the power usage of 5 modern GPUs (or 60 GPUs from 2010). It's also the power usage of a small house, a large pickup truck, or a small sailing boat. It's not that much.<br><br>We can estimate how much energy is required to run AGI by studying how much energy the brain uses. The brain, which is an AGI, uses roughly 20% of the human body's energy at rest. After adjusting for body mass differences, it can be estimated that the brain requires the same amount of energy as the entire human body in a resting state. By this calculation, AGI requires approximately 100 kW of power to operate. This is equivalent to the power consumption of 5 modern GPUs (or 60 GPUs from 2010). It is also comparable to the energy used by a small house, a large pickup truck, or a small sailboat, indicating that the energy requirement for AGI is not substantial.<br><br>Note that tis is a rough estimate, and the brain is highly optimized for power usage. We probably need far more energy to run AGI. <br><br>On the other hand, the brain runs at 5 picojoules per flop, while modern hardware runs at roughly 100 picojoules per flop (5 pJ/FLOP ≈ 0.2 nW/FLOP; 100 pJ/FLOP = 20 nW/FLOP); this is a 20x difference in favor of the brain, but it's hard to calculate how much it'd translate to in terms of AGI since the brain's flops aren't equal to our flops.<br><br># Estimates of global computing power<br><br>We can estimate how much global aggregate computing power we have by looking at international electrical output in terms of GW or TWh, and then calculating roughly how much computing power is equivalent to that.<br><br>Global electricity production is roughly 30 PWh per year. We can estimate the aggregate computing power of the world by assuming that all of this electrical output is used for computing. Assuming a PUE of 1.1–1.2, an average server power efficiency of 450 W per TFLOP, and an average electrical power usage of 100 W per TFLOP, 1 TFLOP consumes roughly 0.1-0.2 TWh per year. Converting global electrical output into computing power, we find that the aggregate computing power of the world is approximately 200-400 zettaflops. This is equivalent to 200-400 billion TFLOPS, or the computing power of 400-800 GPUs.