Human Intelligence vs AI
I had a very candid discussion with an old IITian friend of mine regarding AI and Human Intelligence and I argued that we are highly energy-friendly as compared to any AI
If our brain Processing Speed can be superimposed, then it only requires "4 bulbs of energy consumption" to achieve the processing speed of today's supercomputer.
"Energy consumed by the brain" accounts for 20% of the energy consumed by the human body every day. Ordinary people consume 2000Kcal per day, so the energy consumed by the brain every day is about 400Kcal. The brainpower is actually stable, about 20 watts - It includes:
Energy consumption of "CPU, GPU, hard disk, memory" and other devices;
Energy consumption of "Visual processing, voice processing, language processing" and other functions when they are fully activated.
Let us estimate the energy consumption of AlphaGo in the game with Lee Sedol:
1202pieceofCPU?100watts+176pieceofGPUs?300watts=173kW1202pieceofCPU?100watts+176pieceofGPUs?300watts=173kW
This AI is simple and does not have intelligence beyond Go:
The function of its visual part, "Reading the checkerboard", was done by the engineers of DeepMind;
In order to save energy, it has no function like voice, language processing and so on.
If only CPU and GPU consumption is counted, its power is already 173 kW. (Of course, the current energy consumption of CPU and GPU is far from approaching the theoretical lower limit.)
AlphaGo is a "simple Go function" AI, which consumes about 8650 times the energy of Lee Sedol's Brain. Not to mention that even the "most computationally consuming visual tasks" have not been realized, but its energy consumption has increased thousands of times.
Yes, supporting a powerful AI will inevitably require a strong computer. But once the computing power becomes stronger, its energy consumption will not fall.
Classic computers are irreversible computers. Because NAND gates are irreversible, classic computer constantly erases information when computing. According to the Landauer's Principle, expunge per 1-bit information "need to consume energy/released heat" is kTln2 (k is the Boltzmann constant and T is the ambient temperature). Assuming that each "input data" is processed by N "irreversible logic gates" on average, the theoretical lower limit of energy consumption is NkTln2.
This leads to a huge energy consumption ratio - in doing similar tasks, AI's energy consumption will be many orders of magnitude higher than the human brain's.
Does the human brain not obey the Landauer principle?
Although the human brain is also an irreversible calculation, the energy consumption of the human brain when performing many tasks is low. Especially for those tasks that human survival is critical, such as vision, reinforcement learning, and so on — This is the result of continuous optimization during our long evolution. The folding method, including the structure of the neural network, as small as protein's, has undergone an extensive natural selection. Those inefficient and energy-intensive calculational patterns will be eliminated.
We are not able to clearly know that "what is the human brain based computing unit." However, when dealing with the tasks that humans are good at, the energy consumption of the human brain is much lower than that of electronic computers. This is experimentally verified. In dealing with the tasks that humans are good at, as long as the electronic computers also rely on the irreversible Classical logic gate Operation, its theoretical minimum energy consumption will be higher than the human brain. Its problem is the classic logic gate operation - when every time the NAND gate is erasing information, processing the 1-bit input data according to different tasks, and then need to reuse a large number of NAND gate operations.
to sum up:
Evolution brings "low-energy cost human brain design";
When dealing with tasks that human good at, the human brain's energy consumption is lower than the theoretical minimum energy consumption of electronic computers (due to classical logic gates).
Although both the human brain and the computer obey the Landauer principle, their basic computing units have different designs, so the computer has an unmaskable energy ratio.