Gm.
The human brain runs on something like 20 watts of power. Less than a lightbulb. How many calculations it can do per second is partially unknown, but based on various estimates over the years the processing power is generally believed to be something like one exaflop per second. Some estimates are lower in the petaflops, while others are some orders of magnitude higher. Obviously “software” matters too, not just raw processing ability. The programming of the processor ensures that the processing capability is used efficiently rather than wasted.
The top superconductors crossed the exaflop level within the past few years. However, they run on like 20 megawatts of power; a million times more power than the human brain. They’re extremely large and energy intensive.
As a result, datacenter processing capability reaches something akin to the processing capability of a human brain well before that level of ability can be installed in a human-sized robot with similar energy consumption levels as a human.
Now, robots can offload some of their processing to datacenters, but still at a relatively high cost per calculation for a while, and at the general bandwidth limit of whatever the best wireless rate is in a region at any given time.
For some calculation types, of course computers passed humans long ago. A basic math calculator, for example, beats the best humans at calculating mathematical formulas. But when we talk about human brain “calculations” what it means is that the brain is taking in enormous amounts of information (all five senses at high fidelity, plus other indirect senses like acceleration/balance and other inputs), calculating it to make sense of it, calculating all sorts of things to interact with the environment, and simultaneously running the processes related to sapient thought and general problem solving.
As a result, it’s far easier to get a robot to work on an assembly line more efficiently than a human, or to calculate an insane number of protein folding tests, and things like that, than it is for a robot to be able to operate as effectively as a human in the real world with countless unexpected hazards.
For example, imagine a hypothetical robot handyman. It can drive out to your house and fix any residential electrical, plumbing, or hvac issue, or help with various miscellaneous things (fix drywall, get something out of a tree, carry stuff out of your attic, etc), and then drive back to the station. This is a shockingly hard problem. First they need extremely advanced mechanical bodies. Second they need processors strong enough and cheap enough to safely operate in 3D space with all sorts of unexpected things happening around them (compared to a highly controlled manufacturing floor), now all of these skills, and interact with language.
So, AI can start helping us offload certain types of white collar remote work and expand medical breakthroughs before it can replace human level in-field skilled physical labor. And it can start helping with specific in-field tasks that require less programming, like a robot dog or robot butler to watch your property or come with the owner around town, listen to owner commands and carry some of them out, and follow basic rules when left alone, well before it can fully replace a human for many in-field things.
Anyway, that’s a general framework or napkin math to help think through the order of impacts that AI can have as it goes up orders of magnitude in power and efficiency in the coming years.
