I'm not sure there's a single metric in which you can argue a single human brain has more compute than 1e6 H100s
Bc… we can do training and inference at the same time… while maintaining balance, operating machinery, or any number of tasks which are computationally expensive for digital computers. Maybe that implies we train and infer more efficiently, but it may just as well imply that our brains are more computationally powerful.
Maybe our conscious mind has less compute than H100s, but our entire brain maybe not.
It’s an apples to oranges comparison any which way.
I guess the only way to know for sure is to faithfully simulate a human brain inside one of these clusters in real time...
What is a brain operation? How do you measure that?
My point is that you can’t a make your argument because you’re comparing two subjects that have nothing in common beside the vague colloquial notion of “computing”
What and how your brain computes your response to my comment is wildly different than how digital hardware does.
They’re apples and oranges. No way to compare them on even ground, so no way to make the argument you’re trying to make.
We don’t have a shared metric with which to measure digital computation speed and organic brain computation speed, so your argument doesn’t make sense on the outset.
The computing power of cellular machinery in ~3lb of human tissue which includes neurons, are many orders of magnitude larger than 1e6 or even 1e12 H100s.
Quite literally your finger has more compute power than 1e6 H100s, we just use that computational power to stay alive rather than “think.”
It’s allowed because the energy limits in the Landauer's principle are based on irreversible state changes.
When an Antibody doesn’t match a surface protein there’s computation with zero energy loss because there’s no change in state. It amounts to a “free” if statement as long as the result is false and it’s almost always false. Though in a larger context it’s really inefficient as it’s a random process only effective because of how many Antibodies flood the body and how quickly they bounce around and can preform this computation.
Cells heavily leverage this kind of computation to the point where it’s not generally thought of as computation until something happens.
I guess at some point we should only count computations above some abstraction level otherwise we also undercount the computation power of a gpu.
Are we broadly in agreement? Even if we disagree above the relative computation power of the two
However an abacus does not know how or what to add to achieve a goal.
Can an H100 or a bunch of H100s achieve better biological control than the human brain in real time as a self directed agent?
No way.
Apples and oranges.
Fwiw I think 10e6 h100s can act as a competent self directed agent with the right algorithms yes
But your description of Neurons is off. They output chemical signals, which is why painkillers can dull the signal etc. Many drugs bind to receptors on neurons which then block signals which wouldn’t work if these where electrical connections.
The do respond to electrical stimulation, and it’s often said they use electrical signals to carry information along the body of the neuron but the actual signal only moves up to ~120 m/s a long way from how fast electrical signals move. https://en.wikipedia.org/wiki/Nerve_conduction_velocity
As to simulation of a neuron, the brains got 86 billion neurons and the peripheral nervous system is doing useful work here. A full fidelity simulation of even one neuron is completely off the table simply because it’s an analog system where timing matters to an arbitrary precision, but what’s good enough is really just a question of context. Good enough to test new drugs for depression? No, though good enough to drive a car is a reasonable goal.
The brain is a self conscious biological control system. Incidentally it can do some computing.
A computer is a general purpose digital combinatorics machine that generates a constant output for constant input.
The brain is as if you wrote your software in microcode on a real valued analog computer. The brain does not use an instruction set. Instead it uses a complex spaghetti of somewhat self modifying micro code orginating in a self modifying (over longer time scales) genetic blue print.
Most importantly for human and animal intelligence the brain is motivated in certain directions by emotions. Without motivation intelligence does not exist. One could say intelligence is the application of real control toward a goal.
Calling a program “a code,” as in “I wrote a code for that.” Is common use at Los Alamos and Scandia labs, which I discovered when I taught there. So I have encountered some linguistic oddities in my career. Maybe “compute” has been thriving in the AI world and I haven’t been there to notice?