Are neurons really more efficient at computing than transistors ?
In particular the human brain is an 80-billion core processor with each core running at 300Hz. Its architecture is not so good at FLOPs and other things we typically use computers for, but overall human brains dramatically outperform anything in 2024 except the world's fastest supercomputer.
Meaning at specific tasks, right? Because obviously even the weakest modern computers out perform humans at certain kinds of tasks.
No, they are really not.
A Neuron on its own is capable of exactly one thing, and that's signal integration from it's dendrites. That's it. Wait until the excitation potential reaches the point where the Potassium channels switch and boum: Signal cascade goes down the axon.
The only moderation a single neuron is capable of, is varying signal resistance to filter out repeating noise.
> https://www.quantamagazine.org/how-computationally-complex-i...
Computational complexity doesn't indicate computational capability.
There are any number of systems that stubbornly refuse to be capable supercomputers, despite being incredibly hard to simulate.
Accurately simulating every aspect of the liquid flow through an acorn leaf, with it's billions of micro and nanometer channels, would easily be one of the most difficult tasks in the history of computing, and yet, acorn leaves are not even capable of the simplest thought, and the entire modus operandi of water flow through an entire tree can be accurately summarized in a simple 1-page diagram.
Every single muscle fiber with its trillions of moving cytoskeleton components, has a much higher complexity than a neuron, that doesn't mean muscles are supercomputers.
Add numbers, no of course not, write a book though…
251MJ is what 2000kcal diet for a month works out to. That’s also 2.5kW for 100 servers for 1000 seconds.
Is it true that a human would write a better book in a month than a well-designed algorithm?
Maybe.
The "well designed" plays a big role in this sentence. At some point you're just writing the book with a different "pen".
Somewhere in there there's some efficiencies being achieved that come from being functionally active in ways that it don't make for an easy one to one comparison to transistors.
But what I'm saying is, there is a layer (or layers) of organic function that is abstracted away in machine learning/AI. While the field does borrow ideas liberally from brain architecture (as it well should), you nevertheless have neurons doing things like chemical signalling, metabolic regulation, unique site-specific structural organization, and processes for dynamic reorganization of structure as needed.
That is, not just in the informational representation, which can be sufficiently abstract to model practically anything, but in the hardware itself upon which the information rests. It's there that I suspect there are efficiencies gained by brains over transistors.