CPUs are good at math. They have specialized structures that perform addition, subtraction, multiplication, and division. Of course, a CPU can't just look at a math problem and solve it, on account of the fact that a CPU can't see. A human has to use a complicated, specialized series of technologies (keyboard, motherboard, RAM, monitor, operating system, calculator program, etc.) in order to take advantage of the CPU's capabilities. In other words, a CPU can't encounter a math problem (it can't "encounter" anything, it's a rock) and solve it easily. A math problem out in the world needs to be converted into the CPU's internal representation before it can solve it, and then it needs to be converted out of that representation for the result to be usable.
Neurons are good at detecting serotonin. They have specialized structures that react selectively to serotonin. Of course, a neuron can't just look at a molecule and say if it's serotonin, on account of the fact that a neuron can't see. A human has to use a complicated, specialized series of technologies (I don't know enough about bio, but there's this process where they grow a cell line of neurons in a petri dish that are modified so that they light up after detecting serotonin, and then you can add a solution of the compound you want to test to the dish and use some light-sensitive equipment to see if it's serotonin) in order to take advantage of the neuron's capabilities. In other words, a neuron can't encounter a chemical identification problem (it can't "encounter" anything, it's a single cell) and solve it easily. A chemical out in the world needs to be brought into physical contract with the neuron before it can identify it, and then it needs to be modified to glow for the result to be usable.
Humans aren't good at detecting serotonin. In the roughly 300,000 years of human existence preceding 1935, no human ever managed to detect serotonin. Only through extensive training and technological augmentation was Vittorio Erspamer able to finally detect some. If I gave you a chemical solution, there's pretty much no chance you'd be able to know whether it contained serotonin. Even if you drank it, serotonin doesn't cross the blood-brain barrier, so it wouldn't produce a noticeable effect. If you went further and injected it directly into your brain (and at this point it's not really a human detecting serotonin, it's a human plus a brain needle), how sure are you of your ability to differentiate a rush of serotonin from a dopamine rush? Or an adrenaline rush? What if the mixture contains both serotonin and a GABAergic? This isn't to say humans can't learn to be good at detecting serotonin - I'm sure there's some biology researcher out there with 30 years of experience who can identify it by smell or something. However, we're not inherently good at it like a neuron is, and even the best human is probably worse at serotonin identification than the average serotonin-sensitive neuron.
AIs aren't good at math. For reference, here's what happened when I made up a math problem and typed it into InferKit (formerly Talk to Transformer):
For reference, the correct answer is 2528. This isn't a fluke - researchers have studied AI performance with harder problems and found that they get less than 7% of problems right and you can go play around with it yourself to see how well InferKit does with various types of problems. I'm sure it's possible to train a neural net to be good at math problems - but they'll never be as fast or accurate as the CPUs they run on. Hopefully now you understand why.
Wait why did i make this
- I'm tired of movie depictions of AIs as all being these nerdy supergeniuses who can instantly solve math problems
- The general association between AI and math is annoying to me. Yes, AI algorithms are mathematically definable. So are human ones - your neurons use a very advanced technique known as "addition" to determine whether to fire, and your behavior is emergent those firings and also something involving glia neurobio people please be nice to me
- The worst is when people use this sort of AI-math association as an argument against mechanism
-
There is an interesting digression about what a near-human AI's math ability would be like. Since their mind would
presumably be running on a CPU, they could probably be given a way to directly, explicitly access the CPU for arithmetic problems.
I think this would end up being something like a thought-activated calculator - imagine a voice-activated calculator
(actually, most smartphones count), but instead of having to say "ok google, what's 254 * 118", you can just think it really loudly.
This would probably make you... slightly faster than using a normal calculator!!! Truly a revolutionary mathematical mind.
Interesting digression part 2: isn't the AI having a thought-activated calculator the same as being good at arithmetic?
Is a human having a normal calculator them being good at arithmetic? If a human got a thought-calculator chip
implanted in their brain, would the human be good at arithmetic? I answer "no" to all of the above, because
the AI/human brain consists of many smaller parts with a dense and high-bandwidth web of connections, as does
a calculator (although less so than a sapient mind), but the connection between the two, even for a brain-chip calculator,
is comparatively low-bandwidth, and consists of only a few connections. It's like two computers connected by an ethernet
cable. Here's a visualizer:
remixed from https://commons.wikimedia.org/wiki/File:11-simplex_graph.svg and https://commons.wikimedia.org/wiki/File:7-simplex_graph.svg - After learning more biology, I've realized there's a neat analogy: animals run using chemistry as their substrate, and so we're actually incredibly good at detecting minute concentrations of (some) chemicals: that's what a nose is! I think an AI with the sort of "calculator access circuit" described above could be very good at math in the same way that biological organisms are good at smelling.