I know current learning models work a little like neurons but why not just make a sim that works exactly like how we understand neurons work
Because we don’t understand it.
To clarify:
We don’t even know how human intelligence/consciousness works, let alone how to simulate it.
But we know how an individual neuron works.
The issue with OPs idea is we don’t know how to tell a computer what a bunch of neurons do to create an intelligence/consciousness.
Heck, we barely know how neurons work. Sure, we’ve got the important stuff down like action potentials and ion channels, but there’s all sorts of stuff we don’t fully understand yet. For example, we know the huntingtin protein is critical to neuron growth (maybe for axons?), and we know if the gene has too many mutations it causes Huntington’s disease. But we don’t know why huntingtin is essential, or how it actually effects neuron growth. We just know that cells die without it, or when it is misformed.
Now, take that uncertainty and multiply it by the sheer number of genes and proteins we haven’t fully figured out and baby, you’ve got a stew going.
To understand the complexity of the human brain, you need a brain more complex than the human brain.
We don’t really understand how real neurons learn.
We’ve got some really good theories, though. Neurons make new connections and prune them over time. We know about two types of ion channels within the synapse - AMPA and NMDA. AMPA channels open within the post-synapse neuron when glutamate is released by the pre-synapse neuron. And the AMPA receptor allows sodium ions into the dell, causing it to activate.
If the post-synapse cell fires for a long enough time, i.e. recieves strong enough input from another cells/enough AMPA receptors open, the NMDA receptor opens and calcium enters the cell. Typically an ion of magnesium keeps it closed. Once opened, it triggers a series of cellular mechanisms that cause the connection between the neurons to get stronger.
This is how Donald Hebb’s theory of learning works. https://en.wikipedia.org/wiki/Hebbian_theory?wprov=sfla1
Cells that fire together, wire together.
Name checks out
Trial and error.
That’s kinda the idea of neural network AI
The problem is that neurons aren’t transistors, they don’t operate in base 2 arithmetic, and are basically an example of chaos theory, where a system is narrow enough for outer bounds to be defined, yet complex enough that the amount of “picture resolution” needed to be able to accurately predict how it will behave is currently beyond our scope of understanding to replicate or even theorize on.
This is basically the realm where you’re no longer asking for math to fetch a logical answer to a question and more trying to use it as a way to perfectly calculate the future like an oracle trying to divine one’s own fate from the stars. It even comes with its own system of cool runes!
I fully imagine we will have a precise calculation of Rayo’s Number before we have a binary computer capable of being raised as a human with a fully human intelligence and emotional depth.
More likely I see the “singularity” coming in the form of someone who figures out how to augment human intelligence with an AI neural implant capable of the sorts of complex calculations that are impossible for a human mind to fathom while benefiting from human abilities for pattern recognition to build more accurate models.
If someone figures out how to do this without accidentally creating a cheap 80’s slasher villain, it will immediately become the single most sought after medical device in human history, as these new augmented mind humans will instantly become a major competitive pressure for even most manual labor jobs.
First, we don’t understand our own neurons enough to model them.
AI’s “neuron” or node is a math equation that takes a numeric input with a variable “weight” that affects the output. An actual neuron a cell with something like 6000 synaptic connections each and 600 trillion synapses total. How do you simulate that? I’d argue the magic of AI is how much more efficient it is comparatively with only 176 billion parameters in GPT4.
They’re two fundamentally different systems and so is the resulting knowledge. AI doesn’t need to learn like a baby, because the model is the brain. The magic of our neurons is their plasticity and our ability to freely move around in this world and be creative. AI is just a model of what it’s been fed, so how do you get new ideas? But it seems that with LLMs, the more data and parameters, the more emergent abilities. So we just need to scale it up and eventually we can raise the.
AI does pretty amazing and bizarre things today we don’t understand, and they are already using giant expensive server farms to do it. AI is super compute heavy and require a ton of energy to run. So, the cost is a rate limiting the scale of AI.
There are also issues related to how to get more data. Generative AI is already everywhere and what good s is it to train on its own shit? Also, how do you ethically or legally get that data? Does that data violate our right to privacy?
Finally, I think AI actually possess an intelligence with an ability to reason, like us. But it’s fundamentally a different form of intelligence.
With current technology, a supercomputer capable of that would be absolutely gigantic, immobile, and have an insane power draw. How’re you going to raise a building like a human?
Currently, a mouse brain is about the limit of what we can do. https://www.cell.com/neuron/fulltext/S0896-6273(20)30067-2
There’s actually a Robert Miles video on this very question.
Was wondering if Robert Miles - Children had a music video with a lot of foresight.
https://youtu.be/DvyCbevQbtI
A programmer’s pet peeve is someone who says “why can’t you just…”.
But the fundamental problem with your plan, assuming it’s possible at all - it’s been said that if the brain were simple enough for us to understand then we’d be too simple to understand it - is that you’re going to want to make your AI at least as smart as someone who’s 30-40 years old, which by definition would take 30-40 years.
Simple answer: We don’t have any computer to run that on. While I don’t see any absolute limitations ruling out that approach… The human brain seems to have hundreds or thousands of trillions of connections. With analog electrical impulses and chemistry. That’s still sci-fi and even the largest supercomputers can’t do it as of today. I think scientists already did it for smaller brains like those from flies(?), so the concept should work.
And then there is the question what are you going to do with it. You can’t just kill a human, freeze the brain, slice it and then digitize it by looking at a microscope a trillion times. So you have to make it learn from ground up. And this requires a connection to a body. So you also need to simulate a whole body and the world it’s in on top. To make it learn anything and not just activate random neurons. So that’s going to be sci-fi (like the Matrix) for the near and mid future.
You can’t raise it like a human because is not a human. Are you going to put it the size of baby? Gonna pump it with hormones that change its structure when it becomes a teen?
deleted by creator
Just some thoughts:
-
Current LLMs (chat AIs) are “frozen brains.” (Over-)Simplified, the synapses on the AI’s input neurons are given the 2048 prior words (the “context”) and the AI’s output synapses mean a different word each, so the synapse that lights up most strongly is the next word the AI will say. Then the picked word is added to the “context” and the neural network is executed once more for the next next word.
-
Coming up with the weights of the synapses takes insane effort (run millions of books through the “context” and look if the AI t predicts the next word correctly, if not, change a random synapse). Afaik, GPT-4 was trained on more than 2000 NVidia A100 GPUs for somewhere around 4 to 7 months, I think they mentioned paying for 7.5 Megawatt hours.
-
If you had a super computer that could keep running the AI with live training, the AI’s ability to string up words would likely, and quickly, degrade into incoherence because it would just ingest and repeat whatever went into it. Existing biological brains have these complex mechanisms of distilling experiences and evaluating them in terms of usefulness/success of their own actions.
.
I think that foundation, that part that makes biological brains put the action/consequence in the foreground of the learning experience, rather than just ingesting, is what eludes us. Perhaps at some future point in time, we could take the initial brain structure that grows in a human as the seed for an AI (but I guess then we’d likely have to simulate all the highly complex traits of real neurons, including mixed chemical and electrical signaling and possibly even quantum-level effects that have been theorized).
-
You wouldn’t need to raise it as a baby.
The reason that humans come out as babies in the first place is because if we came out with fully developed brains, our heads would be crushed through the birth canal and the mother would probably die. Therefore, our brains have to mature as we get older which of course takes decades.
Growing up is a biological imperative.
In terms of artificial intelligence or large language models, there would be no need to actually grow in physical size.
Which solidifies the point a person already made here is that it would be a fundamentally different kind of intelligence one that simply needs data input And will not need the ability to grow up as a child would.
AI is a very slow learner still. The base OS for humans is really advanced with hormones biases built in and a initial structure connected to input and outputs.
Sure, it’s possible but we’re not there yet. It could be still 10-100 years until we manage to get a good one, depending on how we don’t know yet.
It’s not a terrible idea by any means. It’s pretty hard to do, though. Check out the Blue Brain Project. https://en.wikipedia.org/wiki/Blue_Brain_Project?wprov=sfla1
ETA: not to mention the brain is a heck of a lot more than a collection of neurons. Other commenters pointed out how we just discovered a new kind of brain cell - the brain is filled with so many different types of neurons (e.g. pyramidal, Purkinje, dopamine-based, myelinated, unmyelinated, internet Ron’s, etc.). Then there’s an entire class of “neuron support” cells called neuralgia. This includes oligodendrocytes (and Schwann cells), microglia, satellite cells, and most importantly, astrocytes. These star-shaped cells can have a huge impact on how neurons communicate by uptaking neurotransmitters and other mechanisms.
Here’s more info: https://en.wikipedia.org/wiki/Tripartite_synapse?wprov=sfla1
Learning models operate like neurons in that they make connections based on experiences (data). But that’s like saying a microwave works like a chef in that it heats up food. We can’t build a microwave that can run a kitchen, design a menu, take a bump in the walk-in, and fire off dishes the way a chef will.
Creating an accurate neuron simulation would probably require much more advanced AI than we already have. Like, real AI, not this piddly, piecemeal shit we have now.
You’re looking at this backwards. We’d need better AI to even start trying to simulate neurons accurately. They’re far more complex.
Currently, AI is capable of analyzing basic chemical and cellular interactions. It’s ok at it.
Actually, we’ve got some pretty sophisticated models of neurons. https://en.wikipedia.org/wiki/Blue_Brain_Project?wprov=sfla1
See my other comment for an example of how little we truly understand about neurons.
Modeling neurons and simulating them with AI are very different things. And, as you say, we still know very little about neurons and the nervous system and the brain itself. How, then, could we even attempt to train an AI to work accurately?
