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
Simulating even one neuron is very complex. Neurons in artificial neuron nets used in machine learning are a gross oversimplification. On top on this you need to get the wiring right. On top on this you need to get the sensorial system right (a brain without input is worthless). On top of this you need an environment. So it’s multiple layers of complexity that we don’t have
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).
-
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.
We don’t really understand how real neurons learn.
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.
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.
I mainly disagree with the final statement on the basis that the LLMs are more advanced predictive text algorithms. The way they’ve been set up with a chatbox where you’re interacting directly with something that attempts human-like responses, gives off the misconception that the thing you’re talking to is more intelligent than it actually is. It gives off a strong appearance of intelligence but at the end of the day, it predicts the next word in a sentence based on what was said previously but it doesn’t do that good job of comprehending what exactly it’s telling you. It’s very confident when it gives responses which also means when it’s wrong, it’s very confidently delivering the incorrect response.
Talk to anyone who consumes Fox News daily and you’ll get incorrect predictive text generated quite confidently. You may also deny them their intelligence and lack of humanity with the fallacies they uphold.
I also think intelligence is a gradient—is an ant intelligent? What about a dog? Chimp? Who gets to draw the line?
It very may be a very complex predictive text generator that hallucinates but I’m concerned that it minimizes its capabilities for better or worse—Its ability to maintain context and has enough plasticity to reason and change its response points to something more, even if we’re at an early stage.
What you’re alluding to is the Turing test and it hasn’t been proven that any LLM would pass it. At this moment, there are people who have failed the inverse Turing test, being able to acerrtain whether what they’re speaking to is a machine or human. The latter can be done and has been done by things less complex than LLMs and isn’t proof of an LLMs capabilities over more rudimentary chatbots.
You’re also suggesting that it minimises the complexity of its outputs. My determination is that what we’re getting is the limit of what it can achieve. You’d have to prove that any allusion to higher intelligence can’t be attributed to coercion by the user or it’s just hallucinating based on imitating artificial intelligence from media.
There are elements of the model that are very fascinating like how it organises language into these contextual buckets but this is still a predictive model. Understanding that certain words appear near each other in certain contexts is hardly intelligence, it’s a sophisticated machine learning algorithm.
Neurons undergo physical change in their interconnectivity. New connections (synapses) are created, strengthened, and lost over time. We don’t have circuits that can do that.
Actually, neuron-based machine learning models can handle this. The connections between the fake neurons can be modeled as a “strength”, or the probability that activating neuron A leads to activation of neuron B. Advanced learning models just change the strength of these connections. If the probability is zero, that’s a “lost” connection.
Those models don’t have physical connections between neurons, but mathematical/programmed connections. Those are easy to change.
That’s a vastly simplified model. Real neurons can’t be approximated with a couple of weights - each neuron is at least as complex as a multi-layer RNN.