Asklemmy

I've trained mine to emulate a LLM. So far the hallucination feature works perfectly. Basic grammar still lacks a bit.

The idea that LLMs are just like how the brain works, except limited by running in a CPU, comes from software engineers - not neuroscientists.

Although there are many analogies that could be made between how CPUs do work and how the brain integrates information, they're actually fundamentally different and use completely different logic.

You could, theoretically, create a computing language to work using neurons. And therefore you could also train machine learning algorithms. But that's like using calculators to sum 2+2 by buying 4 calculators and putting them all together, rather than actually using what a calculator does to get the result, if you get what I mean.

Afaik, an actual neuron is computationally more powerful than a perceptron, so in theory yeah, for sure.

If you're a subscriber to the Chinese Room thought problem, we are already just a bunch of really good "LLMs".

With the way current LLMs operate? The short answer is no. Most machine learning models can learn the probability distribution by performing backward propagation, which involves "trickling down" errors from the output node all the way back to the input. More specifically, the computer calculates the derivatives of each layer and uses that to slowly nudge the model towards the correct answer by updating the values in each neural layer. Of course, things like the attention mechanism resemble the way humans pay attention, but the underlying processes are vastly different.

In the brain, things don't really work like that. Neurons don't perform backpropagation, and, if I remember correctly, instead build proteins to improve the conductivity along the axons. This allows us to improve connectivity in a neuron the more current passes through it. Similarly, when multiple neurons in a close region fire together, they sort of wire together. New connections between neurons can appear from this process, which neuroscientists refer to as neuroplasticity.

When it comes to the Doom example you've given, that approach relies on the fact that you can encode the visual information to signals. It is a reinforcement learning problem where the action space is small, and the reward function is pretty straight forward. When it comes to LLMs, the usual vocabulary size of the more popular models is between 30-60k tokens (these are small parts of a word, for example "#ing" in "writing"). That means, you would need a way to encode the input of each to feed to the biological neural net, and unless you encode it as a phonetic representation of the word, you're going to need a lot of neurons to mimic the behaviour of the computer-version of LLMs, which is not really feasible. Oh, and let's not forget that you would need to formalize the output of the network and find a way to measure that! How would we know which neuron produces the output for a specific part of a sentence?

We humans are capable of learning language, mainly due to this skill being encoded in our DNA. It is a very complex problem that requires the interaction between multiple specialized areas: e.g. Broca's (for speech), Wernicke's (understanding and producing language), certain bits in the lower temporal cortex that handle categorization of words and other tasks, plus a way to encode memories using the hippocampus. The body generates these areas using the genetic code, which has been iteratively improved over many millennia. If you dive really deep into this subject, you'll start seeing some scientists that argue that consciousness is not really a thing and that we are a product of our genes and the surrounding environment, that we act in predefined ways.

Therefore, you wouldn't be able to call a small neuron array conscious. It only elicits a simple chemical process, which appears when you supply enough current for a few neurons to reach the threshold potential of -55 mV. To have things like emotion, body autonomy and many other things that one would think of when talking about consciousness, you would need a lot more components.

Go home, Elon, you're drunk.

The concept of ML comes from neurons/the brain. If we could use the neurons we'd be way ahead, and that's basically the hard part. If it will ever be feasible I don't know.

Brains have a lot more connections and meaningful ways of communicating compared to our silly signals and weights. This may be the barrier to AGI

You could put neurons in a box and wire it up, and implant a partial personality into it and call it a Magi

Cortical Labs certainly hope so: https://wired.me/science/this-startup-grows-brain-cells-on-ai-chips/

But outside of the context of computing on devices: yes, as others have noted, the neurons we're trying to simulate in machine learning models aren't much different than our own. So, just look at any person to see how well neurons are suited to language/etc. workloads (or not, depending how clever the people around you are 😂)

As to ethics, consciousness is an "emergent phenomenon". It seems to arise, near as we can tell, from the interaction of many simple systems. No single cell or cluster thereof in a brain is conscious, but get them all working nearby one another and suddenly... 🎇

Ethically at this point is this neuron array considered conscious in any way?

It’s really a matter of taste, as in how do they taste?

Our current ML Neural Networks work (simplified) like this: A neuron emits a number and the next neuron calculates a new number to emit based on all the values given to it by other neurons as inputs. Our brain can't fire numbers in this way. So there's a fundamental difference. Bridging this difference to create NNs that are more similar to our brains is the basis of the study of Spiking Neural Networks. Their performance so far isn't great, but it's an interesting topic of research.

Calling Cordwainer Smith...

Neurons can't NaN so it would be a very bad use of the technology.

Honestly I've wondered this about shining a laser through some kind of laser-etched glass. Only problem is, I have no idea how to represent something like an activation function using only reflection and such.