LLMs do not have “thoughts” or “intuition” by any definition.
They organise tokens into groups based on statistical probability, and convert those tokens into words and sentences.
They can explain why they resolved a given answer, but he answer is always unsatisfying: this accumulation of tokens seemed to be the most likely response based on my probability matrix.
The problem you’re trying to convey is that humans can not summarise or read or understand a probability matrix with many billions of parameters.
They organise tokens into groups based on statistical probability, and convert those tokens into words and sentences.
Yeah, that’s the way your brain works when you’re not paying attention. That’s why your dreams are like that. It’s just a bunch of vaguely probable stuff based on your daily experiences.
humans can not summarise or read or understand a probability matrix with many billions of parameters.
Uhhhhh, yeah they can. The human brain has 86 billion neurons. If you put all 86 billion of those neurons and their connections into a matrix, you get a mathematical representation of the human brain’s information processing. You’re running the calculations on all 86 billion neuron matrix every second, manually, one neuron at a time.
Like, I’m guessing you’re not really a math person, so let Me explain in simple terms: multiplication is just a shorter way to do addition, right? 6 * 5 is just 6+6+6+6+6, collapsed down. So if you have a 10 year old kid who can count out all the numbers one by one, and a computer that knows some computational tricks to figure out 6 * 5, that’s the same, right?
So your neurons in your brain are biological computers, and that makes them really energy efficient. Electronic computers kinda suck at that efficient stuff. So where your neurons can manually calculate 86 billion parameters many times a second, a computer has to use shortcuts. And vector cross-multiplication is just a more efficient way for a computer to calculate neuron firings. The abstract mathematical operations are the same in principle.
Now there are a lot of differences between your brain and an LLM, and I’ll be happy to explain them. But matrix multiplication is not one of them. That’s stinky nonsense.
That’s like saying a human can’t read the neurons in their own brain
Yeah, that’s what you just said. Matrix multiplication is an abstract representation of neuron weights, remember? And then you said a human can’t read the matrix. Like it was at all relevant. Obviously neither humans nor LLMs can understand the individual neurons that make up their thoughts, nor the matrices that describe those neurons. Obviously.
Are you ready to stop regurgitating canned arguments yet?
There is no “probability matrix”. There’s an input matrix (which you provide it), various neuron weight matrices, and an output matrix (which contains the thing it says in response to your input).
It doesn’t look up a table of things to say like a d100 table in an RPG. You’ve fundamentally misunderstood the technology.
Incorrect.
LLMs do not have “thoughts” or “intuition” by any definition.
They organise tokens into groups based on statistical probability, and convert those tokens into words and sentences.
They can explain why they resolved a given answer, but he answer is always unsatisfying: this accumulation of tokens seemed to be the most likely response based on my probability matrix.
The problem you’re trying to convey is that humans can not summarise or read or understand a probability matrix with many billions of parameters.
Yeah, that’s the way your brain works when you’re not paying attention. That’s why your dreams are like that. It’s just a bunch of vaguely probable stuff based on your daily experiences.
Uhhhhh, yeah they can. The human brain has 86 billion neurons. If you put all 86 billion of those neurons and their connections into a matrix, you get a mathematical representation of the human brain’s information processing. You’re running the calculations on all 86 billion neuron matrix every second, manually, one neuron at a time.
Like, I’m guessing you’re not really a math person, so let Me explain in simple terms: multiplication is just a shorter way to do addition, right? 6 * 5 is just 6+6+6+6+6, collapsed down. So if you have a 10 year old kid who can count out all the numbers one by one, and a computer that knows some computational tricks to figure out 6 * 5, that’s the same, right?
So your neurons in your brain are biological computers, and that makes them really energy efficient. Electronic computers kinda suck at that efficient stuff. So where your neurons can manually calculate 86 billion parameters many times a second, a computer has to use shortcuts. And vector cross-multiplication is just a more efficient way for a computer to calculate neuron firings. The abstract mathematical operations are the same in principle.
Now there are a lot of differences between your brain and an LLM, and I’ll be happy to explain them. But matrix multiplication is not one of them. That’s stinky nonsense.
Humans can not read or understand the probability matrix in which an LLM stores its parameters.
Neither can an LLM. Try asking ChatGPT to run a matrix cross-multiplication, it sucks as badly as if you tried to do it without any paper.
An LLM can’t read the matrix in which it stores its own parameters?
That’s like saying a human can’t read the neurons in their own brain. Sure, looking at them through a microscope, we can not.
You don’t seem to have any idea what you’re talking about.
Yeah, that’s what you just said. Matrix multiplication is an abstract representation of neuron weights, remember? And then you said a human can’t read the matrix. Like it was at all relevant. Obviously neither humans nor LLMs can understand the individual neurons that make up their thoughts, nor the matrices that describe those neurons. Obviously.
Are you ready to stop regurgitating canned arguments yet?
LOL.
We’re discussing whether LLMs can think and have intuition.
My point is, just because we can’t read the probability matrix with which they formulate their answers, does not mean they have intuition.
There is no “probability matrix”. There’s an input matrix (which you provide it), various neuron weight matrices, and an output matrix (which contains the thing it says in response to your input).
It doesn’t look up a table of things to say like a d100 table in an RPG. You’ve fundamentally misunderstood the technology.
It’s not a magic 8 ball.
Oh, silly me. They have intuition and just say the first thing that they think of.