I get what you mean but it feels very different. A catapult is still intuitive without knowing physics. Discovering geometric shapes inside of LLMs feels different.
I'd argue that the step up isn't that big, actually. Most people working in the field of machine learning already had a geometric understanding of the models. That's why it was not surprising to me at all that the LLMs are operations on shapes.
It's a lot like random discoveries in drugs where there are many candidates and only some meet the criteria then the exact mechanism is discovered later. But ML does this at much larger scale and candidate "drugs" are progressively improved by past attempts.
I think of it more like making aliens. We are building an environment (ie criteria) then stepping away to see what evolves to meet it.
Basically all of crops/farming throughout the ages was genetic/bio-engineering without understanding its mechanics. Similarly with domesticating animals.
It's cool, but very straightforward when you think about it. An embedding, the fundamental unit of analysis for a LLM, is simply a point in a N-dimensional space. Every operation in a LLM's architecture is an operation on a collection of points in N-dimensional space, so of course it's all shapes.
The 3D plot takes point+distance over an aggegated (3D instead of 2048D) space and shows the idea of (but not really representative due to the pre-aggregation) the relations as coordinates.
Did humankind ever before invent a technology and then try to find out how it fundamentally works after?
Maybe a few drugs ... but a piece of engineering? This is like finding unknown alien technology.
Definitely. People were using catapults before they understood Newtonian physics, for example.
I get what you mean but it feels very different. A catapult is still intuitive without knowing physics. Discovering geometric shapes inside of LLMs feels different.
I'd argue that the step up isn't that big, actually. Most people working in the field of machine learning already had a geometric understanding of the models. That's why it was not surprising to me at all that the LLMs are operations on shapes.
It's a lot like random discoveries in drugs where there are many candidates and only some meet the criteria then the exact mechanism is discovered later. But ML does this at much larger scale and candidate "drugs" are progressively improved by past attempts.
I think of it more like making aliens. We are building an environment (ie criteria) then stepping away to see what evolves to meet it.
Basically all of crops/farming throughout the ages was genetic/bio-engineering without understanding its mechanics. Similarly with domesticating animals.
It's cool, but very straightforward when you think about it. An embedding, the fundamental unit of analysis for a LLM, is simply a point in a N-dimensional space. Every operation in a LLM's architecture is an operation on a collection of points in N-dimensional space, so of course it's all shapes.
I'm going to be honest with you: this goes over my head. I'm in awe and understand none of that.
How do they decide the dimensions of this N-dimensional shape to use for the 3d plots in the figures?
Exactly what you said up there, I think.
The 3D plot takes point+distance over an aggegated (3D instead of 2048D) space and shows the idea of (but not really representative due to the pre-aggregation) the relations as coordinates.
https://twiiit.com/GoodfireAI/status/2054962242022777189