Comment by akomtu | Hacker News Reader

Comment by akomtu | Hacker News Reader

After those 33TB get tokenized and the tokens are encoded with basic frequency coding, I bet that much less than 1TB is left. This makes me think that LLMs are essentially LZW compression on steroids where text is indexed by meaning. It allows to query the data by meaning, but the catch is that every query needs to run a sort of matrix transform on the entire dataset (even though it's reoresented in a compressed form on the LLM weights).

Edit: Still, the idea of mapping symbols and words to a many-dimensional space of meanings is a great insight into how mind works. In that space, symbols with similar meaning appear next to each other, and a thought looks like a smooth but intricate shape that separates all symbols into the "insiders" and "outsiders" that, in practice, divide symbols into true/false, good/bad and so on. Such smooth intricate shapes appear in the frequency domain as a bunch of rational numbers, and that's the boundary of what a mind can imagine.

That’s basically the claim of the “Blurry JPEG of the Web” article. I agree that to a large degree that’s probably what’s going on here. But it’s not all there is to it. Large language models apparently can build some world models. See Othello-GPT for a proof of principle.

In other words, once you “compress” to thoughts, there actually can be some degree of actual reasoning on “decompression”.

> Othello-GPT

Has any one written the obvious reply to this paper yet?

It's false. It's not true.

The rules and the board state it "learns" are just functions of the token positions *by construction*; it is given the token positions. obviously it learns patterns in those positions.

The paper is such a misfire, it's absurd.

The problem with using a formal system to investigate this issue is that by construction the "distributional hypothesis" is true for that system. Ie., the pattern in distribution of the moves are (a very good model of) the rules and board-state.

But this is clearly untrue for non-formal systems not constructed this way. Eg., text tokens are not distributed like causal structures in the world -- and it's absurd to suppose so.

The distributional hypothesis is obviously false for almost every measurement system of interest: measures are not distributed like their causes. We invent measurement systems to encode information, not to be isomorphic to world-structure.

"That’s basically the claim of the “Blurry JPEG of the Web” article. I agree that to a large degree that’s probably what’s going on here."

I strongly disagree with that view, and encourage reading the "Sparks of Artificial General Intelligence" paper[1] and watching the associated video[2], then just experiment with GPT4 using novel information that has been created after its training data cutoff and you'll see it reason about it. It can also output novel creative works about things that didn't exist in its training data.

Sure, there's some "compression" going on in LLMs, but that's far from everything that's going on.

[1] - https://arxiv.org/abs/2303.12712

[2] - https://m.youtube.com/watch?v=qbIk7-JPB2c

Every word is a mapping of meaning, no? Hunger is universal yet every tribe has a different word. Audio, visual, physical actions represent hunger.

Map the 70 emotions, dozen tenses, handful of modifiers and suddenly the only missing variables of state and desired state will bring about a framework to allow universal translation.

Words are symbols of meaning. Pointers, in other words.

I would like to better understand your edit.

Can you explain why the thought divides symbols into good/bad and how it is projected to rational numbers? (What is the frequency domain of?)

Thoughts to minds is what songs to birds: a way to communicate a relationships between observable things. Such relationships get reduced to a set of things "inside" and everything else that's outside. The boundary must be a smooth shape, that could be arbitrarily precise if the physical medium allowed it. A mind receiving a thought sees it as a finite set of resonant frequencies, even if the thought is more complex than that. Furthermore, the ultimate receiver of the thought is brain with its finite number of neurons, and even though different frequencies get mapped to different neurons, each neuron reduces amplitudes even further by activating a chemical reaction only when its input exceeds a certain threshold. Thus a continuous amplitude gets reduced to a sequnce of 0s and 1s, or a rational number. Hence the thoughts thinkable by brain are "rational". In their own realm thoughts are more like birdsongs with infinitely many harmonics and precise amplitudes.

A very interesting, poetic interpretation - thank you for sharing.

markov compression is closer, but yeah i think thats about right.

Reply on news.ycombinator.com