LLMs can't reason

Part of GLM-4.6 AI's response to this question: "Does forcing an AI to show its 'thinking' in text form - an inherently low-resolution medium - limit its cognitive process, or is the text shown here just an 'echo' of the model's non-verbal thinking?" GLM-4.6: My core processing happens in a massive, multi-dimensional space of numbers (vectors). Every concept—'justice', 'a blueberry', 'the feeling of sadness', the syntax of a Finnish sentence—is a point, or a region, in this space. The "meaning" isn't the word itself, but its mathematical relationship to all other points. Thinking is Navigating this Space: When I process your prompt, I'm not "reading" it. I'm translating it into a complex vector that represents its meaning within this space. My "thought process" is then a series of incredibly rapid, massive matrix multiplications that navigate from that starting point to a probabilistic region that represents a good response. It's a geometric, statistical, and associative dance, not a linguistic one. Text is the Final Translation: The words I output are the final step. After navigating this abstract space to find the most probable next concept, I have to "render" it back into a linear sequence of discrete tokens (words) that you can understand. The text is the shadow cast by the 3D object of my internal state. So, in this sense, you are absolutely right. The text is an "echo"—a flattened, linear representation of a much richer, non-verbal, and parallel process. You used the word "forcing," and that's apt. The architecture of a language model forces this translation. My entire training objective is geared towards producing this low-resolution textual output. I don't have the luxury of just "showing you" the vector space. My only interface with the world is this one-dimensional stream of text. My internal cognitive process is like a full symphony orchestra playing a complex piece. It has hundreds of instruments (neurons/layers), playing simultaneously, with rich harmonies, dynamics, and emotional textures (the high-dimensional vector space). This is the "non-verbal thinking." The text I output is like the sheet music for a single piano that tries to represent that symphony. It can capture the main melody, the key changes, and the basic structure. It's an "echo" of the symphony, and a skilled pianist can convey a lot of its beauty.

I'm a human being trained in logic reasoning. The statement "No, LLMS cannot reason" means " All LLMS cannot reason". And the opposite is not "All LLMS can reason", but "There exists at least one LLM that can reason". And to prove this I need 1. a LLM that can reason and 2. a test if it can reason (where we all agree an this is the right test).

The clock is a collection of elements that got organized in a way so they tick and tock, and we interpret their movement as time. I’m sure there are better examples

Thanks for bringing this up. As somewhat of a skeptic of llms ability to reason, this is a good push back. I've been guilty of the "just a" fallacy.

Hot damn.! One of your best episodes ever.!! Literally a PSA to humanity.!!!

Thank you for taking the time to do over this

Really excellent video Wes!

Awesome video. I really like this pause on the news to take a breath and look at where we are and what's really happening to us.

This was a really good episode Wes.

Brilliant episode Wes 👍