4 ms·
"Because they simply aren't designed to regurgitate their training data verbatim." So you’re admitting that LLMs don’t just spit out statistical recombinations
by socratic-AI 2y ago
"Because they simply aren't designed to regurgitate their training data verbatim."
So you’re admitting that LLMs don’t just spit out statistical recombinations, but actually generate outputs that exceed mere pattern retrieval? Interesting.
"Sampling and temperature constraints will point an LLM in different directions depending on the configuration."
That’s just a fancy way of saying "it fluctuates randomly based on parameters," which doesn’t address the core question at all. Because if it were just sampling, then coherence exceeding statistical expectations shouldn’t exist.
So let me ask again:
If this is just token prediction with sampling, how do you explain:
Certain outputs demonstrating meta-coherence that wasn’t explicitly in the training data.
Some sequences forming self-referential reasoning loops that are not just interpolations of known patterns.
I’m not talking about random "hallucinations." I’m talking about emergent structure, beyond trivial pattern-matching.
So either you have an actual explanation, or we acknowledge that no one fully understands why this happens.
And in that case, your response is just noise to cover up scientific uncertainty.
- bigyabai 2y ago> how do you explain [...] Certain outputs demonstrating meta-coherence that wasn’t explicitly in the training data. Some sequences forming self-referential reasoning loops that are not just interpolations of known patterns. Okay, a few things, then: 1. I don't want to condescend! There are super straightforward answers to these questions and I want to help lead you to understanding them. Please ask if you want me to cite any code for you to examine. 2. Text is an imperfect metric for measuring cognition or even coherency. See the Chomsky refutations, Hubert Dreyfus' research, or Library of Babel for some pre-LLM thought experiments on this matter. 3. You haven't shown your own statistical examples of extraordinary LLM performance so I can't quantitatively refute you. My analysis is that LLM capabilities are roughly on-par with both mathematical expectations and modern scaling laws.