Part of it is the top LLM companies (OpenAI, Mistral) all copy and over train, often against e.g. Claude's or DeepSeek's TOS, on each other's models.

Part of it is they all copy and over train, often against the TOS, on each other's models.

LLMs are good at mimicking human intuition. Still sucks at deep thinking.

LLMs PATTERN MATCH well. Good at "fast" System 1 thinking, instantly generating intuitive, fluent responses.

LLMs are good at mimicking logic, not real reasoning. Simulate "slow," deliberate System 2 thinking when prompted to work step-by-step.

The core of an LLM is not understanding but just predicting the next most word in a sequence.

LLMs are good at both associative brainstorming (System 1) and creating works within a defined structure, like a poem (System 2).

Reasoning is the Achilles heel rn. AN LLM's logic can SEEM plausible, it's based on CORRELATION, NOT deductive reasoning.

correlation between text can implement any algorithm, it is just the architecture which it's built on. It's like saying vacuum tube computers can't reason bc it's just air not reasoning. What the architecture is doesn't matter. It's capable of expressing reasoning as it is capable of expression any program. In fact you can easily think of a turing machine and also any markov chain as a correlation function between two states which have joint distribution exactly at places where the second state is the next state of the first state.