LLM reasoning

I’m trying to understand how to use LLMs for what they are best at, right tool for the job. In one of the short courses was mentioned that “LLMs use inner thought process to resolve a query”(paraphrasing), which is not realy accurate and creates confusion by making people believe that LLMs are actually reasoning.
My understanding is that LLMs generate the next best token, statistically inferred based on the context, attention mechanisms, and other statistical patterns it discovered part of training.
Hence, there is no inner thought process, no reasoning. If we deliver this message then we better understand what LLMs are good at and use them accordingly.
I’d love to hear peoples’ thoughts on this, am I right or should I consider the process of generating next token as actual reasoning, which seems to be the message.
Also, by better understanding what LLMs can do we then have a more through understanding on how to build the agents around LLMs to better solve our problems.

1 Like

I agree with you.

“Reasoning” is a very difficult concept to define.

1 Like

Good to hear Tom. I’d love to hear your opinion on how to convey the message that LLMs don’t reason, they are not even good at simple maths if they haven’t seen that sum before. How do we make people understand that everything that an LLM returns needs to be verified somehow and they can’t be left to their own devices, so to speak? I hear a lot of companies everywhere jumping on this prompt engineering madness as if all our problems can be solved with a good RAG and some well crafted prompts

I don’t have any opinion on how to highlight the limitations of AI technology.