Comment by mythrwy | Hacker News Reader

Comment by mythrwy | Hacker News Reader

Parent Full thread

Perhaps if you use a very restricted definition of symbols. If you consider symbols to be "anything that represents something" (which is the the sense I use the word in) it is fully the case.

That you might not define floating point numbers to be "symbols" aside, the inputs and the outputs are symbols and the intent and purpose of the creation is strictly symbolic.

It's right there in the name "Large Language Models". Language. Not direct experience, not emotion, not anything else. Language. i.e. symbolic representation.

This does not cover the full spectrum of intelligence humans have, and it shows. And yes, the model can spin up Python parse the output and get mathematical intelligence but there is still a big gap.

As I say, I see the holes. I'm just trying to figure out what it is I see and how to describe it. It's particularly difficult because we don't fully understand how human thinking works but I will say I believe human thinking is a lot more than informal statistical correlation.

The latest LLMs (except Qwen and DeepSeek) are MLLMs (multimodal language models). Unless you count RGB values as symbols, they are dealing with more than symbols.

Yes, there are functional gaps between MLLMs and humans. Their long-term memory is an external mechanism that can use RAG-like approaches, context compression or something like that. The models have problems managing those.

The models can't do continual learning. Although there are promising directions (expert cloning in MoE models, and others).

The only mode of learning available to a model while working on a task is in-context learning. This limits the models to concepts that they developed during autoregressive pretraining and the later stages of training. That is a model can't create new concepts as a result of working on a task (the model's maintainers could choose the task to be represented in the training data later though).

But it's all about functionality.

I guess you have the Leibniz's mill intuition. We can look at how those things work, and there are no experiences or intelligence in sight.

It could be the mill intuition, but my thought is nothing along the lines of "computers can't have souls!" or the human mind is supernatural or anything of the sort.

It's gaps in actual thinking or intelligence I notice. A diff between what I can see or understand and what the model sees or understands. Some are very big, and this in spite of the models having much more knowledge and (presumably) less error prone processing.

My thought is that part of it has to do with inherent limitations of using symbolic representation for "thinking" and I suppose humans have other forms of thinking that occur outside of symbolic representation, and that is going to be hard to recreate digitally.

This is my whole point and I'm not trying to win a debate here or prove "LLMs are useless". Just speculating.

Reply on news.ycombinator.com