[jhedwards]

Thinking in humans is prior to language. The language apparatus is embedded in a living organism which has a biological state that produces thoughts and feelings, goals and desires. Language is then used to communicate these underlying things, which themselves are not linguistic in nature (though of course the causality is so complex that the may be _influenced_ by language among other things).

[kenjackson]

This is really over indexing on language for LLMs. It’s about taking input and generating output. Humans use different types of senses as their input, LLMs use text.

What makes thinking an interesting form of output is that it processes the input in some non-trivial way to be able to do an assortment of different tasks. But that’s it. There may be other forms of intelligence that have other “senses” who deem our ability to only use physical senses as somehow making us incomplete beings.

[jhedwards]

Sure, but my whole point is that humans are _not_ passive input/output systems, we have an active biological system that uses an input/output system as a tool for coordinating with the environment. Thinking is part of the active system, and serves as an input to the language apparatus, and my point is that there is no corollary for that when talking about LLMs.

[kenjackson]

The environment is a place where inputs exist and where outputs go. Coordination of the environment in real time is something that LLMs don’t do much of today although I’d argue that the web search they know perform is the first step.

[pixl97]

LLMs use tokens. Tokens don't have to be text, hence multimodal AI. Fee free to call them different senses if you want.

[jsdalton]

Agreed. Many animals without language show evidence of thinking (e.g. complex problem solving skills and tool use). Language is clearly an enabler of complex thought in humans but not the entire basis of our intelligence, as it is with LLMs.

[AlecSchueler]

But having language as the basis doesn't mean it isn't intelligence, right? At least I see no argument for that in what's being said. Stability can come from a basis of steel but it can also have a basis of wood.

[jibal]

LLMs have no intelligence or problem solving skills and don't use tools. What they do is statistically pattern match a prompt against a vast set of tokenized utterances by humans, who do have intelligence and complex problem solving skills. If the LLM's training data were the writings of a billion monkeys banging on typewriters, any appearance of intelligence and problem solving skills would disappear.

[hackinthebochs]

Word embeddings are "prior" to an LLMs facility with any given natural language as well. Tokens are not the most basic representational substrate in LLMs, rather it's the word embeddings that capture sub-word information. LLMs are a lot more interesting than people give them credit for.

[bsoles]

> Thinking in humans is prior to language.

I am sure philosophers must have debated this for millennia. But I can't seem to be able to think without an inner voice (language), which makes me think that thinking may not be prior (or without) language. Same thing also happens to me when reading: there is an inner voice going on constantly.

[zeknife]

Thinking is subconscious when working on complex problems. Thinking is symbolic or spatial when working in relevant domains. And in my own experience, I often know what is going to come next in my internal monologues, without having to actually put words to the thoughts. That is, the thinking has already happened and the words are just narration.

[Bootvis]

I too am never surprised by my brains narration but: Maybe the brain tricks you in never being surprised and acting like your thoughts are following a perfectly sensible sequence.

It would be incredibly tedious to be surprised every 5 seconds.

[hackinthebochs]

I never miss a chance to reference this video. A woman vividly describes he experience of not having an inner monologue: https://www.youtube.com/watch?v=u69YSh-cFXY

[cma]

> which themselves are not linguistic in nature (though of course the causality is so complex that the may be _influenced_ by language among other things).

Its possible something like this could be said of the middle transformer layers where it gets more and more abstract, and modern models are multimodal as well through various techniques.