Language is not essential for the cognitive processes that underlie thought

This is an important result.

The actual paper [1] says that functional MRI (which is measuring which parts of the brain are active by sensing blood flow) indicates that different brain hardware is used for non-language and language functions. This has been suspected for years, but now there's an experimental result.

What this tells us for AI is that we need something else besides LLMs. It's not clear what that something else is. But, as the paper mentions, the low-end mammals and the corvids lack language but have some substantial problem-solving capability. That's seen down at squirrel and crow size, where the brains are tiny. So if someone figures out to do this, it will probably take less hardware than an LLM.

This is the next big piece we need for AI. No idea how to do this, but it's the right question to work on.

[1] https://www.nature.com/articles/s41586-024-07522-w.epdf?shar...

When you look at how humans play chess they employ several different cognitive strategies. Memorization, calculation, strategic thinking, heuristics, and learned experience.

When the first chess engines came out they only employed one of these: calculation. It wasn't until relatively recently that we had computer programs that could perform all of them. But it turns out that if you scale that up with enough compute you can achieve superhuman results with calculation alone.

It's not clear to me that LLMs sufficiently scaled won't achieve superhuman performance on general cognitive tasks even if there are things humans do which they can't.

The other thing I'd point out is that all language is essentially synthetic training data. Humans invented language as a way to transfer their internal thought processes to other humans. It makes sense that the process of thinking and the process of translating those thoughts into and out of language would be distinct.

It sounds like you think this research is wrong? (it claims llms can not reason)

https://arstechnica.com/ai/2024/10/llms-cant-perform-genuine...

or do you maybe think no logical reasoning is needed to do everything a human can do? Tho humans seem to be able to do logical reasoning

I’ll pop in with a friendly “that research is definitely wrong”. If they want to prove that LLMs can’t reason, shouldn’t they stringently define that word somewhere in their paper? As it stands, they’re proving something small (some of today’s LLMs have XYZ weaknesses) and claiming something big (humans have an ineffable calculator-soul).

LLMs absolutely 100% can reason, if we take the dictionary definition; it’s trivial to show their ability to answer non-memorized questions, and the only way to do that is some sort of reasoning. I personally don’t think they’re the most efficient tool for deliberative derivation of concepts, but I also think any sort of categorical prohibition is anti-scientific. What is the brain other than a neural network?

Even if we accept the most fringe, anthropocentric theories like Penrose & Hammerhoff’s quantum tubules, that’s just a neural network with fancy weights. How could we possibly hope to forbid digital recreations of our brains from “truly” or “really” mimicking them?

Can they reason, or is the volume of training data sufficient for them to match relationships up to appropriate expressions?

Basically, if humans have had meaningful discussions about it, the product of their reasoning is there for the LLM, right?

Seems to me, the "how many R's are there in the word "strawberry" problem is very suggestive of the idea LLM systems cannot reason. If they could, the question is not difficult.

The fact is humans may never have actually discussed that topic in any meaningful way captured in the training data.

And because of that and how specific the question is, the LLM has no clear relationships to map into a response. It just does best case, whatever the math deemed best.

Seems plausible enough to support the opinion LLM'S cannot reason.

What we do know is LLMs can work with anything expressed in terms of relationships between words.

There is a ton of reasoning templates contained in that data.

Put another way:

Maybe LLM systems are poor at deduction, save for examples contained in the data. But there are a ton of examples!

So this is hard to notice.

Maybe LLM systems are fantastic at inference! And so those many examples get mapped to the prompt at hand very well.

And we do notice that and see it like real thinking, not just some horribly complex surface containing a bazillion relationships...