I think a hidden tendency towards a belief in Libertarian free will is at the root of people’s opinion that LLMs aren’t capable of reasoning.

I think it’s an emotional and unconscious argument that humans are special, and that by extension—LLMs cannot possibly be doing anything like we are doing.

But if you remember that humans don’t have free will, and that all of our outcomes are either determined or random, it allows us to see LLMs more like us. Which is to say—imperfect but awesome. And then we can switch to speaking purely in terms of capabilities.

So let us say that we’re both deterministic. Or at least mechanistic and practically deterministic because any quantum randomness collapses to deterministic at large scales.

In this model both humans and LLMs are just processors. We're computational devices. We take in inputs, and based on our current state and the state of the environment and the input, we output something.

Cool. So what’s the real question we’re then asking when we ask if LLMs can reason?

First let's remember something. We’re not taking back the human ability to reason just because we are processors, right? No. Let’s not do that. We're still awesome even if we're mechanistic.

In other words, let’s say for the purpose of this that reasoning is consistent with mechanistic/deterministic processing.

Now, let’s find a good definition. Here are some from Merriam-Webster.

Ok, so if we take these all the way down to the base and build back up:

So,

Seems pretty good. And then you have a more common definition based on practicality which is something like:

Regardless of which way we go, we have a couple key sticking points. And they're very tied to my main argument here.

First, the words "think" and "understand"—I would argue—are very much tied to consciousness and Libertarian Free Will. I see these as armaments that LLM-Reasoning skeptics would use to show why LLMs can't be reasoning.

I see them saying something like:

Sounds compelling, but if you break it apart I would argue they're unconsciously binding and confusing experience and understanding vs. actual processing.

In other words, I think they're saying that the thinking and understanding parts are key. As in the human experience of understanding and pondering. They're smuggling these in as essential, when I think they're just red herrings.

Same with "grappling" and "intuition". If we don't have free will, these are all just states of the processing mind that are happening, and our subjective experiences are then being presented with those phenomenon and we're ascribing agency to them.

That's thinking. That's intuition. That's experience. And I think understanding is the same. It's an experience of seeing mappings between concepts and ideas. But in my model the mapping can exist without that subjective experience.

So, I say we take those distractions out of the equation and see what we have left. And what we have left is drawing conclusions, solving problems, and making decisions based on our current model of the world.

The model of the world is the weights that make up the LLM, combined with the context given to it at inference. So it seems to me like we're left with a much simpler question.

Can LLMs draw conclusions, solve problems, and make decisions based on their current model of the world?

I don't see how anyone would say no to that.

Are they perfect? No. Are they conscious? No. Are they "thinking"? I think "thinking" smuggles in subjective experience, so no. But again—those are distractions.

The question is whether LLMs can do this very practical thing that matters in the world, which is drawing conclusions, solving problems, and making decisions.

I think the answer is overwhelmingly and obviously, yes.

As a quick set of examples, we're already using them to:

And a thousand more that we're already familiar with.

Some might say they're not doing "real" things, but just pattern matching and autocompletion.

That's the whole point of what we've been talking about here. That's the whole reason we've explored the argument in this way. We live in a human world where humans have problems and need to solve them.

That’s what logic and reasoning are for.

So what if it's just pattern matching? So what if it's just input + current_state = output. Are humans really all that different? Are we not just as surprised when inspiration—or the very next thought—pops into our minds?

Either way it's a black box information processor with physical limitations.

I think what matters is capabilities. And where capabilities are concerned, LLMs seem remarkably similar and catching up every day.