You shouldn't assume that because you shouldn't assume the converse either.

The LLM will use the chain of thought, as it's being built out like a garden path, to explore different completion possibilities. It's not necessarily logically related what it completes afterwards, but it's definitely influenced.

The search may not be valuable as an artifact in itself. It's likely to be logically unsound especially in parts. But the end result may be useful.

What’s the difference?

At least to me, the difference is that one is ostensibly an explanation of how the AI arrived at the joke, the other is a post-hoc explanation of the joke.

You can be pretty sure the AI isn't doing a post-hoc explanation because the only writable memory it has access to is the tokens it has already output - i.e. the explanation of the joke. Everything else is reset between every token.

As long as it comes up with different jokes different times you ask it (assuming randomness in sampling) - how could it.

The problem is it can’t remember what it hasn’t written but the end result still makes sense, so there has to be some goal after parsing the initial context that the tokens are emitted towards to. This means there’s nothing stopping it from producing an explanation, it might be in there from the very start.

There's no goal. The tokens are computed one at a time from all the previous tokens.

One is orthogonal to the other.

Not one goal though, but a set of goals. Otherwise the random sampling of tokens wouldn't result in it outputting meaningfully different jokes.

I also think it's safe to assume that the set of goals isn't fully resolved, but it's harder to "prove" that.

Given how it is trained specifically (they didn't encourage it to think, they allowed it to) there was a lot of emergent behavior as it trained.

Sort of like chess engines rediscovering classic (named) chess openings. See section 2.2.3 for the training template (it's a single paragraph I can't reproduce here because I'm on my phone)

Example emergent behavior (section 2.2.4 page 8): the model learns to solve more complex problems by spending more time reasoning. It also naturally develops reflection (what have I tried?) and exploration strategies.

Fundamentally, you should think of this as a nn that learned to solve real problems by reasoning about them in written language.

(My favorite part: it defaulted to reasoning in multiple languages. They constrained it to only reason in a single language and this negatively impacted performance! But the hypothesis is that it improves interpretability)

Ever read philosophy? An acquaintance can and will readily mix and match 3 languages to obtain more precision.