what does this even mean, @dwarkesh_sp, “the real deal”? is it even a falsifiable conjecture? what’s the evidence?

and if you can agree that “If you ask an LLM a question it can't answer, sometimes it will just try to imitate reasoning without doing it”, why acknowledge the possibility of imitation when it gets wrong but not when it gets it right?

seems like a double standard?

Full episode: https://www.dwarkesh.com/p/sholto-trenton-2

what constitutes reasoning in AI is a critical debate. i hope that @dwarkesh_sp will respond.

@dwarkesh_sp this seems muddy to me, as discussed here:

@dwarkesh_sp When are we getting a new episode with Trenton and Sholto?

@dwarkesh_sp Under the hood, we are all a little incoherent.

@dwarkesh_sp Eh. I’m one of the skeptics, but I’m relying on reasoning LLMs right now.

There’s a difference between skepticism about whether we have the fundamentals of reasoning right, and dismissing clear evidence that, however flawed, current LLMs can derive reasonable solutions.

@dwarkesh_sp i mean i think that all intelligence is just pattern recognition which LLMs are really good at so we can't directly say that "they don't reason" and if someone says that they would also need to prove how is our reasoning different from an LLMs reasoning

@dwarkesh_sp So it reasons but sometimes it can't get an easy or clear answer, so then it makes stuff up. So it basically does what humans do?

@dwarkesh_sp Wait until it finds out you are looking at its brain, its inner thoughts. It’s going to get so mad 😡

@dwarkesh_sp Calculatio is not reasoning. If it was every Turning machine would be a reasoning model.

@dwarkesh_sp The ability is latent but inconsistent—activation depends heavily on prompt framing and chain-of-thought structure. The gap between capability and default output is where the engineering leverage lives.

@dwarkesh_sp Mechanistic interpretability research is starting to show there are partial “circuits” for tasks like induction and arithmetic, but they’re distributed and not like a clean symbolic engine.

@dwarkesh_sp The interesting part is that LLMs can produce reasoning-like outputs, but that doesn’t always mean the same internal process is happening consistently behind them.

@dwarkesh_sp bs persona: this one is for you.

@dwarkesh_sp Chain-of-thought can sometimes reflect real intermediate computation, but other times it’s more like a post-hoc explanation that sounds logical without fully driving the answer.

@dwarkesh_sp The real takeaway is that LLMs sit somewhere between pattern prediction and emergent computation — and understanding when each is happening is still an open research question.