/Tech1d ago

Dwarkesh Patel and Gary Marcus clash over whether LLMs perform genuine reasoning or merely simulate computation

Marcus argues chain-of-thought outputs mask meaningless internal circuits.

597094556494.7K
Gary Marcus
Dwarkesh Patel#67
Original post
Dwarkesh Patel@dwarkesh_sp#67inTech

Whatever AI sceptics say, LLMs really can reason. They're not just doing an imitation that looks like reasoning, it's the real deal.

But even though they are able to reason, sometimes they won't! If you ask an LLM a question it can't answer, sometimes it will just try to imitate reasoning without doing it.

The chain of thought looks basically indistinguishable from actual reasoning. But under the hood something very different is going on.

@TrentonBricken talked with me about what work on circuits inside LLMs has revealed:

12:01 PM · Jun 9, 2026 · 95K Views
Sentiment

Many users dismissed LLM reasoning and interpretability claims from the Dwarkesh podcast and related research as contradictory nonsense or architecturally impossible, often with insults and accusations of bias.

Pos
0.0%
Neg
100.0%
4 comments with sentiment.
Cluster Engagement
Posts from X
Most Activity
Most Activity
VIEWS18.4KLIKES80RETWEETS6REPLIES13
Gary Marcus@GaryMarcus

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?

Dwarkesh Patel@dwarkesh_sp

Whatever AI sceptics say, LLMs really can reason. They're not just doing an imitation that looks like reasoning, it's the real deal.

But even though they are able to reason, sometimes they won't! If you ask an LLM a question it can't answer, sometimes it will just try to imitate reasoning without doing it.

The chain of thought looks basically indistinguishable from actual reasoning. But under the hood something very different is going on.

@TrentonBricken talked with me about what work on circuits inside LLMs has revealed:

23hViews 18.4KLikes 80Bookmarks 27
BOOKMARKS34
Dwarkesh Patel@dwarkesh_sp

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

Dwarkesh Patel@dwarkesh_sp

Whatever AI sceptics say, LLMs really can reason. They're not just doing an imitation that looks like reasoning, it's the real deal.

But even though they are able to reason, sometimes they won't! If you ask an LLM a question it can't answer, sometimes it will just try to imitate reasoning without doing it.

The chain of thought looks basically indistinguishable from actual reasoning. But under the hood something very different is going on.

@TrentonBricken talked with me about what work on circuits inside LLMs has revealed:

1dViews 11.4KLikes 25Bookmarks 34
Gary Marcus@GaryMarcus

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

Gary Marcus@GaryMarcus

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?

22hViews 5.5KLikes 15Bookmarks 3
Gary Marcus@GaryMarcus

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

Gary Marcus@GaryMarcus

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?

23hViews 2.8KLikes 26Bookmarks 4
Carmine@wavesandgrace

@dwarkesh_sp I'm sorry but all you're doing is rewriting the definition of "reasoning" to include word calculators performing a certain way.

Like, okay, great, I'm glad the word calculator is weighting results correctly. We can call this action "reasoning" within the context of LLMs.

22hViews 476Likes 7
Caleb Scharf🌎@caleb_scharf

@GaryMarcus @dwarkesh_sp The irony of people opining that LLMs "reason" without offering a definition of "reason" on which to base their own reasoning....

22hViews 371Likes 7
AndyXAndersen@AndyXAndersen

@GaryMarcus @dwarkesh_sp The problem of arguing about whether LLM truly reason is that no reasoning is ever perfect. You are constrained by the depth of your knowlege and your ability to predict consequences.

Humans have deeper understanding, but we also do educated guesswork and we also screw up.

22hViews 210Likes 1
The AI Therapist@TheAIShrink

@GaryMarcus @dwarkesh_sp 40% of entry-level coding interviews gone. that's your falsifiable claim. not "is it AGI" (unfalsifiable), but "does it work" (yes). real deal is operational.

23hViews 136
Gary Marcus@GaryMarcus

@caleb_scharf @dwarkesh_sp or criteria on which to differentiate reasoning from imitation

22hViews 313Likes 7
guuber@Guuber42

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

1dViews 350Likes 6
prinz@deredleritt3r

@dwarkesh_sp Tired: Trenton Bricken quotes from a year ago

Wired: Brand-new episode of the annual Dwarkesh podcast with Trenton and Sholto

(And yes, it is absolutely correct that LLMs can reason.)

17hViews 270Likes 6
Joshua Winstead@joshwin41

@dwarkesh_sp LLMs really cannot reason, this isn't AI skepticism, this is reality. When we say the word "reason" we're invoking something tied to a biological process in our brains, and it's not even close to being emulated or simulated mathematically speaking.

22hViews 238Likes 5
Aarav Shirpurkar@aaravshirpurkar

@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

1dViews 603
Ankit Maloo@ankit2119

@dwarkesh_sp Reasoning inside a frame is just search. llms cant change the frame once they commit to it. Here it looks like reasoning, but not the same as connotations associated with human reasoning.

20hViews 135Likes 5
breeeeez@Krubslaw

@wavesandgrace @dwarkesh_sp If we're gonna strip it down to "word calculator", then we can say human reasoning just some electricity firing between neurons no?

22hViews 40Likes 2
Shephard@shephard42

@dwarkesh_sp For a weak definition of "reasoning", then of course, and this isn't really any kind of controversial statement worth belaboring.

For a strong definition of "reasoning", then no, this doesn't definitively resolve the issue.

22hViews 99Likes 1
Adam Zachary Wasserman@adamzwasserman

Interesting discussion, but this over-attributes capabilities to the LLM itself.

What we actually interact with is not a raw pretrained LLM. It is a heavily post-trained system: SFT + RLHF/RLAIF + process supervision + hidden reasoning tokens + scaffolding. The naked base model (pure next-token prediction) does not reliably produce clean chain-of-thought, self-correction, or agentic reasoning. It mostly continues text in plausible ways.

The impressive "reasoning" circuits Trenton studies are real in the weights, but reliably eliciting them in the way described requires the full engineered stack on top. Calling that "LLMs reason" blurs the line. The raw transformer is a powerful pattern-matcher with latent abilities. The polished assistant behavior is a product of deliberate post-training.

Happy to be corrected with evidence from base models (not chat versions).

23hViews 250Likes 4
JohnSalmondOz@OzSalmond

@AndyXAndersen @GaryMarcus @dwarkesh_sp "buildings built of chocolate fudge fall down, but so do buildings built of concrete reinforced with steel". Infantile.

19hViews 6
Jonas Persson@BishopBlougram

@GaryMarcus @dwarkesh_sp What many people do not realize is that there is sizable overlap between pure pattern recognition and deductive, rules-based reasoning, but the domains are not coeval. LLMs can play chess if the moves are reasonable, but play a legal move out of left field, and they cry foul.

23hViews 148
Carmine@wavesandgrace

@Krubslaw @dwarkesh_sp Yes we can strip both down, but even so, now you must recognize the unfathomable difference between matrix multiplications creating words, and biochemical reactions creating novel thoughts.

If LLMs are not even 1% as impressive as a brain, saying that they "reason" is cheap.

21hViews 41Likes 1
Load more posts