It's nice to see someone else that recognizes that the Emperor is naked. I find it somewhat disturbing how many people seem to fall for the Eliza-effect illusion that LLMs are thinking.
The idea that LLMs are useless "bullshit machines" is very 2022. We live in the world today of 2026 where Donald Knuth and Terrance Tao, neither of whom have any patience for hype, use LLMs to help them craft mathematical proofs, I get not liking AI, and getting more satisfaction by doing things oneself. I get frustration on how the AI boom has caused the RAM and graphics cards we want to buy to become unavailable/unaffordable. I get concerns over how such technology is being used by the police and military. But in 2026 the attitude that they are useless "bullshit machines" is as absurd as the articles in the early 2000s that still claimed the Web was just a fad that could be safely ignored.
> One computer scientist speculated that his LLM had attained sentience.
> How did he reach that conclusion? Basically, he asked “Are you conscious?”, the machine responded “Yes”, and that was that.
Oh, come on now. This is referring to Blake Lemoine, and while I doubt his conclusions, he wasn't being as simplistic as all that. He's not completely stupid.
> LLMs operate in the plane of words, not in the world of physical phenomena that science investigates. They don’t reason, synthesize evidence, or draw upon the previous literature. They can generate text that looks like a paper but mistaking this for science is a cargo-cult fallacy.
I’m genuinely interested in someone countering the following evidence that supports the authors.
Plane of words: broadly correct. Everything is flattened to tokens and token sequences, and the training data is dominated by text tokens.
Reasoning: CoT tokens are mostly just tokens, more appropriately called intermediate tokens, and are largely disconnected from the end result. Including them improves the end result (user satisfaction), but does not imply reasoning. See for example Turpin 2023, Mirzadeh 2024, Pournemat 2025, Palod 2025.
Synthesising evidence: You can achieve SOTA summaries with LLMs, but this involves, for example, using a harness to generate dozens of summaries with different models, separately using some kind of vector embedding model to compare results to the original, and selecting the best match. This is not how most people are using LLMs for summaries. While this is being slowly RLVR’d in post-training, a one-shot naive summary underperforms more complex methods significantly.
I think I know the examples you’re talking about. They don’t show much in terms of reasoning.
The Erdős problems have turned out to be largely brute force or finding older results.
The Feb 2026 GPT-5.2 theoretical physics paper was a result of “dialogue between physicists and LLMs”, called “grad student level” by experts in the field, used a “custom harnessed” “internal OpenAI” model with “20 hours of reasoning”. Quotes from OpenAI blog.
The Matthew Schwartz physics paper with Claude this March involved “51,248 messages across 270 sessions, producing over 110 draft versions and consuming 36 million tokens”, and the actual contribution was Schwartz finding an error in Claude’s solution.
It's nice to see someone else that recognizes that the Emperor is naked. I find it somewhat disturbing how many people seem to fall for the Eliza-effect illusion that LLMs are thinking.
horrible design on the website please just give me a block of text to read
What an obnoxious website. I'm not clicking through your silly javascript animated slideshow one sentence at a time just to read an article.
That's surely a professional, non-sensationalist, title that's appropriate for university professors.
Was this written in 2022?
Looks like early 2025. Anything in it in particular that you see as comically out of date?
The idea that LLMs are useless "bullshit machines" is very 2022. We live in the world today of 2026 where Donald Knuth and Terrance Tao, neither of whom have any patience for hype, use LLMs to help them craft mathematical proofs, I get not liking AI, and getting more satisfaction by doing things oneself. I get frustration on how the AI boom has caused the RAM and graphics cards we want to buy to become unavailable/unaffordable. I get concerns over how such technology is being used by the police and military. But in 2026 the attitude that they are useless "bullshit machines" is as absurd as the articles in the early 2000s that still claimed the Web was just a fad that could be safely ignored.
Previously:
Feb 2025 https://news.ycombinator.com/item?id=42989320
From the same authors of "Calling Bullshit: The Art of Skepticism in a Data-Driven World"[0], Carl Bergstrom and Jevin West.
[0]https://callingbullshit.org/
> One computer scientist speculated that his LLM had attained sentience.
> How did he reach that conclusion? Basically, he asked “Are you conscious?”, the machine responded “Yes”, and that was that.
Oh, come on now. This is referring to Blake Lemoine, and while I doubt his conclusions, he wasn't being as simplistic as all that. He's not completely stupid.
Clearly just an anti AI rant packaged up in a fancy suit.
This website is not worth your time.
> LLMs operate in the plane of words, not in the world of physical phenomena that science investigates. They don’t reason, synthesize evidence, or draw upon the previous literature. They can generate text that looks like a paper but mistaking this for science is a cargo-cult fallacy.
This is clearly wrong
I’m genuinely interested in someone countering the following evidence that supports the authors.
Plane of words: broadly correct. Everything is flattened to tokens and token sequences, and the training data is dominated by text tokens.
Reasoning: CoT tokens are mostly just tokens, more appropriately called intermediate tokens, and are largely disconnected from the end result. Including them improves the end result (user satisfaction), but does not imply reasoning. See for example Turpin 2023, Mirzadeh 2024, Pournemat 2025, Palod 2025.
Synthesising evidence: You can achieve SOTA summaries with LLMs, but this involves, for example, using a harness to generate dozens of summaries with different models, separately using some kind of vector embedding model to compare results to the original, and selecting the best match. This is not how most people are using LLMs for summaries. While this is being slowly RLVR’d in post-training, a one-shot naive summary underperforms more complex methods significantly.
What? Reasoning models are inventing proofs for unsolved open problems in mathematics. That is my benchmark for reasoning.
I think I know the examples you’re talking about. They don’t show much in terms of reasoning.
The Erdős problems have turned out to be largely brute force or finding older results.
The Feb 2026 GPT-5.2 theoretical physics paper was a result of “dialogue between physicists and LLMs”, called “grad student level” by experts in the field, used a “custom harnessed” “internal OpenAI” model with “20 hours of reasoning”. Quotes from OpenAI blog.
The Matthew Schwartz physics paper with Claude this March involved “51,248 messages across 270 sessions, producing over 110 draft versions and consuming 36 million tokens”, and the actual contribution was Schwartz finding an error in Claude’s solution.
It is clearly correct, but it sometimes works anyway. Which is the thing that one needs to accept even as not a fan of LLMs.
A 10,000 word website about bullshit machines.
See what they did there ?