# LLMs can't reason

Source: https://www.youtube.com/watch?v=VJUK_NIma6Q
Recap page: https://rapidrecap.app/video/VJUK_NIma6Q
Generated: 2025-11-02T09:01:50.869+00:00

---
## Quick Overview

The assertion that Large Language Models (LLMs) cannot reason is flawed because the tests used to disqualify them, such as the 4-minute mile or the ability to run Doom on a microwave, rely on physical capabilities or subjective, emotional responses that are irrelevant to logical reasoning demonstrated by LLMs.

**Key Points:**
- The speaker refutes the claim that AI cannot reason by challenging common arguments used by skeptics, such as citing physical tasks like running a 4-minute mile or running Doom on a microwave.
- The speaker argues that these challenges are irrelevant tests for reasoning, as they test physical capability or the ability to interact with the physical world, not logical inference.
- The speaker points out the bias in human responses regarding AI-generated content, referencing a study where human-written poetry was preferred over AI poetry until the source was revealed, causing ratings to drop.
- The speaker introduces a diagram showing that human creativity/art is characterized by high craft and low reliance on AI art, while AI output is often high in AI Art and low in craft (according to some), suggesting a false dichotomy.
- The speaker cites Ilya Sutskever's tweet that valuing intelligence above all else leads to a bad time, suggesting that human bias and emotion (like fear or pride) influence the debate against AI reasoning.
- The speaker concludes that the ability to think/reason is defined by logic (as per dictionary definitions), which LLMs can utilize, unlike subjective experiences or physical embodiment.
- The speaker suggests that the common critique that AI lacks 'soul,' 'depth,' or 'presence' is an emotional, rather than logical, argument against the capability of LLMs to perform complex reasoning tasks.

![Screenshot at 00:04: The speaker introduces the central conflict by displaying a tweet predicting that critics will fail to propose an actual test for AI reasoning, instead falling back on vague dismissals like "AI can't reason because it's just a...".](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-00-04.png)

**Context:** The video addresses the ongoing debate in AI circles regarding whether Large Language Models (LLMs) truly possess reasoning capabilities, contrasting this debate with common, often flawed, analogies and evidence presented on social media. The speaker uses specific examples from Twitter discussions, including a tweet by Ilya Sutskever and a study on poetry preference, to frame the argument that skepticism about LLM reasoning is often rooted in emotional bias or irrelevant comparisons to physical tasks.

## Detailed Analysis

The speaker argues against the notion that LLMs cannot reason by deconstructing the arguments used by skeptics, many of which rely on irrelevant comparisons to human physical capabilities or subjective experiences. The speaker first addresses the prediction that no one can propose a valid test for AI reasoning, highlighting Jeff Ladish's challenge for a concrete test, such as whether an AI can reason, understand, or think like a human. The speaker then demonstrates why examples like running Doom on a microwave or a human running a 4-minute mile are poor tests for reasoning, as they test physical function, not logic. The speaker then shifts to a discussion about art and creativity, referencing social media reactions where AI-generated content (like John Wick animation or music) is initially favored, but human bias causes ratings to drop once the AI origin is revealed. This leads to a diagram contrasting AI Art (low craft) versus physical Sculpting (high craft), suggesting that people value the 'craft' aspect of human creation. Finally, the speaker addresses Ilya Sutskever's tweet about valuing intelligence too highly, suggesting that emotional reactions and fear drive resistance to accepting AI reasoning. The speaker concludes by showing the dictionary definition of 'reason' (using logic to form judgments) and asserting that LLMs can perform this, while they inherently lack subjective experience, soul, or presence, which are often cited as reasons for their supposed reasoning failure.

### The Reasoning Debate

- AI Can't Reason vs. AI Can Reason
- Wes Roth challenges the premise that LLMs cannot reason by examining common skeptical arguments
- The speaker uses the failure of LLMs to perform physical tasks (like running a 4-minute mile or running Doom) as an example of an irrelevant test against reasoning capability.

### Human Bias in Artistic Judgment

- AI-generated content vs. Human Creation
- The speaker cites social media examples where AI-generated art (John Wick animation) and music are initially praised but then criticized when the AI origin is revealed, demonstrating a negative bias against AI creations.

### The Craft vs. AI Art Triangle | Sculpting is placed high on the craft scale, while AI Art is low, suggesting that the emotional attachment to human effort ('craft') drives much of the skepticism.


### Defining Reason

- Logic as the Core Requirement
- The speaker examines the definition of 'reason' (power of the mind to think, understand, and form judgments by a process of logic) and argues that LLMs can fulfill this criterion through complex pattern matching and logical processing, even if they lack subjective experience.

### The 'Special Sauce' Argument

- Missing Human Qualities
- The speaker lists qualities attributed to humans that AI supposedly lacks: soul, depth, presence, grit, aura, and panache, concluding that these are emotional arguments rather than logical refutations of AI's reasoning ability.

![Screenshot at 00:04: The speaker displays a tweet predicting that critics will fail to propose an actual test for AI reasoning, instead falling back on vague dismissals like "AI can't reason because it's just a...".](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-00-04.png)
![Screenshot at 00:10: The speaker discusses Jeff Ladish's challenge on Twitter: "For people who think that AIs aren't really reasoning, don't really understand things, or can't really think, what do you mean?"](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-00-10.png)
![Screenshot at 00:44: The screen displays a thumbnail for a video featuring Scott Aaronson discussing 'JustAism' and AGI, referencing experts in the field.](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-00-44.png)
![Screenshot at 02:09: The speaker draws a simple diagram on the blackboard to illustrate the 'Can LLMs Reason?' question, showing a human figure with a checkmark and a rock with an X mark.](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-02-09.png)
![Screenshot at 03:41: The speaker transitions to a more difficult test case: 'TELL TIME?', contrasting it with the simpler rock example.](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-03-41.png)
![Screenshot at 04:42: An image showing the disassembled components of a digital watch, used to argue that a mere collection of parts \(like a clock or digital watch\) cannot inherently perform a complex function like telling time without the underlying mechanism.](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-04-42.png)
![Screenshot at 06:43: The speaker introduces the Drake meme format to contrast the common dismissal \("LLMs CAN'T REASON"\) with the speaker's preferred stance \("LLM CAN REASON"\).](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-06-43.png)
![Screenshot at 11:11: The screen shows a tweet from Ilya Sutskever stating, "if you value intelligence above all other human qualities, you're gonna have a bad time."](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-11-11.png)
![Screenshot at 15:19: The speaker shows a series of tweets arguing against AI art, exemplified by a response to George Rebeiz's post about AI animation, calling it 'soulless.'](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-15-19.png)
![Screenshot at 27:24: A screenshot of a ChatGPT response admitting, "No. I don't have subjective awareness—I model patterns in data to produce useful answers," highlighting the AI's self-awareness of its mechanism.](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-27-24.png)
![Screenshot at 31:17: A slide appears illustrating the Iceberg Model of writing, showing that AI writing only accesses the 'Text' level, while human writing includes deeper elements like 'Subtext,' 'Author Intention,' 'Contextual Impact,' and 'Dialectical Participation.'](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-31-17.png)
![Screenshot at 36:36: The slide changes to a list of abstract qualities that AI supposedly lacks, such as 'Lived Experience,' 'Presence,' 'Spunk,' 'Grit,' 'Soul,' 'X-Factor,' 'Aura,' 'Vibe,' 'Energy,' and 'Frequency.'](https://ss.rapidrecap.app/screens/VJUK_NIma6Q/00-36-36.png)