# LLMs won't get us to AGI. HRM might.

Source: https://www.youtube.com/watch?v=jHGGVMOp1qY
Recap page: https://rapidrecap.app/video/jHGGVMOp1qY
Generated: 2025-08-15T03:02:07.789+00:00

---
## Quick Overview

The video argues that current Large Language Models (LLMs) like GPT-4 are not the right path to Artificial General Intelligence (AGI), suggesting that a more structured, reasoning-based approach might be. The presenter explains that LLMs, while powerful, primarily rely on pattern matching from vast datasets, lacking a true understanding of causality or complex reasoning needed for AGI. The proposed alternative, referred to as HRM (likely Hierarchical Reinforcement Learning or a similar reasoning-focused model), aims to build models that can decompose problems and reason through solutions, similar to how humans learn and solve novel problems.

**Key Points:**
- Current LLMs like GPT-4 are not on the right path to AGI due to their reliance on pattern matching rather than true reasoning.
- AGI requires causal reasoning, planning, and extrapolation, capabilities that LLMs currently lack.
- The presenter advocates for an alternative approach, possibly involving hierarchical reasoning (HRM), which focuses on problem decomposition and step-by-step reasoning.
- Simply scaling up LLMs (e.g., from GPT-3.5 to GPT-4) is insufficient for achieving AGI.
- LLMs excel at interpolation (predicting within known data patterns), while AGI requires extrapolation (solving novel, unseen problems).
- The proposed HRM models are more computationally efficient for complex problems and are designed to 'think' before acting, unlike current LLMs.
- Concerns exist about the validity and verification of some current LLM approaches, as the underlying mathematical principles are not always fully understood or demonstrated.

![Screenshot at 00:01: The presenter, wearing glasses and a green t-shirt, directly addresses the camera, setting the stage for a discussion on AI and AGI.](https://ss.rapidrecap.app/screens/jHGGVMOp1qY/00-00-01.png)

**Context:** The video discusses the current state of Artificial Intelligence, specifically focusing on Large Language Models (LLMs) and their potential to achieve Artificial General Intelligence (AGI). The presenter, a non-expert in mathematics but with an understanding of AI implementation, aims to explain why current LLM architectures might not be the correct path towards AGI and proposes an alternative approach.

## Detailed Analysis

The video contends that current LLMs, despite their impressive capabilities in tasks like text generation and translation, are fundamentally limited in their ability to achieve Artificial General Intelligence (AGI). The presenter explains that LLMs operate by identifying patterns in massive datasets, which allows them to predict the next word or token in a sequence. However, this pattern-matching approach does not equate to genuine understanding, reasoning, or common sense. The video highlights that LLMs struggle with tasks requiring deep causal reasoning, planning, and adapting to novel situations, which are considered crucial for AGI. The presenter proposes that a different approach, potentially involving structured reasoning and hierarchical learning (referred to as HRM), is more likely to lead to AGI. This alternative model would focus on decomposing problems into smaller, manageable parts and using reasoning mechanisms to solve them, mimicking human cognitive processes more closely. The video criticizes the current trend of simply scaling up LLMs (e.g., from GPT-3.5 to GPT-4) as insufficient for true AGI, suggesting that architectural changes and a focus on reasoning are more critical. The presenter implies that while LLMs are good at interpolation (filling gaps within known data), AGI requires extrapolation (solving novel problems), which necessitates a different kind of intelligence architecture.

### Core Argument

- LLMs are pattern matchers, not true reasoners, and thus not on the path to AGI
- AGI requires causal reasoning, planning, and extrapolation, which LLMs lack

### Limitations of LLMs

- Struggle with novel problems, lack of true understanding, reliance on massive datasets for pattern matching

### Proposed Alternative (HRM)

- Focus on problem decomposition, reasoning mechanisms, mimicking human cognition, likely involving hierarchical learning

### Critique of Scaling

- Simply increasing LLM size (e.g., GPT-3.5 to GPT-4) is insufficient for AGI

### Key Distinction

- LLMs excel at interpolation (within known data), AGI needs extrapolation (solving novel problems)

![Screenshot at 00:01: A person wearing glasses and a green shirt, speaking directly to the camera.](https://ss.rapidrecap.app/screens/jHGGVMOp1qY/00-00-01.png)
![Screenshot at 00:15: The person gestures with their hands, emphasizing a point about the limitations of current LLMs.](https://ss.rapidrecap.app/screens/jHGGVMOp1qY/00-00-15.png)
![Screenshot at 00:40: Close-up of the person's face as they explain the concept of reasoning in AI models.](https://ss.rapidrecap.app/screens/jHGGVMOp1qY/00-00-40.png)
![Screenshot at 01:15: The person gestures to illustrate the difference between LLMs and reasoning-based models.](https://ss.rapidrecap.app/screens/jHGGVMOp1qY/00-01-15.png)
![Screenshot at 01:50: The person uses hand gestures to represent problem decomposition and step-by-step solutions.](https://ss.rapidrecap.app/screens/jHGGVMOp1qY/00-01-50.png)
![Screenshot at 02:30: The person explains the architecture of proposed AI models.](https://ss.rapidrecap.app/screens/jHGGVMOp1qY/00-02-30.png)
![Screenshot at 03:15: The person touches their chin thoughtfully while discussing the challenges of achieving AGI.](https://ss.rapidrecap.app/screens/jHGGVMOp1qY/00-03-15.png)
![Screenshot at 04:00: The person holds up their hands, palms facing outward, to illustrate a concept.](https://ss.rapidrecap.app/screens/jHGGVMOp1qY/00-04-00.png)
![Screenshot at 04:40: The person gestures with their hands to emphasize the continuous nature of AI learning processes.](https://ss.rapidrecap.app/screens/jHGGVMOp1qY/00-04-40.png)
![Screenshot at 05:15: The person makes a circular motion with their fingers, possibly representing iterative learning or refinement.](https://ss.rapidrecap.app/screens/jHGGVMOp1qY/00-05-15.png)