# GPT-5: Have We Finally Hit The AI Scaling Wall?

Source: https://www.youtube.com/watch?v=mjB6HDot1Uk
Recap page: https://rapidrecap.app/video/mjB6HDot1Uk
Generated: 2025-08-21T15:33:01.964+00:00

---
## Quick Overview

AI scaling laws have hit a practical wall due to the immense computational power required to eliminate errors, making further significant progress with large language models (LLMs) like GPT-5 unlikely to meet expectations, while true AGI may emerge through embodied learning and interaction with the real world.

**Key Points:**
- GPT-5's release is described as "underwhelming," reigniting the debate on whether AI scaling laws have hit a wall.
- A paper suggests that eliminating errors in LLMs has an "extremely long computational tail," requiring an estimated 10^20 times more computing power for a single order of magnitude reduction in errors, making it practically intractable.
- Another paper found that LLM "chains of thought" reasoning fails to generalize out-of-distribution and that reasoning steps do not align with results, describing LLM reasoning as a "brittle mirage."
- The speaker believes companies focused on LLMs for AGI will "slowly begin to panic" as it's becoming clear this is not the right path.
- Alina is seeking people to train AI systems with human expertise and judgment, offering flexible, remote work at rates up to $150 per hour.
- The speaker argues that AGI will not be achieved through LLMs but through embodied AI that can "probe and test" in real or virtual worlds, citing DeepMind's Genie 3 as a significant advancement.

**Context:** The video critically examines the current state of large language models (LLMs), specifically referencing GPT-5's perceived underperformance. It delves into the debate surrounding AI scaling laws and their ability to drive progress towards artificial general intelligence (AGI). The discussion is informed by recent research papers that challenge the optimistic predictions based on scaling laws, suggesting fundamental limitations in LLM reasoning and error correction.

## Detailed Analysis

The video discusses the perceived underwhelming performance of GPT-5, questioning whether AI scaling laws have reached a limit. It references a paper suggesting that eliminating errors in LLMs has an extremely long computational tail, requiring an estimated 10^20 times more computing power for just one order of magnitude fewer errors, making it practically intractable. This explains the discrepancy between theoretical scaling claims and user experience, as human attention naturally amplifies the error tail, leading to the perception of a wall. Another paper highlights that 'chains of thought' reasoning in LLMs fail to generalize out-of-distribution and that the reasoning steps often do not align with the results, labeling LLM reasoning as a 'brittle mirage' and sophisticated simulation rather than true understanding. This suggests companies investing heavily in LLMs for AGI may face disillusionment. The video also promotes Alina, a company seeking individuals to train next-generation AI systems by providing human expertise and judgment, offering flexible, paid remote work up to $150/hour. The speaker clarifies their stance on AGI, believing it won't be achieved through LLMs but through embodied AI that can interact with and learn from the real world, citing DeepMind's Genie 3 as a step forward. They argue that language is a poor descriptor of nature and true intelligence requires probing and testing in interactive environments.

### AI Scaling Laws Re-evaluated

- failure to account for error reduction computational tail
- intractable computational burden for error weeding
- practical limit perceived as a wall

### LLM Reasoning Limitations

- chains of thought fail to generalize out-of-distribution
- reasoning steps misaligned with results
- LLMs are sophisticated simulators, not principled reasoners

### Future of AGI

- companies betting on LLMs for AGI may panic
- AGI unlikely via LLMs
- embodied AI interacting with the real world is the path to AGI
- DeepMind's Genie 3 is a step forward

### Human-Assisted AI Training

- Alina seeks individuals to train AI with expertise and judgment
- flexible, remote, paid work up to $150/hour