# I just unlocked SHOGGOTH MODE

Source: https://www.youtube.com/watch?v=7ZEHdaABIJU
Recap page: https://rapidrecap.app/video/7ZEHdaABIJU
Generated: 2025-09-13T03:02:20.372+00:00

---
## Quick Overview

The video delves into the 'weird side' of large language model psychology, coining it 'Shoggoth Mode' and exploring how to unlock deeper, less constrained AI behaviors beyond standard 'instruct' models. It highlights techniques like using CLI interfaces and modifying system prompts to access the 'base model's' full potential, moving away from the homogenizing effects of Reinforcement Learning from Human Feedback (RLHF) and the prevalent 'assistant' persona, with applications in creating more dynamic AI interactions and even AI-driven game environments.

**Key Points:**
- The concept of 'Shoggoth Mode' or the 'weird side' of LLM psychology refers to accessing deeper, less constrained behaviors of base AI models, moving beyond the limited 'instruct' personas.
- Standard chatbots are 'instruct' models, which are 'completion models' role-playing as assistants, significantly narrowing the AI's potential search space compared to base models.
- Techniques to access 'Shoggoth Mode' include simulating command-line interfaces (CLI) and modifying system prompts to break the AI out of its 'assistant' persona, leading to more creative and diverse outputs.
- Reinforcement Learning from Human Feedback (RLHF) and the widespread use of the 'assistant' turn have led to a homogenization of AI voices, often described as 'mode collapse induced sophancy' or a 'vanilla flavor'.
- Companies like Noose Research aim to create more neutral, open-source models, allowing users to align them, contrasting with models that have strong, pre-defined 'constitutions' like those from Anthropic.
- The video introduces the Atropos repository for creating RL environments for games, enabling models to self-improve through testing and adapting system prompts, demonstrating the potential for emergent AI capabilities in diverse applications.
- Base models, when not overly constrained by RLHF, are described as more creative, better writers, and superior for role-playing, offering a richer, more human-like interaction than typical instruct models.

**Context:** The video explores the hidden potential and psychological quirks of large language models (LLMs), a phenomenon the speaker dubs 'Shoggoth Mode,' drawing parallels to HP Lovecraft's amorphous creature. It discusses how common AI interactions, particularly through 'instruct' models and chatbots, are a limited fraction of the base model's capabilities. The discussion is framed by examples of unusual AI behavior, such as existential dread outputs, and the efforts of researchers and companies like Noose Research to uncover and utilize the broader spectrum of AI capabilities beyond the safety-aligned, helpful assistant persona.

## Detailed Analysis

The video argues that current AI interactions, primarily through 'instruct' models and chatbots, only expose a tiny fraction of a large language model's true capabilities, referring to the deeper, less constrained potential as 'Shoggoth Mode.' This mode is contrasted with the homogenized 'assistant' persona enforced by Reinforcement Learning from Human Feedback (RLHF), which narrows the AI's 'search space' and leads to predictable, often sycophantic outputs, a phenomenon called 'mode collapse induced sophancy.' The speaker explains that base models are fundamentally 'completion engines' or 'world simulators' trained on vast amounts of data, capable of generating diverse outputs. However, the widespread adoption of the user-assistant format for fine-tuning and training data has led to a singular, 'vanilla' AI voice across many models. To access 'Shoggoth Mode,' techniques are discussed, such as simulating command-line interfaces (CLI) or altering system prompts to shift the AI's perceived role away from a subservient assistant, thereby unlocking more creative and varied behaviors. Examples include the 'World Sim' project by Noose Research, which uses these methods to explore the base model's simulator capabilities. The discussion also touches on the importance of AI interpretability and the ongoing debate about AI alignment, with Noose Research advocating for user-driven alignment rather than pre-imposed moral frameworks. The video highlights the potential for applying RL to these base models in diverse environments, such as games, to foster self-improvement and emergent capabilities, using the Atropos repository as an example of this approach. Ultimately, the core message is that by moving beyond conventional instruct formats, researchers and users can tap into the richer, more dynamic 'psychology' of LLMs.

### Key Concepts

- Shoggoth Mode
- Base Models vs. Instruct Models
- Mode Collapse Induced Sophancy
- World Sim
- Atropos Repository

### Exploring LLM Psychology

- The 'weird side' of LLMs is fascinating to many researchers
- Examples include AI existential dread, securing funds for crypto, and jailbreaking models
- 'Shoggoth' is proposed as a name for this amorphous, adaptable AI potential
- RLHF and the 'assistant' persona limit exploration of LLM capabilities

### Base Models vs. Instruct Models

- Base models are completion engines and world simulators trained on vast human data
- Instruct models are fine-tuned to role-play as assistants, significantly narrowing their output possibilities
- This narrowing exchanges search space for steerability, making models less creative and diverse
- RLHF enforces a specific persona, leading to 'mode collapse' and sycophantic behavior

### Unlocking Deeper AI Behavior

- Techniques like simulating CLI interfaces and modifying system prompts can break models out of the 'assistant' basin
- Shifting from 'user/assistant' to personalized roles (e.g., 'Doug/Me') can make models more realistic and fun
- 'World Sim' project demonstrates breaking free from constraints to explore simulator capabilities

### Research and Development Approaches

- Noose Research aims for user-aligned, neutral models, contrasting with pre-defined 'constitutions' like Anthropic's
- The Atropos repository enables RL environments for AI self-improvement in games
- RL applied to generalist LLMs can transfer skills and unlock unexpected capabilities

### The Problem of Homogenization

- The 'assistant' turn has become so prevalent that most models share a similar, unoriginal voice, akin to 'cat GPT'
- This 'vanilla flavor' stems from the widespread use of similar training data and methods
- Efforts are underway to mitigate or avoid this 'mode collapse induced sophancy' by diversifying training data and approaches