Skip to content

DataCTE/LatentExplorer

Repository files navigation

Subjective LLM Latent Explorer

Philosophy

This project is founded on a crucial understanding: intellectual qualities are not superposable and cannot be measured as linear surfaces. We reject the notion of qualifying the unqualifiable and acknowledge that many aspects of intelligence and cognition defy objective quantification.

Our aim is to create an LLM "benchmark" that fully embraces the subjective context of exploring intelligence. We recognize that traditional ML approaches often fall into the trap of trying to quantify the unquantifiable, relying on arbitrary assumptions and the flawed premise that quality is objectively known or measurable.

Approach

Rather than attempting to produce scores or rankings, this tool serves as a latent explorer of an LLM's internal representations. It employs:

  • Open-ended questions
  • Philosophical scenarios
  • Introspective prompts
  • Abstract concept exploration

The goal is not to measure or compare, but to reveal and examine the unique, subjective landscapes of different LLMs' cognitive processes.

Key Features

  1. Philosophical Probing: Engage LLMs with deep, open-ended philosophical questions that have no "correct" answers.

  2. Scenario Exploration: Present complex, ambiguous scenarios to observe how LLMs navigate ethical and logical dilemmas.

  3. Concept Mapping: Encourage LLMs to freely associate and connect abstract concepts, revealing their internal conceptual structures.

  4. Introspection Prompts: Ask LLMs to reflect on their own thought processes, exploring the limits of their self-awareness.

  5. Creative Synthesis: Challenge LLMs to combine disparate ideas in novel ways, showcasing their creative capabilities.

  6. Paradox Navigation: Present logical paradoxes and observe how LLMs grapple with inherent contradictions.

  7. Metaphor Generation: Prompt LLMs to explain complex ideas through original metaphors, revealing their ability to draw unexpected connections.

Output

The output of this tool is not a score or a ranking, but a rich, qualitative exploration of each LLM's unique cognitive landscape. Results are presented as:

  • Narrative summaries
  • Conceptual maps
  • Highlighted excerpts showcasing novel ideas or approaches
  • Reflections on the LLM's apparent strengths, limitations, and quirks

Usage

This tool is not for benchmarking in the traditional sense. Instead, use it to:

  • Gain insights into the philosophical and creative capacities of different LLMs
  • Explore how LLMs handle ambiguity, paradox, and open-ended questioning
  • Reveal unexpected capabilities or limitations of LLMs
  • Inspire new directions for AI research and development

Disclaimer

This project intentionally avoids quantitative comparisons between LLMs. Any apparent "performance" differences should be viewed as subjective observations rather than objective measures. The true value lies in the unique perspectives and ideas generated through this exploration.

Personal Opinion and Critique of Modern LLM Benchmarking

For an in-depth discussion on the philosophical issues with current LLM benchmarking practices and the motivation behind LatentExplorer, please read RANT.md

Contributing

We welcome contributions that align with our philosophy of embracing subjectivity and avoiding false quantification. Please feel free to suggest new philosophical prompts, scenarios, or analysis approaches that further our goal of exploring the unquantifiable aspects of artificial intelligence.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Remember: In the realm of intelligence and cognition, the most profound insights often lie beyond the reach of numbers and metrics. Our task is not to measure, but to explore, wonder, and learn.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages