Chain of Recursive Thoughts: Make AI think harder by making it argue with itself (github.com)

Article summary

The article discusses a method called Chain of Recursive Thoughts (CoRT) that makes AI models think harder by arguing with themselves repeatedly. This approach involves generating initial responses, evaluating alternatives, and refining the output through multiple rounds of thinking. The author claims that CoRT significantly improves the performance of AI models, especially for small models, in tasks such as programming. The method is made available through a web UI and a Python script.

Main themes

• AI model improvement
• Recursive thinking
• Workflow design
• Abbreviation schemes
• AI personalities
• Consciousness and competing processes
• Prompt design and iteration

What commenters say

• Some commenters are experimenting with similar approaches, such as designing workflows with multiple agents and critics to improve AI output.
• Others suggest using existing tools and services, like n8n, to build and manage complex AI workflows.
• There is a debate about the effectiveness of abbreviation schemes, with some arguing that they are useful and others claiming that they are confusing and obfuscate the message.
• Some commenters believe that telling AI models to be blunt or critical can actually downgrade their performance and lead to hallucinations and inaccurate statements.
• Others propose that using multiple AI models or personalities in parallel or series can lead to more novel and high-quality ideas.
• A few commenters draw parallels between the concept of recursive thinking in AI and the human brain's ability to generate consciousness through competing processes.
• Some suggest that using metrics to score AI ideas and iterating until a certain threshold is met could be a useful approach to improve AI output.
• There is also a discussion about the limitations of AI models and how they can be gamed or improved through careful prompt design and iteration.