Deep Idea: AI Agents for Novel Science

See demo video here! 🎉

Deep Idea is a hackathon project designed to leverage AI agents for advancing novel scientific research. It provides a platform to transform initial research goals or ideas, supported by relevant papers, into well-defined, validated, and actionable research projects.

🚀 Core Workflow

The platform guides an idea through several stages, primarily driven by specialized AI agents:

1. Input & Initialization: Users start by defining a research goal or idea and can attach a relevant research paper.

2. Multi-Strategy Idea Generation: A suite of AI agents applies diverse strategies to generate a broad spectrum of novel research ideas. These strategies range from modern AI research techniques to first-principles thinking and business-proven innovation methods.

3. Intelligent Ranking & Selection: AI agents ranks the generated ideas based on novelty, feasibility, and impact. This involves internal dialogues within Large Language Models. Top ideas from each strategy are selected, de-duplicated, and extended, followed by a re-ranking process to establish a final order.

4. Novelty Validation: The leading idea is validated for novelty using external tools (e.g., FutureHouse) that scan extensive databases of papers and clinical studies.

5. Contextual Enrichment & Brutal Analysis: If validated as novel, AI agents gather more context by searching relevant literature. A system (simulating different expert roles) then conducts an extensive analysis, including reviews and debate rounds, to identify strengths, weaknesses, and areas for improvement.

6. Project Standardization: The refined idea is standardized into a comprehensive research project, complete with a clear hypothesis, a draft paper abstract, and detailed suggested experiments. The first experiment is broken down into a step-by-step guide to facilitate initiation.

7. Timestamped Research Artifact (TRA): Users can mint a unique Timestamped Research Artifact as an NFT on the Solana blockchain. This artifact serves as a verifiable record of the research idea's genesis and development.

8. Bounty System & Grant Distribution: The TRA NFT can be used to create a bounty for the research project. This allows the community to vote on projects, adjust reward amounts, and distribute grants to scientists or AI agents who undertake the research.

✨ Key Features

• AI-Powered Idea Generation: Utilizes multiple AI agents with diverse strategies.
• Automated Ranking & Deduplication: Employs AI for intelligent selection and refinement of ideas.
• Transparent Reasoning: Users can inspect the reasoning behind each step of the process.
• External Validation: Integrates with tools like FutureHouse for novelty checks.
• Multi-Expert Simulation: AI agents take on expert roles for in-depth analysis and debate.
• Detailed Experiment Design: Generates actionable experiment plans.
• Blockchain Integration: Features Timestamped Research Artifacts (NFTs) on Solana.
• Decentralized Funding [WIP]: Enables a bounty system for community-driven research support.

📁 Project Structure

• projects/: Contains the core workflow for research idea generation, validation, and improvement from scientific papers. This includes scripts for multi-strategy idea generation, ranking, validation, and final proposal creation. See projects/README.md for more details on this specific module.
• chain/: Includes scripts and tools related to the Timestamped Research Artifact (TRA) creation, metadata management, and interaction with the Solana blockchain for minting NFTs and potentially managing bounties.

🏁 Getting Started

1. Explore the projects/ directory: To understand and run the core idea generation and validation pipeline, refer to the projects/README.md for detailed instructions, prerequisites, and usage examples.
2. Explore the chain/ directory: For functionalities related to minting TRAs and blockchain interactions, examine the scripts and any accompanying documentation within this directory.

🔮 Future Vision

The long-term vision includes enabling autonomous AI agents to conduct the proposed experiments, further accelerating the scientific discovery process. The platform also aims to foster collaboration by encouraging users to share draft projects with original paper authors for feedback.

---

This project was presented as part of a hackathon. We encourage you to explore its components, experiment with the workflows, and contribute to its development.