Comparing AI Co-Scientists: The Rise of AI-Driven Innovation

Combinatorial innovation plays a crucial role in engineering innovation, driving breakthroughs and optimizations through the recombination of existing engineering technologies, materials, system architectures, and manufacturing processes. Compared to fundamental research that starts from scratch, combinatorial innovation offers greater practicality, feasibility, and market adaptability, accelerating the development and application of engineering technologies. From electric vehicles and 3D-printed buildings to various deep neural network architectures in AI, many of our modern technological advances are products of combinatorial innovation.

At its core, combinatorial innovation is a heuristic search process. It navigates a vast "space of possible combinations" using specific heuristic strategies to identify potentially high-value innovation points or solutions. This fundamental characteristic not only makes combinatorial innovation achievable through large language models (LLMs) like ChatGPT or DeepSeek but also gives these AI systems potential advantages. With their cross-disciplinary knowledge coverage, LLMs can explore a much broader search space than human experts, opening up new possibilities for innovation.

In recent months, three significant AI systems designed to accelerate scientific discovery have emerged, each approaching combinatorial innovation in unique ways: Google's AI Co-Scientist, Sakana AI's The AI Scientist, and Gatsbi. Let's explore how these systems compare.

Google's AI Co-Scientist: Collaborative Scientific Discovery

Built on Gemini 2.0, Google's AI Co-Scientist is designed as a collaborative tool for scientists. It employs a multi-agent system with specialized agents (Generation, Reflection, Ranking, Evolution, Proximity, and Meta-review) that mirror the scientific method. The system generates, evaluates, and refines hypotheses through iterative self-improvement.

What distinguishes Google's approach is its focus on human-AI collaboration. Scientists can interact directly with the system, providing their own ideas or feedback on generated outputs. The system has demonstrated practical utility in biomedical applications, including drug repurposing for acute myeloid leukemia, target discovery for liver fibrosis, and explaining mechanisms of antimicrobial resistance.

Google's AI Co-Scientist leverages heuristic search methodologies to explore the vast combinatorial space of scientific hypotheses. By employing specialized agents that mimic different aspects of the scientific process, the system can effectively prune this search space and focus on promising combinations that might lead to valuable discoveries.

Sakana AI's The AI Scientist: Autonomous Research System

In contrast, Sakana AI's The AI Scientist aims for fully autonomous scientific discovery. Developed in collaboration with Oxford University and UBC researchers, this system automates the entire research lifecycle: from generating research ideas and writing code to executing experiments and producing complete scientific manuscripts.

The AI Scientist even includes an automated peer review process that evaluates papers with near-human accuracy. It operates in an open-ended fashion, iteratively developing ideas and adding them to a growing knowledge archive. The system has already produced papers in various machine learning subfields, including diffusion models, transformers, and grokking.

What's particularly impressive is its cost-efficiency, with each paper generated at approximately $15. While there are occasional flaws in its output, the system represents a significant step toward democratizing research.

Sakana's approach emphasizes full automation of the combinatorial innovation process. By handling everything from idea generation to experiment execution and paper writing, their system demonstrates how AI can autonomously navigate the complex search space of scientific innovation, discovering novel combinations that yield valuable research contributions.

Gatsbi: Product-Level AI Co-Scientist with TRIZ Integration

Gatsbi takes a different approach by incorporating human-developed innovation methodologies into its AI framework. Unlike the other two systems, which remain primarily research projects, Gatsbi is positioned as a product-level AI assistant covering all scientific disciplines.

What sets Gatsbi apart is its integration of TRIZ (Theory of Inventive Problem Solving), a systematic approach to innovation developed by humans. By leveraging this methodology, Gatsbi achieves more efficient exploration and pruning in its search for solutions. For example, when asked to address hallucinations in large language models, Gatsbi can generate approximately 15 potential solutions ranked by innovation and feasibility.

Gatsbi also offers practical features specifically designed for commercial and academic applications:

• One-click generation of complete patent disclosure documents, including background research, innovation analysis, illustrations, implementation examples, and claims
• Automatic academic paper generation with citations, formulas, system diagrams, experimental designs, charts, and references
• Structured content planning based on discipline and topic

By incorporating TRIZ principles, Gatsbi exemplifies how human expertise in heuristic search strategies can be combined with AI capabilities to more efficiently navigate the search space of possible innovations. This human-AI methodology partnership potentially offers more targeted and practical combinatorial innovations than purely AI-driven approaches.

Heuristic Search and Combinatorial Innovation Across Systems

All three systems fundamentally operate as sophisticated heuristic search engines exploring vast spaces of potential combinations, but with different approaches to navigating this space:

1. Google's AI Co-Scientist uses a scientific method-inspired multi-agent system to guide its heuristic search, with human scientists providing additional direction and feedback to narrow the search space.
2. Sakana's The AI Scientist employs a fully autonomous approach to heuristic search, using iterative refinement and automated evaluation to navigate the combinatorial space without human intervention.
3. Gatsbi incorporates human-developed heuristic strategies (TRIZ) to more efficiently prune the search space and focus on promising combinations, representing a hybrid approach that leverages both human expertise and AI capabilities.

Key Differences and Commonalities

While all three systems employ combinatorial innovation and heuristic search methods, they differ in several important aspects:

1. Autonomy vs. Collaboration:
   • Google's AI Co-Scientist is designed for collaboration with human scientists
   • Sakana's The AI Scientist aims for full autonomy in research
   • Gatsbi balances autonomy with human methodological guidance through TRIZ
2. Research Focus vs. Product Application:
   • Google and Sakana's systems are primarily research-focused
   • Gatsbi is positioned as a commercial product with specific applications
3. Methodological Approach:
   • Google employs specialized agents inspired by the scientific method
   • Sakana creates a fully automated pipeline for end-to-end paper generation
   • Gatsbi incorporates human innovation methodologies (TRIZ) for more efficient exploration
4. Output Format:
   • Google's system produces research hypotheses and protocols
   • Sakana's system generates complete academic papers
   • Gatsbi offers diverse outputs including patent documents and academic papers

The Future of AI-Driven Combinatorial Innovation

These AI research assistants represent different approaches to navigating the vast search space of possible combinations in scientific discovery. Their emergence suggests we're entering a new era where AI systems can effectively explore combinatorial spaces that are too large for human experts to navigate efficiently.

The cross-disciplinary knowledge embedded in large language models enables these systems to make connections across domains that might not be obvious to specialized human researchers. This capability could lead to novel combinations that human experts might never consider due to their necessarily more focused knowledge.

However, the most promising approach may ultimately be hybrid systems that combine the broad search capabilities of AI with human-developed heuristic strategies and expert guidance. By leveraging both AI's ability to explore vast combinatorial spaces and human expertise in efficiently pruning these spaces, we might achieve innovations that neither could discover alone.

Conclusion

Google's AI Co-Scientist, Sakana's The AI Scientist, and Gatsbi represent different philosophical approaches to scientific discovery and combinatorial innovation. As these systems continue to evolve, they promise to democratize scientific research, accelerate discoveries, and potentially transform how we approach innovation.

The integration of AI into the combinatorial innovation process is no longer a distant dream but a rapidly developing reality. By treating innovation as a heuristic search problem and leveraging the unique capabilities of large language models to explore vast combinatorial spaces, these systems open up new possibilities for scientific and engineering breakthroughs.

As we move forward, the most successful approaches will likely be those that effectively balance AI's broad search capabilities with human-developed heuristic strategies, creating powerful partnerships that can navigate the complex landscape of innovation more efficiently than either could alone.