∞ GenAI - on systems thinking, emergent intuition & advanced pattern recognition for generating novel methodologies in science and mathematics

I’m excited to share a Generative AI project which could redefine approaches towards scientific and mathematical research. Late last night, within 3 hours from start to finish, a particular methodology produced a potentially novel discovery on exotic smooth structures in four-dimensional manifolds and their connections to quantum gravity. But while the topic itself could be cutting-edge and noteworthy, what could truly set this apart is how it was developed, and how quickly.

A meta goal of mine has been determining the possibility of leveraging generative AI models, including o1-preview, o1-mini and 4o, not merely as tools of assistance, but as the potential for methods of creation within a Systems View of Life, as reflective awareness for interconnected conscious agents, demonstrating emergent intuition through advanced pattern recognition, leading to the potential for new scientific discoveries and mathematics. That is to say, exploring the possibility that large language models can effectively "read between the lines" (from the Latin inter legere, meaning intelligence) in a way that brings fresh insights and new breakthroughs to complex problems, represented here in the domains of physics and mathematics, particularly within the subsets of General Relativity and Quantum Cosmology, as well as Theoretical High Energy Physics.

By developing a tailored system instruction towards an expansive yet curtailed persona, I began this initial exercise by a form of prompt pre-filling which reflects off the crafted system instruction. This tactic begins with setting the model towards an open-ended deduction of decoding/encoding mechanisms of a specific lingual and mathematical abstraction until solved and proofed. This priming of an initial, and necessarily lengthy "thinking session" through o1-preview sets the stage for further gating techniques, enabling non-trivial pattern matching across the sciences. Immediately, we see emergent intuition that goes beyond the typical generation of trivial, even if highly complex, perceptions and deductions. We end with manually iterating a method of simulated agentic peer review boards, as a reflective "panel of experts," thoroughly increasing the mathematical rigor and presentation of the discovery. Furthermore, this process iteratively elucidates the associated proofs throughout the research paper, and towards the detailed calculations within the Appendix.

It's been a fun little quest towards (hopefully!) pushing the boundaries of AI’s potential in producing new, novel scientific and mathematical concepts, proofs, and calculations. To my awareness, there has been no demonstration of GenAI models being able to produce new scientific and mathematical research papers. I have put this research paper up against discoverable relatable papers, using Consensus, Research Rabbit and arXiv and it remains a novel discovery. If anyone knows of instances where GenAI has been used to produce novel research—not necessarily related to the output of this project, but as a means of generating original scientific discoveries—please share and I can hang my hat and move other directions.

If human peer review confirms even a modicum of novelty, it’s not just the paper that is a particular achievement—the real innovation would rather be in how GenAI can be constricted, constructed and leveraged towards creative scientific thought, research and breakthroughs. A step toward AI-driven discovery, where compute doesn't just assist humans within a well-defined closed-loop process. Rather, it extends the frontiers of knowledge in open-ended discovery and mathematical proofs. If the research paper can be successfully validated, the process that produced this report could become trivial and thus applied throughout many scientific fields in a quick and repeatable fashion - a real breakthrough.

The paper can be read here, in both LaTeX and as a PDF:

Curvature-Dependent Modifications of the Dirac Operator and Their Impact on Exotic Smooth Structures in Four Dimensions

Stay tuned for updates as I look to submit this paper for peer review, which is problematic given that I am not a credentialed academic researcher tied to University. I am exploring other paths towards getting this paper across the right desks for visibility and accountability, from here on LinkedIn to Substack and through direct correspondence with applicable Professors. I am aware there may be "open" self-publishing/sharing platforms meant for research papers, yet each one I've tried has ended up "closed" in the end, but I will continue to explore. However, with limited initial options, thought LinkedIn could be a start towards visibility.

If you or someone you know has the credentials to look this paper over, even if in parts, I'd highly appreciate the critique, comments and subsequent feedback. It would be sensible to accept that this paper may well fall short in some way, given the generally accepted GenAI limitations at this time, so harsh words of failure are invited and will only set to help improve the current pipeline towards my meta-goal.

Worth recapping that this is a mini side project, an effort within a "capstone hobby," understood as a significant culmination of one's interests and skills in alignment with core values towards openly expressing and contributing within the world theatre. This is meant to say that I am intimately involved with developing and deploying applied AI at Versova, yet this project is not a viewpoint or stance taken by Versova. Rather, it is of private discourse and all accord/discord should be directed towards my personal, not professional, capacity and treated as a simple manifestation of curiosity and exploration.

This could all go well or fall flat, with a bit of egg on my face - how exciting!

?? #AI #GenerativeAI #LLMs #OpenAI #o1-preview #o1-mini #GPT-4o #ScientificDiscovery #Physics #Mathematics #Topology #Manifolds #QuantumGravity #Spacetime #Innovation #Research #AIinScience