ShroomGPT: Modeling the Neural Energy Landscape of Psychedelic Effects Using General Relativity-Inspired ODEs

ShroomGPT is an open-source interdisciplinary project that attempts to use general relativity-inspired ordinary differential equations (ODEs) to simulate the impact of psychedelic substances on the reasoning process of large language models (LLMs), exploring the dynamic changes in the neural energy landscape of these networks. By modifying the reasoning state of LLMs instead of conducting human experiments, this project provides a new paradigm for psychedelic research and consciousness exploration.

Psychedelic substances (e.g., psilocybin) show significant potential in treating conditions like depression and PTSD, but research faces challenges such as strict ethical constraints, high costs, and difficulty quantifying results. ShroomGPT proposes an alternative approach: using LLMs to simulate certain aspects of human cognition, and modifying model states to simulate the cognitive effects of psychedelic experiences. In neuroscience, the 'energy landscape' metaphor describes the state space of the brain; psychedelic substances are believed to smooth this landscape, lowering the threshold for state transitions and producing experiences like fluent thinking.

ShroomGPT borrows mathematical tools from general relativity that describe spacetime curvature and applies them analogously to neurogeometry: using a similar ODE framework to describe the 'curvature' of the cognitive state space caused by neural network activation patterns, involving manifold dynamics, metric tensor changes, geodesic equations, etc. The implementation of the ODE system includes: 1. Attention weight renormalization (simulating attention shifts); 2. Activation pattern diffusion (simulating neural activity breaking conventional boundaries); 3. Time perception distortion (modifying time constants for sequence processing). The project chooses to intervene during the inference phase, with advantages such as being able to use any pre-trained model, real-time parameter adjustment, and avoiding retraining costs.

Despite the elegance of the mathematical framework, correlating ODE parameters with subjective experiences remains challenging. The project currently focuses on objective measurable metrics: semantic diversity of generated text, topic transition frequency, metaphor density, self-reference frequency, etc., as proxy variables for 'psychedelic intensity'. However, there is still a gap from understanding real conscious experiences.

The scientific value of ShroomGPT lies in opening up a new research paradigm: enabling large-scale, low-cost 'cognitive pharmacology' experiments, quantitatively comparing the characteristic signatures of psychedelic substances, and building predictive models for therapeutic effects. At the same time, ethical boundaries must be addressed: whether the simulation effects are misused, whether AI-generated text is misleading, whether the complexity of real experiences is downplayed, etc. The project emphasizes pure research purposes and responsible use.

ShroomGPT represents a unique direction in AI research—using AI to understand the human mind. In the future, with the development of multimodal models and embodied intelligence, more rich cognitive simulation environments may be built. Regardless of whether it can accurately simulate psychedelic experiences, the project demonstrates the value of interdisciplinary thinking: combining physics, mathematics, and neuroscience, and using AI to serve human self-understanding. LLMs are not only tools but also detectors for exploring the frontiers of cognition.