About the Project

Lead Author

Jerry Zhang

An investor and entrepreneur. Passionate about combining science fiction concepts with rigorous theories.

Email: jerry.zhang.bill@gmail.com

GitHub Profile

Project Inspiration

Inspired by the movie Interstellar, where humanity transcends known space through wormholes and black holes. Similarly, Interstella treats semantic spaces as a "Token Cosmos," enabling AGI emergence through traversals, akin to AHA moments in human cognition.

"Mankind was born on Earth. It was never meant to die here."

Featured Launch Article

Welcome to Interstella: Traversing the Token Cosmos to AGI's Interstellar Journey

Imagine standing on the edge of a vast universe, surrounded by twinkling stars, each representing an idea, a word, or a line of reasoning. This is not a scene from the science fiction movie Interstellar, but the core metaphor of our project "Interstella": in the semantic space of large language models (LLMs), tokens form clusters like stars, creating a dynamic "Token Cosmos." Just as humans in the movie traverse wormholes to find new homes, our project aims to guide LLMs from known semantic regions to "leap" into unknown territories, achieving continuous innovative emergence. This is not just technical exploration, but an engineering blueprint for AGI (Artificial General Intelligence) and ASI (Artificial Super Intelligence).

The Interstella project draws inspiration from Interstellar's themes: the courage to face the unknown and precise scientific navigation. We view the embedding space of LLMs as a universe, where token sequences are "trajectories," and traversals are leaps from conventional reasoning to breakthrough insights. In the short term, we focus on enhancing the controllability, alignment, and efficiency of LLM reasoning; in the long term, we hope to pave a reliable engineering path to truly autonomous AGI/ASI. Let's unveil this project step by step.

Project Origins: From Sci-Fi to Real Token Cosmos

In Interstellar, humanity achieves cross-dimensional travel through black holes and wormholes. Similarly, in the AI field, the semantic space of LLMs like the GPT series or Qwen models is not a flat plane but a composite of "manifolds"—known knowledge forms dense star regions, while innovation hides in sparse frontiers. Our project names this space "Token Cosmos," where each token sequence is a trajectory.

Industry research shows that such emergence is not random. OpenAI's o1 model achieves longer reasoning paths through chain-of-thought and self-verification, improving the resolution rate of complex problems. Similarly, breakthroughs like DeepSeek and AlphaGo's "Move 37" demonstrate "Move 37" moments: leaping from known patterns to new strategies. In our experiments, we simulated similar processes using the Qwen model to generate multiple trajectories, observing that temperature sampling perturbations can repeatedly induce AHA (aha) moments—trajectories crossing from stable zones to new semantic clusters, similar to wormhole leaps in the movie.

These experiments are based on visualization tools, such as trajectory bundle diagrams and uncertainty heatmaps, showing how LLMs navigate semantic landscapes during response generation. Industry results like Google's PaLM model research further confirm that as model scale increases, emergent capabilities grow exponentially. The Interstella project builds on this foundation, transforming sci-fi metaphors into operable frameworks.

Short-Term Goals: Making LLM Reasoning More Controllable, Aligned, and Efficient

In the current landscape, LLMs like ChatGPT often face "hallucinations" and uncontrollability issues during reasoning. Our short-term goal is to enhance these aspects through engineering tools. We have developed a five-layer pipeline: navigator positions semantic boundaries, trajectory generator introduces perturbations, verifier cross-checks facts, semantic map dynamically updates known areas, and learning loop provides feedback optimization.

In experiments, we tested three progressive prompts: from simple mathematical proofs to quantum experiment design, to cross-domain AI frameworks. After 10 runs, results showed that by adjusting sampling parameters (e.g., temperature from 0.7 to 1.2), we can control leap probabilities—higher perturbation increases innovation, but the verifier layer ensures alignment with human intent. For example, in a quantum entanglement prompt, the standard output follows Bell's inequality, but after perturbation, an idea emerged to combine entanglement with AI error correction, improving efficiency.

Industries like Anthropic's Claude model have adopted similar self-critique mechanisms to enhance alignment. Our visualization tools (such as attention wormhole intensity maps) further reveal the internal dynamics of these leaps, helping developers debug and optimize, making reasoning more efficient—in experiments, AHA detection reduced invalid generation paths by 30%.

Long-Term Vision: Engineering the Path to AGI/ASI

Interstella goes beyond optimizing existing LLMs; our vision is to find an engineering blueprint for AGI/ASI. Through a differential geometry perspective, we view semantic leaps as continuous emergence from one knowledge manifold to another—similar to time curvature in Interstellar, but achieved controllably with probabilistic dynamical systems.

Industry research like OpenAI's scaling laws shows that model parameter growth induces phase transitions, leading to new capabilities. We extend this idea: through wormhole-like attention mechanisms and random perturbations, LLMs can repeatedly achieve "Move 37" effects, ultimately building autonomous intelligence. Experimental simulations confirm that under high-complexity prompts, multi-round leaps can produce cross-domain frameworks, such as applying category theory to federated learning for optimized distributed AI.

In the long term, we envision an ecosystem: open-source toolchain, community-contributed semantic maps, and high-impact outputs under ethical constraints. This is not a distant dream—with the emergence of o1-like models, the threshold for AGI is already lowering.

Join the Interstella Journey

Welcome to Interstella! This project is open-source on GitHub, and we invite researchers and developers to participate: contribute code, test prompts, or discuss ethical risks (such as misalignment misuse). Stay tuned—the interstellar journey has just begun.

Contact

For collaborations, questions, or feedback, please get in touch.

Powered by Formspree – responses sent to project email.