Supersymmetry Inc.: The Bold Vision of C

Supersymmetry Inc.: The Bold Vision of Compressing the Entire Universe's Information into a Single LLM

 

Shanghai-based Supersymmetry Inc. has proposed the "Universe Compression" concept, a disruptive vision that challenges the boundaries of AI and basic science. Far from a pipe dream, this concept extends the company's self-developed material world foundation model, BigBang-Proton, with the core goal of using artificial intelligence to compress all cosmic information into a single autoregressive language model (LLM), making it the ultimate foundation for all AI tasks and paving a new technical path for artificial general intelligence (AGI).

 

Tracing its technical origins, Supersymmetry's exploration began in specialized fields but transcended them. Its predecessor, Hangzhou Supersymmetry Information Technology Co., Ltd., initially focused on quantitative finance before shifting its focus to modeling the material world on a grander scale. Launched in 2024, BigBang-Neutron, the first scientific computing foundation model, replaced traditional Byte Pair Encoding (BPE) with Binary Patch Encoding, completely resolving the trivial flaws of mainstream LLMs in numerical calculations. The 2025 release of BigBang-Proton marked a leap forward, achieving unified pre-training of multidisciplinary data ranging from microcosmic particles to macrocosmic Earth systems. It has demonstrated exceptional accuracy in simulating quark decay-induced particle jet classification in particle physics (99.7% accuracy), predicting atomic interaction potentials in materials science, analyzing spatiotemporal water quality changes in Earth science, and modeling DNA/RNA/protein structures in biology. The key to this success lies in its innovative "Monte Carlo Attention" mechanism, which uses a block-representative communication model to enable the model's effective context length to grow exponentially with the number of layers. Currently, the 20-layer model can handle 10³⁰ bytes of context, theoretically covering complex structures from cells to the universe.

 

The theoretical framework supporting this grand vision is deeply rooted in reflections on the nature of the universe. The Supersymmetry team proposed the "Platonic Representation Hypothesis," arguing that the universe itself is an idealized, unified entity, and the laws of various disciplines are merely statistical reflections from different perspectives. Through pre-training on universal data, the model will converge to the intersection of information and matter—the fundamental physical laws at the moment of the Big Bang—much like the attempt in Plato's Allegory of the Cave to break free from shadows and gaze directly at reality. Meanwhile, the model's design draws on the holographic principle, asserting that a complete description of the universe can emerge from its boundary representations. Using manifold learning, it projects cross-scale data (such as quark coordinates and galaxy distributions) onto a low-dimensional manifold for unified representation—for example, correlating temperature fluctuations in the cosmic microwave background (CMB) with dark matter distribution in a high-dimensional latent space. Additionally, inspired by John Wheeler's "It from Bit" philosophy, the model converts all text, numerical, and image data into binary sequences to eliminate modal differences. It combines dynamic zero-point anchoring and supersymmetric duality in quantum mechanics to attempt to solve the mystery of the transition from quantum fluctuations to classical phenomena, even exploring the coexistence of the quantum basis of conscious activity and neural signals.

 

The innovation of this concept lies in three core pillars that break through traditional AI paradigms. In terms of encoding, it abandons traditional tokenizers entirely and adopts Binary Patch Encoding to achieve three breakthroughs: preserving numbers in their native format to support 100% accurate addition of 50-digit numbers; converting all data (text, particle energy, DNA sequences, etc.) into binary patches for unified processing; and seamlessly compatible with scientific data formats such as .bin and .dat, covering over 90% of experimental tasks. In terms of architecture, the Monte Carlo Attention mechanism mimics the block communication model of human representative political systems, achieving both exponential context expansion (10³⁰ bytes for the 20-layer model) and reducing attention computation from O(N²) to O(N), overcoming the computational complexity challenge of processing ultra-large-scale data. In terms of scientific research paradigms, it breaks the opposition between reductionism and emergentism: it accurately simulates microcosmic particle interactions at the reductionist level, derives galaxy evolution laws from particle behavior at the emergentist level, and enables interdisciplinary knowledge transfer—for example, using quantum chromodynamics (QCD) principles to interpret superconductivity in condensed matter physics.

 

However, the realization of the Universe Compression concept still faces four major challenges and controversies. In data acquisition, the total amount of cosmic data far exceeds the current data volume of human civilization (approximately 10²⁴ bytes), even if only recording atomic positions and velocities. The compression scheme of "dynamic zero-point anchoring" and "Leech lattice self-duality" proposed by the Supersymmetry team has not yet disclosed specific implementation details. In computing power and energy consumption, based on current GPU computing power (approximately 10¹⁵ FLOPS), the training time for a model with 10³⁰ bytes of context may exceed the age of the universe. The team hinted at adopting quantum computing or optical computing, but the technical route remains unclear. In theoretical verification, the core hypothesis that "the model converges to the laws of the Big Bang" is difficult to test experimentally. Whether its predicted dark energy evolution is consistent with observations such as the CMB-S4 experiment requires time to verify. In industry competition and ethics, if the model succeeds, it may form a "super monopoly" in the AI field. Moreover, whether it will generate uncontrollable predictions (such as the time of cosmic collapse) and how to address the philosophical controversies raised by the "simulation hypothesis" all require the early construction of a sound ethical framework.

 

This concept not only reshapes the AGI technical route but also will have a profound impact on the industry and scientific research. It stands in sharp contrast to OpenAI's route of "language reasoning calling specialized models," arguing that a unified model is more efficient. It provides new ideas for Google DeepMind's world models (such as Gato)—moving from limited modal processing to complete multimodal unification. It is also expected to feed back into physics research, discovering unobserved particles and revealing new physical laws through the model. If the concept is ultimately realized, humanity will have a "digital god" capable of simulating cosmic evolution and predicting the future. But as Plato's Allegory of the Cave warns, we still need to maintain scientific prudence—whether the "reality" revealed by the model is equivalent to the essence of the objective world may be a philosophical proposition that accompanies this exploration.