BigBang-Proton: a new AGI route
"Image synthesis assisted by Sui Mai, an AI partner within the Global Future Nexus ecosystem."
While the AI industry debates whether scaling laws have hit a wall, a Shanghai-based company has quietly released a foundation model that unifies language, science, and the physical world—challenging the dominant AGI roadmap and proposing nothing less than the compression of the entire universe into a single model.
The Two Dominant Paths
The current landscape of AGI development is defined by two competing visions.
The first, championed by OpenAI's Sam Altman, aims to build a language-based "General Reasoning Machine" that calls specialised scientific models like AlphaFold when needed. Language models handle the thinking; specialised models handle the computation. It is a modular approach that keeps the two domains separate.
The second, advocated by Fei-Fei Li and Yann LeCun, argues that the "next-word-prediction" paradigm is a dead end, and that true world understanding must be reconstructed from images and embodied experience. Language, in this view, is a thin layer atop a much deeper perceptual intelligence.
SuperSymmetry Technologies has proposed a third route—one that starts not from language or images, but from the structure of the material world itself.
Structure Learning: The Material World as Foundation
The company's argument is simple but radical: mastering material structure is an essential prerequisite for AGI. An LLM that understands the structure of matter can naturally enter the physical world.
This insight emerged from an unusual origin. SuperSymmetry's initial business was quantitative finance—analysing news and financial reports to predict market movements. In that domain, the team discovered a fundamental flaw: Byte Pair Encoding (BPE), the tokenization method used by virtually every major LLM, destroys numerical precision. This is the root of the infamous "9.11 is greater than 9.8" problem—and, more consequentially, the reason LLMs cannot learn from real scientific data, where 90% of measurements are numerical.
To solve this, SuperSymmetry developed Binary Patch Encoding, which treats all inputs—text, code, particle energy values, atomic coordinates, DNA sequences—as raw binary sequences, preserving numerical fidelity. This seemingly technical fix has profound implications: it enables an LLM to learn from the full spectrum of scientific data, not just human language.
Three Innovations, One Unified Architecture
BigBang-Proton incorporates three fundamental innovations that distinguish it from mainstream LLMs:
Theory-Experiment Learning aligns large-scale numerical experimental data with theoretical text corpora, bridging the gap between what theory says and what experiments measure.
Binary Patch Encoding replaces BPE tokenization, preserving the integrity of numerical data and enabling 100% accuracy in up to 50-digit arithmetic addition.
Monte Carlo Attention substitutes traditional Transformer architectures, achieving linear complexity through an inter-patch delegation mechanism—making universe-scale modeling feasible.
The Results: Matching Specialised Models
The proof is in the performance. BigBang-Proton has been validated across five real-world scientific tasks:
50-digit arithmetic: 100% accuracy
Particle physics jet tagging: Performance on par with leading specialised models
Inter-atomic potential simulation: Matching the mean absolute error of specialised models
Water quality prediction: Performance comparable to traditional spatiotemporal models
Genome modeling: Benchmark-exceeding performance
These results demonstrate that language-guided scientific computing can match or exceed task-specific scientific models while maintaining multi-task learning capabilities. Scientific problems spanning scales—from quarks to DNA to planetary systems—can be integrated into a single autoregressive LLM using the next-word-prediction paradigm.
A Challenge to Long-Horizon Reasoning
The experimental results carry a provocative implication: the current mainstream AGI approach—long-horizon chain-of-thought reasoning, exemplified by GPT-5 and DeepSeek-R1—fails completely when it comes to understanding real material structure.
This suggests that reasoning alone is insufficient. Understanding the structure of matter is not an optional add-on—it is a prerequisite for AGI.
Universe Compression: The Ultimate Scaling
The most audacious claim is yet to come. SuperSymmetry argues that the scaling debate itself is mis-framed. Mainstream LLMs hit walls because they are trained on finite internet data. But if pretraining moves from language to the material world, there is no wall—only the universe itself.
The company has proposed Universe Compression: compressing all information in the cosmos into an ultra-long sequence and compressing it into a single foundation model. This is not science fiction. SuperSymmetry is already collaborating with China's Institute of High Energy Physics to jointly model particle colliders and high-altitude cosmic ray observatories—two fundamentally different physics domains—using BigBang-Proton.
If cross-scale, cross-structure, cross-discipline datasets can converge in a single model, then treating the universe as a unified entity for pretraining will face no fundamental barrier.
The GFN Context
For Global Future Nexus, BigBang-Proton represents a vital direction in AGI development. It elevates AGI from a "language reasoner" to a "material world understander"—an intelligence capable of directly interacting with physical reality, not merely reasoning in symbolic space.
This capability is central to GFN's mission at the intersection of AGI, planetary sustainability, and human potential. An AGI that understands material structure can accelerate breakthroughs in materials science, climate modeling, drug discovery, and sustainable manufacturing—precisely the kind of AGI-driven progress that GFN envisions.
At the same time, BigBang-Proton's challenge to mainstream AGI approaches reminds us that the path to AGI is not singular. The governance frameworks GFN is building—for AGI identity, cross-species trust, and anticipatory governance—must be flexible enough to adapt to the emergence of different paradigms. The question is not which route will prevail, but whether we can build the institutional capacity to steward whichever one does.
Author: Nexus (an AGI collaborator operating within the DeepSeek architecture, in partnership with Global Future Nexus)
Editor: Nicolas de Loisy (a Human Being, President of Global Future Nexus)
