自回归科学基座模型 BigBang-Proton，提出实现 AGI 的新路线

Core Insights - The article discusses the advancements made by the company 超越对称 (Super Symmetry) in developing the BigBang-Proton model, which integrates various scientific disciplines and challenges existing AGI approaches [1][2][4]. Group 1: BigBang-Proton Model Overview - BigBang-Proton successfully unifies multiple scientific problems across different scales, from subatomic particles to macro-level Earth systems, using a next-word prediction paradigm [2][4]. - The model addresses the limitations of current AGI technologies, such as GPT-5 and DeepSeek-R1, which struggle with understanding real-world material structures [2][4]. - The company proposes that material structure learning is essential for achieving AGI, allowing LLMs to engage with the physical world [4][5]. Group 2: Innovations in Pre-training Methodology - BigBang-Proton introduces three fundamental innovations: Binary Patch Encoding, a theory-experiment learning paradigm, and Monte Carlo Attention [9][12][19]. - Binary Patch Encoding replaces traditional tokenizers, allowing for unified processing of language, numerical, and scientific data, thus enhancing numerical analysis capabilities [11][12]. - The theory-experiment learning paradigm aligns numerical experimental data with theoretical knowledge, covering over 90% of experimental research tasks [13][14]. Group 3: Performance Metrics and Comparisons - BigBang-Proton demonstrates superior performance in arithmetic tasks, achieving 100% accuracy in addition and 98% in subtraction, significantly outperforming other models like DeepSeek-R1 and ChatGPT-o1 [36][38]. - In particle jet classification tasks, BigBang-Proton achieves an accuracy of 51.29%, competing closely with specialized models [44]. - The model also excels in material property predictions, achieving a mean absolute error of 0.043 eV/atom, outperforming many traditional machine learning methods [54][56]. Group 4: Applications in Scientific Domains - The model is applied to lake water quality prediction, achieving a mean absolute error of 0.58 μg/L, demonstrating its capability in environmental science [58][59]. - In genomic modeling, BigBang-Proton surpasses the performance of the leading model Evo, achieving a perplexity of 2.8 with significantly fewer training tokens [66][70]. - The model effectively predicts the functional impact of mutations on proteins and non-coding RNAs, showcasing its potential in biological research [71][72]. Group 5: Future Implications and Theoretical Insights - The company envisions that the pre-training of LLMs can extend to the entire universe, proposing a concept of "universe compression" to consolidate vast amounts of information into a single model [5][79]. - The advancements made by BigBang-Proton could lead to breakthroughs in various fields, including finance, engineering, and scientific research, by addressing the limitations of current LLM architectures [8][38].