LoopDot 研究人工智能
最底层的技术。
LOOPDOT (ailoopdot.com) IS A RESEARCH TEAM WORKING AT THE FOUNDATIONAL LAYER OF ARTIFICIAL INTELLIGENCE — THE ARCHITECTURES, SYSTEMS, AND MECHANICS BENEATH EVERY APPLICATION.
在循环迭代中，打磨基础。
智能系统的能力边界，不由参数量决定，而由架构、系统与理论的联合优化决定。LoopDot（ailoopdot.com）不做应用层的缝合，只解决尚未被解决的底层问题——每一处结论都要能在数学上立住、在工程上跑通。
CAPABILITY IS BOUNDED NOT BY PARAMETER COUNT, BUT BY THE JOINT OPTIMIZATION OF ARCHITECTURE, SYSTEMS, AND THEORY. WE DON'T STITCH TOGETHER APPLICATIONS — WE SOLVE PROBLEMS THAT HAVEN'T BEEN SOLVED AT THE FOUNDATION. EVERY CLAIM HOLDS MATHEMATICALLY AND RUNS IN PRODUCTION.
五个核心方向
模型架构Model Architecture
研究稀疏专家路由、亚二次复杂度注意力与非自回归解码拓扑，在参数规模与推理延迟的帕累托前沿上寻找可行解。
SPARSE MOE ROUTING, SUB-QUADRATIC ATTENTION, NON-AUTOREGRESSIVE DECODING — SEARCHING THE PARETO FRONTIER BETWEEN PARAMETER COUNT AND INFERENCE LATENCY.
大规模训练系统Training Systems
构建千卡级分布式训练管线，研究显存分片、梯度检查点与弹性容错机制，把每一块 GPU 的算力压榨到极限。
THOUSAND-GPU TRAINING PIPELINES — MEMORY SHARDING, GRADIENT CHECKPOINTING, AND ELASTIC FAULT TOLERANCE TO EXTRACT EVERY CYCLE OF COMPUTE.
推理与效率优化Inference & Efficiency
通过 KV 缓存压缩、投机解码与低比特量化，把大模型的服务成本压到工程可承受的边界。
KV-CACHE COMPRESSION, SPECULATIVE DECODING, LOW-BIT QUANTIZATION — DRIVING SERVING COST DOWN TO THE EDGE OF WHAT'S ENGINEERING-FEASIBLE.
学习算法与理论Learning & Theory
从缩放定律到泛化边界，用信息论与优化理论解释模型为什么有效，而不只是它有效。
FROM SCALING LAWS TO GENERALIZATION BOUNDS — EXPLAINING WHY MODELS WORK, NOT JUST THAT THEY DO.
对抗性智能与战略决策Adversarial Intelligence & Strategic Decision Theory
研究智能体在持续对抗环境下的策略学习与均衡求解，为人机协同决策与军事战略博弈建立统一的数学框架，已有阶段性成果发表。
STRATEGIC LEARNING AND EQUILIBRIUM-FINDING UNDER CONTINUAL ADVERSARIAL PRESSURE. GROUNDED IN GAME THEORY RATHER THAN WEAPONS ENGINEERING.
论文与技术报告
Adaptive Adversarial Intelligence
A Unified Mathematical Theory of Learning Competition in Military Artificial Intelligence
1 Institute of Defence and Strategic Studies (IDSS), S. Rajaratnam School of International Studies (RSIS), Nanyang Technological University (NTU), Singapore
2 LoopDot AI Research
ABSTRACT Artificial intelligence is transforming military decision making from static policy optimization into continual learning against adaptive intelligent opponents.