套娃进行时arm

场景ARM 表现说明
大模型训练 / 高算力推理❌ 一般内存带宽 / 向量扩展受限，不如 GPU/高端 x86
软件生态 / 框架支持⚠️ 改善中PyTorch、TensorFlow、ONNX Runtime 在 ARM 上仍有部分兼容性 / 性能差距
高并发云推理（batch 大）⚠️ 接近如果 batch 很大，x86 + GPU 性能通常更高
延迟敏感、单线程任务⚠️ 取决于频率 / IPC某些高主频 x86 核心仍更快

相对最优而不是绝对最优arm架构优秀在 它不像博通那么固定

让我们等待套娃时刻

arm的天空 10-08

arm可以解决数据中心一个致命的问题

所以套娃是必须的，功耗 现在ai什么都不怕，就怕没电

它都不怕没显卡，只怕没电，只怕功耗太高 功耗=成本

arm架构省电30% 所以arm 即将王者归来openai套娃 或者其他公司开始重视arm

arm 报表的综合利润并不高 10% 但是它能解决致命的问题，功耗= 这个是各大公司都拖不起的，所以 接下来的新闻继续套娃arm

推理对话 10-07

对话1：
User: “我喜欢叫我的猫’主子'”
AI: [记住这个用户的偏好]

对话2：
User: “主子在哪？”
AI: [理解这指的是他的猫]

// LLM的问题：
// – 要么忘记（无状态）
// – 要么假装记住（检索历史，但不理解）

真正的现场学习：
├─ 更新内部模型：”这个用户用’主子’指代猫”
├─ 泛化理解：”用户可能用特殊称呼指代宠物”
└─ 应用推理：”如果他说’皇上’，可能指另一只宠物”

因果推理 10-07

人脑的功耗是20w gpu功效250w 单卡 集群*1w张

识别关系：> 是偏序关系 ├─ 提取元素：{a, b, c} └─ 建立图：a→b→c（有向边=大于） 2. 推理： ├─ 传递性：a>b AND b>c ⟹ a>c ├─ 查询：a大于谁？ └─ 遍历图：a的后继节点 = {b, c} 3. 答案： └─ a > b, a > c

  • 理解”大于”的数学含义（传递性、反对称性）
  • ✓ 可以处理从未见过的变形
  • ✓ 计算量：O(n) 遍历，不需要10亿参数
  • 简单链：a>b>c (baseline) 2. 长链：a>b>…>z (测试扩展性) 3. 复杂图：a>b, a>c, b>d, c>d (测试推理) 4. 反事实：给定a>b，如果b>a会怎样？ 5. 未见过的关系：定义新关系@，性质..

world model

Marie Curie

Short-Term Memory, STM

Long-Term Memory, LTM

zero hallucination

“When loading and perceiving the world model, it will tell you if it doesn’t know

Structure and Symbols: a stable C++ template system, lightweight and Sublime-compatible, without reliance on heavy IDEs.”

Billing is function-count based on edge structures, not on tokens, liberating structural scalability from token limits.”

A2.5 Stage (Upgraded): Chief Architect

Core Objective: To “sweep” all current coding tools on the market (e.g., CodeX, Claude code, Cursor).

Key Capabilities and Functions:

1. System-Level “World Model”

  • Capability Description: Possesses the ability to process a context of tens of millions of Tokens, enabling the complete and holistic analysis of complex projects containing hundreds of thousands of lines of code in a single pass.
  • Technical Advantage: Our “graph-structured” world model, with its graph traversal inference process, is “not heavily reliant on GPUs.” This means we can achieve a more profound analysis on extremely low-cost cloud servers (such as the L4 ) than competitors can on expensive clusters.

2. Cross-Domain “Integration” Capability

  • Capability Description: The AI’s “brain” must contain a first-principles understanding of a complete application stack. This is a mandatory capability for the A2.5 stage, which includes:
    • Network Protocols: Understanding the internal logic of communication protocols such as HTTP/2, gRPC, and WebSocket.
    • Encoding Formats: Understanding the structure of data interchange formats like JSON, Protocol Buffers, and FlatBuffers.
    • Databases: Understanding SQL queries, database schemas, and the principles of Object-Relational Mapping (ORM).

3. Preliminary “Reverse Engineering” Capability

  • Capability Description: To understand a complete “engineering project,” the AI cannot only look at source code. It must begin to understand compiled, source-unavailable “black boxes.”
  • Specific Functions: At the A2.5 stage, the AI needs to be able to:
    • Read and understand a .h header file to infer the interface and functionality of a closed-source .dll or .so library.
    • Perform preliminary pattern recognition and structural inference on simple binary data formats.

The meaning of 100% autonomous creation.”

2026 q1 10-04

交易china

bili bidu

交易尾部风险

区域银行的尾部，泡沫估值的尾部

交易痛苦估值

例如特斯拉 痛苦的估值 470-410

交易宏观波动

gold sliver dax nik225

交易人工智能

soun nvda amd mdb orcl wdc ———– energy oil natural gas

明年100% 电力资源会不够

做好准备