Deep technical content explained for humans. From first principles to production systems.
From collaborative filtering to production-grade systems
Vectors, matrices, and transformations — the language of ML
Every ML algorithm, derived from first principles
From word embeddings to GPT and beyond
Teaching agents to learn from experience
Standing on the shoulders of pretrained giants
The mathematics of change
Where physics meets neural networks
Predicting the future from the past
Infrastructure for the modern world
How computers actually work
Algorithms and data structures, one problem at a time
DashScope/Bailian in production: Qwen LLM, Qwen-Omni, Wanxiang video, Qwen-TTS
PAI in production: DSW notebooks, DLC distributed training, EAS serving, Designer/QuickStart
Stitching the seven modules into one repo, running terraform apply once, and watching a complete agent runtime — VPC, ECS, RDS, OpenSearch, OSS, LLM gateway, SLS observability, cost alarms — come up in seven minutes. Real apply output, the module DAG, and the starter repo to fork.
把七个 module 拼到一个仓库，跑一次 terraform apply，看一个完整的 Agent runtime——VPC、ECS、RDS、OpenSearch、OSS、LLM 网关、SLS 观测、成本告警——七分钟内起来。真实 apply 输出、module DAG、可 fork 的起手仓库。
Logs to SLS, traces to ARMS, metrics to CloudMonitor — all provisioned in HCL so a new env comes pre-instrumented. The four alarms that actually catch real incidents and the SLS-driven cost dashboard that tells you which agent is burning your budget before payday.
日志进 SLS、Trace 进 ARMS、指标进 CloudMonitor——全部用 HCL 配，新环境天生带观测。真实救过我项目的四条告警，加上 SLS 驱动的成本看板，发薪日之前告诉你哪个 Agent 在烧预算。
Centralise LLM API access through one gateway: per-agent quotas, request logging, and zero secrets outside KMS. Terraform-provisioned API Gateway plus self-hosted LiteLLM on ECS, with DashScope/OpenAI/Anthropic keys rotating automatically through KMS Secrets Manager.
把所有 LLM 访问收敛到一个网关：按 Agent 限流、请求落 SLS 日志、KMS 之外不留密钥。Terraform 配 API Gateway + ECS 上自托管 LiteLLM，DashScope/OpenAI/Anthropic 的 key 通过 KMS Secrets Manager 自动轮转。
An agent has three kinds of memory and they map onto three Aliyun services: PolarDB/RDS for sessions, OpenSearch (vector edition) or pgvector for embeddings, OSS for artifacts. Real Terraform for each, plus the lifecycle and backup rules that keep the bill flat.
Agent 有三种记忆，分别落到三个阿里云服务上：会话用 PolarDB/RDS，embedding 用 OpenSearch 向量版或 pgvector，产物用 OSS。每一层的真实 Terraform，加上让账单不暴涨的 lifecycle 和备份规则。
The three places an agent's main loop can live on Aliyun: a long-running ECS instance with pm2, a Kubernetes pod on ACK, or a Function Compute invocation. The cost-crossover model I use to pick between them, and a real cloud-init bootstrap that goes from bare Ubuntu to running agent in 90 seconds.
Agent 主循环在阿里云上有三个合理落点：长跑 ECS + pm2、ACK 上的 Kubernetes Pod、Function Compute 触发式调用。我用来选择的成本拐点模型，加一段真实的 cloud-init 脚本，从裸 Ubuntu 到 Agent 运行 90 秒搞定。