Profile: how to choose your path

For people who define AI products and do not build the systems themselves. How it differs from FDE / AI Engineer: L5 goes up to "high", because judging which frontier will become a product is the job. L4 stays, but only to "know the cost and failure modes of each part". L2 stops at these two sections: LLM: how the model works and Calling model APIs: token, streaming, tools, cost.

Path: LLM: how the model works → Calling model APIs: token, streaming, tools, cost → Prompting: instructions you can measure → Evals: knowing whether it got better (the engineering concept a PM most needs to understand: a requirement without an eval cannot be accepted) → RAG: look up first, then answer → Agent: a model with tools in a loop → Cost: every call has a price tag → Safety: the model acts on untrusted text, then AI history: how to read the three waves → Weekly reads: the frontier feed (≤ 5) → Look up when needed: frontier reference list (≤ 15) → Role family: how AI jobs are divided. On the site, this maps only to the first two stages of the foundations track.

For senior engineers who already tune models in production, and who need to turn LLM features into a product and be accountable for the results.

Why the weights are set this way: L4 is highest. RAG, agent, MCP and evals are quoted straight from the JD, and interviews ask about systems you have built. L2 is high. Interviews keep pushing down to "why", so you need to implement GPT from scratch and explain the path from pretraining to alignment. L3 is medium. Stop at being able to talk through serving system design and having deployed one once. You don't write CUDA. L6 is high, but only in the sprint. L0 / L1 as needed.

The path (23 stops, in order; each stop is a node):

1. Set up the big picture first: LLM：模型怎么工作 → 调用模型 API：token、流式、工具、成本 → Prompting：能度量的指令 → 结构化输出：让模型的回答能被程序消费.
2. Build the first production system: Embedding：把意思变成向量 → RAG：先查再答 → Agent：循环里带工具的模型 → MCP：给每个系统一个统一的插口.
3. Make it measurable and operable: Evals：知道它有没有变好 → 可观测性：知道线上发生了什么，并能回到评测 → 成本：每一次调用都有价签 → 安全：模型会照着不可信的文本行动.
4. Go down and fill in the "why": Transformer → 预训练与 scaling → 微调：教它你的任务 → 后训练：用奖励塑形.
5. Go down and fill in the systems: 解码与推理 → Serving：自己跑模型 → Kubernetes：在客户的地方跑.
6. Go up to the industry and the career: AI 历史：怎么读三次浪潮 → 每周必看：前沿信息流（≤ 5） → 岗位族：AI 岗位怎么分 → 行为面与价值观轮.

How this maps to the site tracks: Step 1 = the first two stages of the foundations track. Steps 2–3 = ai-engineer track "Build LLM features" and "Measure and improve" + fde track "Build for a customer" and "Prove it". Step 4 = ai-engineer track "Train and post-train" + fde track "Adapt the model". Step 5 = ai-engineer track "Serve" + fde track "Deploy". The site tracks have no step 6 (L5 / L6 appear only on the layer pages of the reading guide, not in the skill guide or the tracks).

For people building inference engines, GPU performance, and training platforms. It is the exact inverse of FDE / AI Engineer: L3 is highest, L4 is lowest. From L2 you only need the "model internals" half (architecture, decoding, MoE), not post-training.

Path: Linear algebra → PyTorch: train a network yourself → Transformer → Decoding and inference → Long context and MoE, then all of L3: GPU basics: why it, and where the bottleneck is → Serving: running models yourself → Parallel training: why one card is not enough → Compute economics: where training money goes, how inference is priced → Kubernetes: running it at the customer's site. This path has the thinnest resources in this library (the L3 candidates only reach "can explain it" depth, and CUDA and kernel optimization have no nodes). The site has no matching track. If you really take this path, first fill in GPU MODE and the CS336 systems lectures (see the Not picked entries and recommendations in GPU basics: why it, and where the bottleneck is).

For people who train models, build recommenders or do MLOps. How it differs from FDE / AI Engineer: L1 goes from “medium” to “high”: training, generalization and metrics are daily work. L4 drops to “medium”: you only need to know how to use RAG and agent. L0 goes up to “medium”: you must be able to derive optimization and probability, not just follow them.

Path: lay the foundation first (Math: fill in when needed → Linear algebra → Probability and statistics → Optimization), then go from classical ML to deep learning (Classical machine learning: learning from samples → Loss and backpropagation → Generalization → Classic architectures: CNN, RNN / LSTM, residuals → PyTorch: train a network yourself), then move to the training side of LLM (Tokenizer → Transformer → Pretraining and scaling → Fine-tuning: teach it your task → Post-training: shaping with rewards), close with Evals: know whether it got better, and finally fill in systems (Parallel training: why one card is not enough → Serving: run the model yourself). On the site this maps to the full foundations track + the last two stages of the ai-engineer track.