Skip to content
AGI Study Notes
HomeFoundationsInterview prepEngineering practiceTech historyCareer
中文
EN / 中

English or 中文

Use 中文 / EN above to switch the same note.

点上方的 中文 / EN,切换同一篇笔记。

Interview prep
Find a route for your goal →
  • Study map3 notes
      • Foundations catalog & guide
      • Find your route
      • Coverage & gaps
  • 01 · Understand models
  • Foundations21 notes
    • ›Learning path4
      • Learning map
      • Study & review guide
      • Transformer
      • Transformer lab
    • ›01 · Inputs & sequences7
      • Linear to neural
      • Tokenization
      • Tokenizer algorithms
      • Embeddings & similarity
      • RNN & LSTM
      • Seq2Seq
      • BPTT & gates
    • ›02 · Build a Transformer8
      • Overview
      • Multi-head attention
      • Residual connections
      • Normalization
      • FFN & SwiGLU
      • Vanilla Transformer
      • BERT
      • Decoder-only
      • Labs from scratch
      • Component reading map
  • Architectures & model families26 notes
    • ›Reading model reports9
      • Model family deep dives
      • Reading a model report
      • Reading GPT
      • Reading Llama
      • Qwen 1–2.5
      • Qwen deep dive
      • DeepSeek deep dive
      • DeepSeek-V4: compression & training
      • Gemma deep dive
    • ›Components: gates, memory & depth4
      • Gated Attention
      • Gated DeltaNet
      • Engram
      • Attention Residuals
    • ›MoE: experts & routing7
      • Overview
      • Router & top-k
      • Load balancing
      • Fine-grained & shared
      • LatentMoE
      • System cost
      • Review
    • ›Looped: depth & sharing6
      • Overview
      • Why it helps reasoning
      • Adaptive depth
      • Huginn · Ouro · MoR
      • Costs & limits
      • Review
  • Multimodal learning12 notes
      • Multimodal learning
      • ViT: images to tokens
      • CLIP alignment
      • SigLIP & SigLIP 2
      • Vision to language
      • BLIP & Q-Former
      • LLaVA & DeepSeek-VL
      • Qwen-VL: resolution & position
      • Omni: streaming audio & video
      • OCR & context compression
      • Qwen-Image: generation & editing
      • VLM fine-tuning & checks
  • 02 · Train & evaluate
  • Pretraining & training mechanics18 notes
    • ›Training foundations4
      • Generalization & diagnosis
      • Activations & initialization
      • SGD to AdamW
      • Muon
    • ›01 · Objectives, data & updates5
      • Training route
      • LM objective
      • Pretraining pipeline
      • One training step
      • Multi-token prediction
    • ›02 · Precision, memory & parallelism4
      • Precision & memory
      • Distributed training
      • Tensor parallelism, worked through
      • DDP, FSDP2 & training stacks
    • ›03 · Scaling laws & hyperparameters (draft)5
      • Overview
      • Compute-optimal size and data
      • Training past Chinchilla
      • LR, batch size & schedule
      • Review
  • Deep RL22 notes
    • ›Learning route2
      • Start here
      • Q-learning, PPO & DPO
    • ›01 · Foundations6
      • MDPs & Bellman
      • MC, TD & boundaries
      • Policy gradients
      • Actor–Critic & GAE
      • Natural gradient & TRPO
      • Fitted Q & convergence
    • ›02 · Core algorithms7
      • DQN & Double DQN
      • DDPG → TD3
      • SAC & maximum entropy
      • Model-based RL
      • Which model to learn
      • Planning & LQR
      • Planning meets learning
    • ›03 · Data & experiments7
      • Imitation & rewards
      • Offline RL & OPE
      • OPE estimators
      • Exploration & hierarchy
      • Experiments & debugging
      • LLMs & agents
      • Why RL remains difficult
  • Post-training27 notes
    • ›01 · Demonstrations, adaptation & distillation7
      • Overview
      • SFT
      • Rejection sampling & filtering
      • Fine-tuning & LoRA
      • Prompt / Prefix / Adapter
      • LoRA & QLoRA
      • Distillation: answers & distributions
    • ›02 · Preferences, rewards & objectives7
      • Reward models
      • After PPO
      • DAPO: sampling & weighting
      • Clipping & gradients
      • GSPO & ASPO
      • SAO & asynchronous training
      • Verifiable rewards
    • ›03 · The RLHF training process10
      • Overview
      • RL for language models
      • Value & advantage
      • On- vs off-policy
      • Stages & models
      • Reference & Critic
      • Three KL estimators
      • PPO clipping
      • Then: GRPO · DPO · RLVR
      • Evaluation & review
    • ›04 · Experiments & capability retention3
      • Infrastructure
      • One base model
      • Alignment tax
  • Evaluation15 notes
    • ›Evaluation basics7
      • Overview
      • Evaluation stack
      • BLEU, ROUGE, edit distance & perplexity
      • Benchmarks & long-context tests
      • Comparisons & slices
      • Is the metric robust?
      • Evaluation online (draft)
    • ›LLM-as-a-Judge8
      • Overview
      • Writing criteria
      • Scoring modes
      • Probability scores
      • Biases & workflow
      • Calibration & thresholds
      • RAG, agents & memory
      • A minimal workflow
  • 03 · Inference & applications
  • Generation, context & inference10 notes
    • ›01 · Generation & caching6
      • Inference route
      • Decoding
      • KV cache & inference
      • A request through the service
      • vLLM & SGLang
      • DFlash
    • ›02 · Long context & efficient attention4
      • RoPE & long context
      • NoPE & sequence order
      • FlashAttention & paging
      • MLA & sparse attention
  • Data, memory & retrieval12 notes
    • ›Data & feedback3
      • Data & feedback
      • What data teaches
      • Feedback to objectives
    • ›Representation & memory2
      • Representation & memory
      • Memory lifecycle
    • ›Search & retrieval7
      • Search & retrieval
      • TF-IDF & BM25
      • Dual encoders & objectives
      • Qwen Embedding & BGE-M3
      • Vector indexes: IVF, PQ & HNSW
      • Hybrid retrieval & reranking
      • RAG: evidence & evaluation
  • Agents9 notes
      • Overview
      • Prompting, ICL & CoT
      • Building blocks
      • One agent or several (draft)
      • Tools, MCP & skills
      • Deep Research: questions & evidence
      • Settings
      • Model selection & routing
      • Review
  • AI systems & experience6 notes
    • ›Systems4
      • Overview
      • Agent observability
      • Human in the loop
      • Decision support (draft)
      • Model experience
      • Personal AGI
  • AI infrastructure (draft)16 notes
    • ›From silicon to product10
      • Overview: one request
      • How chips are made
      • CPU architecture
      • GPU architecture
      • NVIDIA generations
      • CUDA
      • Performance: roofline & MFU
      • The framework layer
      • Kubernetes & GPU clusters
      • Review
    • ›Quantization (draft)6
      • Overview
      • Formats & error
      • Weight-only & PTQ
      • FP8 & KV-cache quantization
      • Measuring the loss
      • Review
  • Perception, robots & driving (draft)14 notes
      • Start here
    • ›Perception & geometry6
      • Overview
      • From 3D point to pixel
      • Estimation & tracking
      • Detection & segmentation
      • Geometry foundation models
      • Reconstruction & rendering
    • ›Autonomous driving4
      • Overview
      • BEV & occupancy
      • Forecasting & planning
      • End-to-end & closed loop
    • ›Robot learning3
      • Overview
      • Learning from demonstrations
      • VLAs & sim-to-real
  • 04 · Frontier watch
  • Frontier31 notes
      • Start here
    • ›World models7
      • Overview
      • Learning in imagination
      • Predicting representations
      • Video you can act in
      • Robots & driving
      • Do they understand?
      • Review
    • ›Recursive self-improvement5
      • Overview
      • Generate, verify, retrain
      • Code, not weights
      • Measuring AI speeding AI
      • Review
    • ›Test-time compute (draft)5
      • Overview
      • Sampling, voting & verifiers
      • Training long chains
      • Comparing & reading numbers
      • Review
    • ›Agent RL (draft)4
      • Overview
      • Environments & rewards
      • Credit & training systems
      • Review
    • ›Interpretability (draft)5
      • Overview
      • Features & superposition
      • Sparse autoencoders
      • Circuits & interventions
      • Review
    • ›Diffusion LMs (draft)4
      • Overview
      • What masked diffusion trains
      • Steps, caching & speed
      • Review
  • Start here9 notes
      • Preparation home
      • Interview speedrun
      • Starting out
      • Study plan
      • Preparing for RE / MLE interviews
      • Code comprehension
      • Five short exercises
      • Python idioms
      • Objects & calls
  • ML / LLM fundamentals5 notes
      • Fundamentals review guide
      • Basics
      • ML math
      • Transformer follow-ups
      • Question bank
  • ML coding8 notes
    • ›PyTorch: tensors to training6
      • PyTorch reading path
      • Tensors & storage
      • Operations & shapes
      • Cross-entropy & loss
      • Autograd
      • Batches to training loops
    • ›From formulas to code2
      • Start with code
      • Whiteboard coding
  • Algorithm patterns15 notes
      • LeetCode pattern map
      • Complexity & tools
      • Hash maps & prefix sums
      • Pointers, windows & search
      • String matching · KMP
      • DFS / BFS walkthrough
      • Binary search trees · BST
      • Stacks, heaps & lists
      • Monotonic queues
      • Backtracking & DP
      • Knapsack & coin change
      • Sequence DP · LCS / LIS
      • State-machine DP
      • Sorting, intervals & greedy
      • Primality & sieves
  • Start here
      • Overview
  • An ML system end to end3 notes
      • 1 · System map & ownership
      • 2 · Data, training & artifacts
      • 3 · Serving, diagnosis & feedback
  • Build something small5 notes
      • Start here
      • 1 · This site & CI/CD
      • 2 · A full-stack reading list
      • 3 · Evaluation reports & CI
      • 4 · What to build next
  • System design5 notes
      • From needs to design
      • A recommendation feed
      • A RAG knowledge base
      • An assistant with memory
      • Reading & trade-offs
  • Inside recommender systems10 notes
      • Series introduction
      • 00 · Architecture map
      • 01 · One request
      • 02 · Candidate retrieval
      • 03 · Ranking & diversity
      • 04 · Modern recommenders
      • 05 · Metrics & experiments
      • 06 · Why two towers
      • 07 · Component choices
      • 08 · Serving lifecycle
  • A RAG knowledge assistant3 notes
      • Project introduction
      • Documents & retrieval
      • Evidence & evaluation
  • Post-training experiments7 notes
      • Project introduction
      • 1 · SFT & frameworks
      • 2 · Data to batches
      • 3 · Data & objectives
      • 4 · GPUs & training cost
      • 5 · Save & resume
      • 6 · Experiments & release
  • How the field developed2 notes
      • Technology tree
      • RL timeline
  • Job search7 notes
      • Read this first
      • What to prepare
      • 25-26 internship search
      • 26-27 new-grad search 🚧
      • One interview
      • Mindset
      • Metrics
Chapter overview Preparation guide

Interview preparation

1 min readNote 1 of 9

ML / LLM fundamentals, coding practice, algorithms, and system design. Pick the part you need; there is no need to start from the beginning.

Pick an area

Find priorities for your role ↗
01 / CONCEPTSML / LLM fundamentalsExplain mechanisms, assumptions, and tradeoffs 02 / IMPLEMENTATIONML codingTensors, training loops, attention, and tests 03 / ALGORITHMSPython & LeetCodeLanguage basics, complexity, and reusable patterns 04 / DEBUGGINGCode comprehensionRead an implementation, find a problem, verify a fix 05 / DESIGNSystem designClarify requirements, compare options, handle failure 06 / PROJECTSProjects & experienceYour contribution, the evidence, and what didn't work

These are public concepts and original exercises, not confidential company questions—or a checklist everyone has to finish.

Quick reviewStudy planFull libraryAll learning routesEngineering practiceCareer reflections

How could this page be better?

OutdatedSomething's wrongUnclearSuggest an additionPage problem

Quoting

Select a passage first to quote it.

Opens a prefilled GitHub issue in a new tab. Nothing is public until you confirm there. Needs a GitHub account; leave out personal information.

GitHub opened in a new tab. Confirm there to send it. Thank you!

Edit this page ↗ Write a new note ↗ Page discussions How to contribute

Interview preparation · Review contact: @tianyi-zhang-02

Last updated 2026-10-14 · View source
On this page

Search notes

    ↑ ↓ navigate · Enter open · Esc closeSearch stays on this site