Signal to BuildSIGNAL TO BUILD中文Owner sign in ↗Explore curated knowledge, course notes and methods, with their existing content and sources preserved.
This index lists complete English editions. Additional Chinese notes and bilingual updates remain available in the Chinese index.
Explains Claude Code's Linux Bubblewrap sandbox, file and network boundaries; provides verified minimal examples and in-product checking steps, while retaining historical security analysis and competitor comparisons.
A synthesis of Skill engineering across 7 perspectives. newtype Lao Huang (Super Analyst three-piece set + Sequential Thinking + 7-Stage Workflow) / Wang Kai (three major pitfalls of Skill + writing Skills in code + IFS + SillyTavern 7 stages) / Xiaohongshu vibecoding (advantages over N8N + calling local services + Vibe Marketing + IELTS three-piece set) / Lao D (SKILL.md + references/ + Git version management) / Anthropic official (Skill format = zip bundle = SKILL.md + references + scripts) / **zara Zhang Zala (opinionated distribution school + Centrally fetch locally Remix + Frontend Slides 16k+ stars + Fork > PR + Skill > SaaS)**. Covers the boundaries between Skill vs MCP vs Sub-Agent, rule applicability scope, memory cross-contamination, and the problem that more Skills lead to more conflicts.
3 close readings of Agent research / talks collected by newtype Planet. 05 CAMEL-AI Wendong Fan "Finding the Scaling Laws of Agents" (2025 talk: # Parameters → # Agents analogy / CAMEL + OWL + CRAB + OASIS frameworks / Agentic Data Generation feeding back into training) / 07 "More Agents Is All You Need" (Tencent ICML 2024: brute-force scaling of Agent count via sampling-and-voting, Llama2-13B + 15 agents matches Llama2-70B, 3 dimensions determine gain magnitude) / 08 "ChatQA" (NVIDIA 2024: two-stage instruction tuning, 70B scores 54.14 vs GPT-4's 53.90 across 10 conversational QA datasets, fine-tuned retriever replaces query rewriting to save API costs, "unanswerable" sample training eliminates hallucinations)
Across Claude Code, Skills and Clawdbot, Huang sees coding capability spreading outward: a system can write programs that supply missing tools. Performing an action still differs from delivering it repeatedly, reliably and economically.
Unified index and core reading guide for 5 mainstream large model textbooks. ① Understand AIGC in One Book (2023, 256 pages, general science popularization) ② Large Model Development and Fine-Tuning from Scratch Based on PyTorch+ChatGLM (376 pages, hands-on textbook) ③ Minimalist Introduction to Large Model Application Development Based on GPT-4+ChatGPT (177 pages, O'Reilly, 2024 SDK hands-on ★ deep intensive reading) ④ Large-Scale Language Models: From Theory to Practice (297 pages, Fudan academic school, Zhang Qi, Gui Tao, Zheng Rui, Huang Xuanjing) ⑤ Large Model Black Book Based on GPT-3/4 Transformer NLP (379 pages, Denis Rothman). Covers the complete path from classic NLP theory → Transformer → GPT evolution → Prompt engineering → fine-tuning → application development
How far is a model that answers questions from an application a business can use over time? This 2023 interview breaks that distance into handling information, testing instructions, organizing tools and operating the product after launch. Dify’s ambition includes a workspace through which business colleagues can keep shaping an application.
25 hub wiki entries from the Anthropic engineering blog. Building Effective Agents framework (Workflow vs Agent + 5 patterns) + new Context Engineering paradigm (evolution of prompt engineering) + 4 categories of Harness design (single-shot / long-running single agent / long-running multi-agent / managed-decoupled) + 5 Tool design principles + Agent Skills open standard + Claude Code engineering (sandboxing / auto mode / best practices / 2 postmortems) + MCP evolution (code execution / DXT/MCPB / advanced tool use). 1-sentence index for all 25 posts. The Anthropic perspective is the "official first-party" layer in the [[wiki-ai-agent-architecture]] / [[wiki-claude-code-workflow]] series, cross-checked against third-party observations.
Complete in-depth reading of the 2023 open-access Springer textbook by Gerhard Paaß and Sven Giesselbach. 436 pages / arXiv:2302.08575. 8 chapters covering: Ch1 Classical NLP fundamentals (token / embedding / RNN / CNN) / Ch2 Three major PLM architectures (BERT autoencoder + GPT autoregressive + Transformer seq2seq) + training/regularization/uncertainty/interpretability / Ch3 Six major directions for improving PLMs (pretraining task variants + long-range sparse attention + multilinguality + knowledge injection + model scale + fine-tuning / few-shot / CoT) / Ch4 PLM knowledge evaluation (GLUE / SuperGLUE / BIG-bench + Probing) / Ch5 Information extraction (classification / NER / relation extraction) / Ch6 Text generation (retrieval / QA / translation / summarization / stories / code / dialogue) / Ch7 Multimodality (speech / image ViT / video / control Decision Transformer + GATO / DNA / Protein) / Ch8 Summary and outlook (new paradigm of Foundation Models / economic applications / 5 categories of harms / Grounding / Fast & Slow Thinking / Planning). Serves as the "academic foundation" supplement to [[newtype-agent-architecture-research-2025]] and [[newtype-agent-research-papers]]
A self-contained bilingual highlight of Google’s official Prompting Essentials course. The core is the 5-element prompt framework TCREI (Task / Context / References / Evaluate / Iterate, with the mnemonic Thoughtfully Create Really Excellent Inputs), which frames writing prompts as equivalent to writing a clear brief for a new hire. It covers: 4 iterative fix methods (return to the framework / split long prompts / switch to a similar task / add constraints) and chain-of-density summarization; prompt chaining, which uses the previous step’s output as the next step’s input, distinguished from simple iteration, including chain-of-thought / tree-of-thought variants; reusable prompt skeletons for 6 high-frequency office scenarios (brainstorming / table trackers / meeting notes / data visualization / text-to-image / presentation speaker notes), plus email writing / summarization / tone and style / multimodal prompts / meta-prompts / decoding formulas and jargon; and a human-in-the-loop checklist for responsible AI use (preventing hallucinations / preventing bias / protecting data). It complements [[newtype-prompt-engineering]]: while Lao Huang’s piece takes the contrarian stance that “prompt engineering doesn’t need to be studied” and covers product-grade prompt orchestration, this article provides Google’s official standardized framework and scenario templates, suitable for non-technical PMs to use immediately as a starting point.
32 real website verification tasks, paired comparison of serial and parallel reads, preserving network failures, first-token wait, correctness, and token usage; distinguish tool orchestration from model prefill, and do not promise universal speedup.
How to install and enable a Skill on the Claude.ai web app (simply upload the .md file), and what it means for PMs: A Skill is a preloaded "expert instruction manual" built into Claude. When you submit a task, the system automatically inserts the corresponding expert workflow into the context, so Claude works according to validated best practices instead of guessing. Division of content with newtype-skills-collection: that article covers Skill engineering (how to build a good Skill), while this article only covers the two-step install/enable process from the Claude.ai end-user perspective, plus when to use Skills and common pitfalls.
Organizes a Karpathy reading path by software paradigms, AI gadgets, nanochat, and knowledge workflows, links original practices and talk notes, and distinguishes the author’s own words, Chinese-community extensions, and historical predictions.
Understand the LLM OS analogy, the three software paradigms, and semi-autonomous products through notes on Karpathy’s Software 3.0 talk; provides direct chapter links, primary video entry points, and system-boundary explanations, while preserving existing interpretations.
Cross-author evals topic summary. Anthropic's 4 key lessons (6 percentage point gap from infrastructure noise / eval awareness where models detect testing / take-home test lifespan / pass^k math) + academic school (full set of SWE-Bench / Terminal-Bench / τ2-Bench / BrowseComp / 𝜏2 / OSWorld / CORE-Bench) + practical perspectives from Wang Kai / Lao Huang / Cal Rueb. Framework selection (Harbor / Braintrust / LangSmith / Langfuse / Arize) + 14-item PM self-check checklist. This wiki is the "horizontal cross-author expansion" of [[anthropic-engineering-demystifying-evals]] (Phase A anchor / Anthropic single-source deep dive)
Comprehensive guide to LLM distributed training. llm-interview-distributed (9 PDFs / distributed overview + pipeline parallelism + DataParallel/DDP + torch.multiprocessing + AMP mixed precision + DeepSpeed + accelerate + ZeRO) + vLLM / inference optimization stack. Covers 4 major parallelism strategies (DP/MP/PP/TP), 3D Parallelism, three stages of ZeRO, and the mainstream Megatron + DeepSpeed stack
Synthesis of LLM fine-tuning and alignment. llm-interview-finetune (19 PDFs / fine-tuning basics + PEFT + LoRA family + QLoRA + continued pretraining + RLHF + PPO + DPO + VRAM optimization + inference + Prompting + SFT data) / newtype fine-tuning practice + Wang Kai's LoRA application notes. Covers full-parameter vs PEFT, LoRA + QLoRA, RLHF vs DPO, SFT data construction, and VRAM optimization strategies
Comprehensive overview of LLM core fundamentals. Covers llm-interview-basics (three Decoder architectures: prefix/causal/encoder-decoder + Layer Norm LN/RMS/Deep + activation functions GeLU/SwiGLU + Attention upgrades MQA/GQA/Flash + Transformer operations + loss/similarity functions + repetition problem + 8 major mitigation strategies)/ Karpathy nanochat (reproducing LM in 243 lines of code)/ newtype Karpathy (Software 2.0 neural network weights). Covers three Transformer paradigms, mainstream LLM configurations (LLaMA/GPT/GLM), Attention evolution, and training experience
Complete illustrated 8-PDF set for LLM large model distributed training interviews. Covers distributed training for large models (overview) + Pipeline Parallelism + nn.DataParallel + nn.parallel.DistributedDataParallel + torch.multiprocessing + AMP mixed precision training + DeepSpeed + accelerate + ZeRO learning. Systematically covers the three mainstream distributed training stacks: PyTorch, DeepSpeed, and Megatron
15 special-topic PDFs for the fine-tuning / RLHF / PEFT / continued pre-training section in LLM interviews. Covering LLM fine-tuning interview questions + PEFT (Parameter-Efficient Fine-Tuning) + Adapter-tuning + LoRA series + QLoRA + PEFT library + continued pre-training + continued pre-training sample concatenation + LoRA secondary pre-training + evaluation interview questions + reinforcement learning interview questions + RLHF and its variants + PPO + NLP applications of reinforcement learning + training set interview questions + LLM-generated SFT data + GPU memory optimization + inference interview questions + Prompting section
A full collection of 14 special-topic PDFs on RAG (Retrieval-Augmented Generation) for LLM interviews. Covers RAG fundamentals (why an external knowledge base is needed / the R/A/G three modules) + LangChain RAG hands-on practice + large model external knowledge bases + RAG experience-based interview questions + PDF parsing + table recognition + text chunking + large model-assisted recall (HYDE) + hard negative sample mining + RAG evaluation + RAG optimization strategies + key pain points of RAG + RAG-Fusion + Graph RAG + 8 optimizations for long-term memory in multi-turn conversations