AI HOT CONCEPTS HANDBOOK
Concept maps · How the tech works · Case studies · Paper deep-dives · Engineering practice · Career paths
50+ pages aren't a library to read cover to cover — they're a route map you can combine as needed
Interviewing in 1–3 months: go straight to JD benchmarking, resume rewrites, and the interview question bank — every page is mapped to what big tech actually tests.
View the route →Have time and want solid fundamentals: work week by week from the guide to hands-on projects, with a clear checkpoint at the end of each week.
View the route →Working builders shouldn't read this cover to cover: use the problem → page index to look up exactly what you need and get back to work.
View the route →If you only read ten pages, read these
AI, machine learning, deep learning, generative AI, and agents—five headline terms whose containment hierarchy and boundaries this page clarifies with an overview table, a layered stack diagram, four common confusions, and a job-market skill-stack comparison.
GuideA top-down dissection of an AI application into five layers — application, capability, model, architecture, and infrastructure — reconnected through the dual lenses of data flow and the training, post-training, and inference lifecycle, making this page the site-wide content map where every layer points to pages that go deeper.
GuideThe handbook's 50+ pages are not a linear textbook to read cover to cover but a roadmap you can combine on demand; this article lays out three routes for three kinds of readers — a 2-week job-hunting sprint, an 8-week systematic deep dive, and an always-ready desk reference — topped off with a one-line index of all 51 pages on the site.
GuideThe AI concepts driving today's industry and public conversation—LLMs, Transformers, prompts, RAG, agents, multimodal and diffusion models—mapped with a city metaphor, a why-now backstory, a four-layer framework, core relationship chains, and a guide to using the handbook.
GuideAI agents are systems with an LLM as the "brain" that can plan autonomously, call tools, remember, and act to complete multi-step tasks — this page covers how they differ from chatbots, their core components, a typology, and the engineering and evaluation challenges.
Core ConceptsFine-tuning continues training a pretrained model for new objectives while PEFT updates only a small share of parameters; this article explains fine-tuning vs. prompting vs. RAG, the spectrum from full fine-tuning (FFT) to QLoRA, the low-rank decomposition behind LoRA, instruction data, and the complete end-to-end workflow.
Core ConceptsThe core engine of the generative AI wave — how LLMs learn through next-token prediction, the three-stage training pipeline, emergent abilities, architecture families, leading models, hard limits, and the trade-offs between prompt, RAG, and agent approaches.
Core ConceptsLLM evaluation is the discipline of systematically measuring a model's capabilities, behaviors, and risks — from benchmarks like MMLU and GSM8K to LLM-as-a-judge and red-teaming, this article explains why benchmark scores don't equal real capability and how to turn evaluation into an engineering feedback loop.
Core ConceptsThe craft of guiding LLM output through carefully designed instructions and examples — the first language of working with LLMs — covering role prompting, chain-of-thought, ReAct, and when to choose prompts over fine-tuning or RAG.
Core ConceptsRAG is the architecture that has an LLM retrieve external knowledge first and then answer from what it found — this page walks through the full offline-indexing and online-retrieval pipeline, maps the variant family, compares fine-tuning and long context, and distills hard-won production lessons.
Core ConceptsThe 2017 Transformer — an architecture built entirely on attention and the foundation of nearly every modern large model — broken down into QKV attention, multi-head, positional encodings, the three architecture variants, O(n²) complexity, and the KV cache.
Core ConceptsReleased on 2022-11-30, ChatGPT hit one million users in 5 days and 100 million in 2 months, turning the large language model from a research toy into a mass-market product, and this article unpacks its technical evolution, product design, industry impact, and controversies to distill the lesson that productization, more than raw model capability, decides success.
Case StudiesIn January 2025, DeepSeek-R1 took on OpenAI's o1 as an open-source model and turned "reasoning models" from a lab term into a global hotspot; this article breaks down the chain-of-thought mechanism, R1's reinforcement learning and distillation techniques, the evolution of reasoning models, and their industry impact, limitations, and controversies.
Case StudiesFrom Attention Is All You Need to DPO, this is a paper-by-paper deep dive into the ten classic papers that shaped modern AI — covering the problem each solved, its core method, key results, historical impact, and the traps hidden inside it — so you can build a complete map from "reading papers" to "understanding the principles" to "putting them into practice."
Paper Deep-DivesWhy AI's hottest concepts demand reading the papers — this page explains the first-hand versus second-hand difference between concept pages and papers, lays out the module's four-step method of map, deep-dive, frontier, and FAQ with a quick scan of the paper list, and points each type of reader to their first step.
Paper Deep-DivesBuild a runnable personal-knowledge-base RAG system from scratch in Python — document loading and chunking, embedding-based indexing, hybrid vector and BM25 retrieval, and prompt assembly with generation — with fully runnable code plus tuning, evaluation, common pitfalls, and next steps.
Hands-On GuidesBuild, from scratch, an agent that can call tools to complete tasks — function definitions and function calling schemas, the ReAct tool loop, memory, and planning — all threaded through a "check the weather and add it to the schedule" example, with framework selection, engineering essentials, evaluation, common pitfalls, and directly runnable code.
Hands-On GuidesAI applications rarely fail because the model isn't strong enough — they fail on high-frequency pitfalls across four stages (selection and understanding, engineering implementation, evaluation and launch, and organizational process), so this page collects 14 anti-patterns and 3 real-world failure case studies with a review-ready self-check checklist.
Hands-On GuidesA panoramic guide to interviews for in-demand AI roles with high-frequency real questions explained — nine knowledge modules and roughly 50 questions covering Transformer/LLM principles, prompting, RAG, agents, fine-tuning and alignment, inference and deployment, and evaluation, plus end-to-end scenario questions and tactics for live coding, project deep-dives, and asking the interviewer good questions.
Careers & Job MarketA resume for a trending AI role is not a laundry list of experiences but a chain of evidence aimed at the target job, and this article covers the most common ways resumes die, a four-quadrant benchmarking table (model theory / engineering / evaluation / domain understanding), the context-action-result project formula, how to build personal projects from zero, and how emphasis differs across algorithm, application, and product roles.
Careers & Job MarketA quick-reference guide to models and leaderboards — comparison tables of the major closed- and open-source LLMs as of mid-2025, how to read leaderboard rankings without being misled (rank is not suitability), scenario-based selection guidance, plus embedding models, vision models, and where to get them.
ResourcesBrowse by section, or search directly
From the 1950 Turing Test to the 2025 era of large language models and agents, AI has weathered three booms and two winters; this article uses a single timeline of milestones to make sense of 75 years of technical evolution, plus three threads that help you remember the years that really matter.
AI, machine learning, deep learning, generative AI, and agents—five headline terms whose containment hierarchy and boundaries this page clarifies with an overview table, a layered stack diagram, four common confusions, and a job-market skill-stack comparison.
A top-down dissection of an AI application into five layers — application, capability, model, architecture, and infrastructure — reconnected through the dual lenses of data flow and the training, post-training, and inference lifecycle, making this page the site-wide content map where every layer points to pages that go deeper.
The handbook's 50+ pages are not a linear textbook to read cover to cover but a roadmap you can combine on demand; this article lays out three routes for three kinds of readers — a 2-week job-hunting sprint, an 8-week systematic deep dive, and an always-ready desk reference — topped off with a one-line index of all 51 pages on the site.
The AI concepts driving today's industry and public conversation—LLMs, Transformers, prompts, RAG, agents, multimodal and diffusion models—mapped with a city metaphor, a why-now backstory, a four-layer framework, core relationship chains, and a guide to using the handbook.
AI agents are systems with an LLM as the "brain" that can plan autonomously, call tools, remember, and act to complete multi-step tasks — this page covers how they differ from chatbots, their core components, a typology, and the engineering and evaluation challenges.
AI safety and governance is about identifying, measuring, and mitigating the risks of AI systems so that AI stays controllable, trustworthy, and accountable — this article covers a four-layer risk taxonomy, six technical countermeasures, the global regulatory landscape, and a practical checklist for engineers.
Alignment is the process of making a model's capabilities serve human intent — being capable, and behaving the way people want — and this article breaks down the three stages of RLHF, DPO's implicit reward optimization, Constitutional AI, and other alignment techniques, along with the alignment tax and reward hacking.
Diffusion models learn data distributions by forward noising and reverse denoising and now dominate text-to-image and text-to-video generation; this page covers DDPM, latent diffusion, conditioning, and sampling acceleration, compares GAN/VAE/autoregressive routes, and maps applications and risks.
Fine-tuning continues training a pretrained model for new objectives while PEFT updates only a small share of parameters; this article explains fine-tuning vs. prompting vs. RAG, the spectrum from full fine-tuning (FFT) to QLoRA, the low-rank decomposition behind LoRA, instruction data, and the complete end-to-end workflow.
Train once, infer countless times — inference is the biggest bill an LLM application will ever pay; this article breaks down the core inference optimization techniques, including KV Cache, INT4 quantization, distillation, speculative decoding, continuous batching, and FlashAttention, and closes with a serving framework comparison and an end-to-end tuning path.
Entity-relation-property triples turn world knowledge into a reasoning-ready graph structure; this page covers knowledge graph fundamentals, a head-to-head comparison with vector databases, and the four KG×LLM integration patterns (GraphRAG, knowledge injection, KGQA, LLM-assisted construction) plus their engineering challenges.
The core engine of the generative AI wave — how LLMs learn through next-token prediction, the three-stage training pipeline, emergent abilities, architecture families, leading models, hard limits, and the trade-offs between prompt, RAG, and agent approaches.
LLM evaluation is the discipline of systematically measuring a model's capabilities, behaviors, and risks — from benchmarks like MMLU and GSM8K to LLM-as-a-judge and red-teaming, this article explains why benchmark scores don't equal real capability and how to turn evaluation into an engineering feedback loop.
How AI models come to understand and generate text, images, audio, and video — modality encoders, CLIP-style alignment, Q-Former fusion, and unified token spaces — plus the leading models from GPT-4o to Qwen-VL, and multimodal hallucination, benchmarks, and applications.
The craft of guiding LLM output through carefully designed instructions and examples — the first language of working with LLMs — covering role prompting, chain-of-thought, ReAct, and when to choose prompts over fine-tuning or RAG.
RAG is the architecture that has an LLM retrieve external knowledge first and then answer from what it found — this page walks through the full offline-indexing and online-retrieval pipeline, maps the variant family, compares fine-tuning and long context, and distills hard-won production lessons.
The 2017 Transformer — an architecture built entirely on attention and the foundation of nearly every modern large model — broken down into QKV attention, multi-head, positional encodings, the three architecture variants, O(n²) complexity, and the KV cache.
Vector databases store embeddings and retrieve by semantic similarity via approximate nearest neighbor search, forming the storage backbone of RAG systems; this page covers embedding basics, semantic vs. keyword search, similarity metrics, ANN indexes, product selection, and engineering pitfalls.
From its dominant CASP14 win to a database of 200M+ structures and a Nobel Prize in Chemistry, this article breaks down AlphaFold's MSA + Evoformer technical approach and charts the AI for Science paradigm shift, including AlphaFold3 and RFdiffusion.
Released on 2022-11-30, ChatGPT hit one million users in 5 days and 100 million in 2 months, turning the large language model from a research toy into a mass-market product, and this article unpacks its technical evolution, product design, industry impact, and controversies to distill the lesson that productization, more than raw model capability, decides success.
In January 2025, DeepSeek-R1 took on OpenAI's o1 as an open-source model and turned "reasoning models" from a lab term into a global hotspot; this article breaks down the chain-of-thought mechanism, R1's reinforcement learning and distillation techniques, the evolution of reasoning models, and their industry impact, limitations, and controversies.
From Copilot's 2021 preview to Cursor and Claude Code, AI coding assistants are reshaping software development — this piece breaks down how code LLMs are pretrained and complete code, HumanEval-style evaluation, RAG-based repository understanding, and agentic development, while critically examining the productivity evidence, code quality, and copyright controversies.
Manus's March 2025 launch ignited a boom in general-purpose agents; this article breaks down the product form, technical foundations, and competitive landscape of autonomous task agents, and examines the controversies around their success rates, costs, and safety.
Using Midjourney as an entry point to unpack the 2022-2023 text-to-image explosion: the diffusion-model and CLIP foundations behind the three-way race of DALL·E 2, Stable Diffusion, and Midjourney, the product iteration methodology, the open-source ecosystem, and the copyright controversy.
How Perplexity turned Retrieval-Augmented Generation into a mass-market product: the mechanics of its "search + citations" conversational answer engine, the paradigm gap with traditional search, and AI search's impact on the content ecosystem and traffic distribution.
From collaborative filtering to LLMs, this article reviews the traditional recommender system stack, details four paths for integrating LLMs—feature understanding, conversational, agent-based, and generative—and analyzes the industrial challenges of latency, cost, explainability, and evaluation.
A retrospective on how OpenAI's Sora, from its early-2024 teaser to its official launch, ignited the text-to-video explosion, covering video diffusion models and the DiT architecture, the spacetime patching technical foundation, the product landscape spanning Sora, Kling, Runway, Veo, and Pika, and the physical limitations, deepfake governance, and evaluation challenges that remain.
Speech AI's "listening" and "speaking" — from OpenAI's open-source Whisper ASR to neural TTS and voice cloning, unpacking Transformer audio encoding, autoregressive and diffusion-based speech synthesis, and surveying the product landscape, agent applications, and the risks of voice-cloning fraud and privacy.
From Attention Is All You Need to DPO, this is a paper-by-paper deep dive into the ten classic papers that shaped modern AI — covering the problem each solved, its core method, key results, historical impact, and the traps hidden inside it — so you can build a complete map from "reading papers" to "understanding the principles" to "putting them into practice."
Why AI's hottest concepts demand reading the papers — this page explains the first-hand versus second-hand difference between concept pages and papers, lays out the module's four-step method of map, deep-dive, frontier, and FAQ with a quick scan of the paper list, and points each type of reader to their first step.
Before reading any hyped AI paper, set your discipline — read with questions, apply a three-pass method, watch for red-flag signals, and keep a note template — then work through ten high-frequency Q&As that untangle everyday sticking points like impenetrable math, endless reading lists, English difficulties, and misread benchmarks.
The ocean of papers is vast and deep, and diving in blind will only drive you away — this page lays out five thematic reading routes (LLMs, generative AI, application engineering, alignment and safety, plus a less-but-better fallback), each with a reading order, a self-check for when you finish, and a suggested time budget.
With Attention Is All You Need as the root node, this page lays out four evolutionary branches — large language models, generative models, engineering applications, and alignment — into a searchable paper map covering 20+ key papers from 2014–2025, complete with paper-concept-case mappings and reading priorities.
The eight hottest frontier threads in AI for 2024–2025 — reasoning models, long context, efficient architectures, multimodality, agents, video and world models, safety and alignment, and agentic RL plus the AI Scientist — each broken down by current state and landmark works, with sober judgments on which will last and which are hype, plus a low-cost, low-noise way to track it all.
Build a runnable personal-knowledge-base RAG system from scratch in Python — document loading and chunking, embedding-based indexing, hybrid vector and BM25 retrieval, and prompt assembly with generation — with fully runnable code plus tuning, evaluation, common pitfalls, and next steps.
Build, from scratch, an agent that can call tools to complete tasks — function definitions and function calling schemas, the ReAct tool loop, memory, and planning — all threaded through a "check the weather and add it to the schedule" example, with framework selection, engineering essentials, evaluation, common pitfalls, and directly runnable code.
Build a regression-ready evaluation system for your AI application, starting from business goals and using golden datasets, exact match and semantic similarity, RAGAS, LLM-as-a-judge, and CI integration, with runnable Python code throughout.
Turn a trained LLM into a stable, inexpensive, fast production service, covering the cloud API vs. self-hosting decision, AWQ/GPTQ quantization export, vLLM launch configuration, OpenAI-compatible API integration, Redis caching and rate limiting, TTFT/TPOT monitoring and alerting, plus cost optimization, security and compliance, and troubleshooting common pitfalls.
Fine-tune your own model with Qwen/Llama and LoRA on a consumer GPU: decide whether fine-tuning is warranted, prepare VRAM and data, then follow complete runnable code from QLoRA training and merge/export through inference validation and evaluation to launch.
AI applications rarely fail because the model isn't strong enough — they fail on high-frequency pitfalls across four stages (selection and understanding, engineering implementation, evaluation and launch, and organizational process), so this page collects 14 anti-patterns and 3 real-world failure case studies with a review-ready self-check checklist.
A copy-paste-ready library of prompt templates and a hands-on debugging guide: eight categories of ready-to-use templates covering role prompting, JSON output, few-shot, chain-of-thought, RAG Q&A, and code review, plus a complete debugging methodology for diagnosing failure modes, running A/B comparisons, and versioning prompts.
A panorama of the AI job market — role-by-role comparison of seven job types (LLM algorithm engineer, RAG/Agent application engineer, prompt engineer, AI product manager, AI infra engineer, data & eval engineer, AI safety engineer), a role-to-content mapping table, market trends, and how to use the four career pages.
A panoramic guide to interviews for in-demand AI roles with high-frequency real questions explained — nine knowledge modules and roughly 50 questions covering Transformer/LLM principles, prompting, RAG, agents, fine-tuning and alignment, inference and deployment, and evaluation, plus end-to-end scenario questions and tactics for live coding, project deep-dives, and asking the interviewer good questions.
Turns the high-frequency keywords in job descriptions for hot AI concepts—Transformer, pre-training, prompts, RAG, agents, LoRA, RLHF, quantization, evaluation, and multimodality—into a review checklist you can work through item by item, pairing each with likely interview follow-ups and a page on this site, plus a rundown of the hidden test points that never appear in the JD but always come up in interviews.
A synthesis of open roles at major tech companies in China and abroad across trending AI fields, covering six role types—LLM algorithm engineer, RAG application engineer, Agent engineer, AI Infra/inference optimization engineer, AI product manager, and multimodal algorithm engineer—with 2-3 sample job descriptions dissected per role plus a JD-jargon decoder and a requirements mapping table to help you reverse-engineer a learning plan from job descriptions.
A resume for a trending AI role is not a laundry list of experiences but a chain of evidence aimed at the target job, and this article covers the most common ways resumes die, a four-quadrant benchmarking table (model theory / engineering / evaluation / domain understanding), the context-action-result project formula, how to build personal projects from zero, and how emphasis differs across algorithm, application, and product roles.
A curated resource map for the AI hot-concepts space — six categories covering official docs and APIs, learning courses, must-read books, high-quality blogs, open-source tools and frameworks, and communities and leaderboards — with every entry real and verified at its official source.
A reference catalog of LLM datasets and tools: quick-lookup dataset tables organized by purpose (pretraining, instruction tuning, preference, evaluation, RAG, multimodal), plus tool recommendations for frameworks, vector databases, annotation, and deployment — with copyright and compliance notes.
Core glossary of AI hot concepts: accurate one-sentence definitions of 80+ terms across six categories — model architecture, training & optimization, application paradigms, data & retrieval, generative models, and evaluation & safety — plus a quick-reference table of high-frequency interview abbreviations.
A quick-reference guide to models and leaderboards — comparison tables of the major closed- and open-source LLMs as of mid-2025, how to read leaderboard rankings without being misled (rank is not suitability), scenario-based selection guidance, plus embedding models, vision models, and where to get them.
