Guides
Last reviewed
May 9, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v3 ยท 3,530 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 9, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v3 ยท 3,530 words
Add missing citations, update stale details, or suggest a clearer explanation.
See also: Terms, Models and Applications
The Guides section of AI Wiki collects how-to articles, explainer pages, reading lists, and reference material that help readers move from a first encounter with artificial intelligence toward practical work with large language models, prompt design, retrieval pipelines, fine-tuning, and safety. Articles in this hub are written for self-directed learners, software engineers, researchers, students, and writers who want a clear path through the wiki's deeper topic pages. Each section below points to existing articles in AI Wiki and indicates which ones are short reference pages and which are longer explainer articles. As of 2026-05-09, AI Wiki contains more than 2,000 entries, and this gateway is updated as new guide-style content lands in the database.
This is a top-level index, not a tutorial in itself. Each linked title takes the reader to a dedicated wiki page covering that topic in depth. The order of sections roughly mirrors a learning path: start with foundational vocabulary and concepts, move to prompt design, then build practical applications with LLMs, then study how models are trained and aligned, and finally consult curated reading lists, podcasts, talks, and conference summaries.
Readers who only want a quick reference can jump to the Acronyms and Abbreviations pages, both of which list common short forms used across AI research and industry writing. Readers who want a structured curriculum can follow the path described in the "Suggested learning paths" table further down the page.
New readers usually need three things first: a working vocabulary, a sense of what current models can and cannot do, and a basic mental model of how a transformer processes text. The articles in this section are written to give that foundation without assuming a research background.
| Article | What it covers | Approximate length |
|---|---|---|
| Artificial intelligence | Definitions, history, and the main subfields of AI | Long explainer |
| Large Language Model | What an LLM is, how training works, common architectures | Long explainer |
| LLMs | Overview of major model families and their use cases | Long explainer |
| Transformer | The neural network architecture behind modern LLMs | Long explainer |
| Attention Is All You Need | Summary of the 2017 paper that introduced the transformer | Long explainer |
| Tokenization | How text is split into tokens before a model sees it | Long explainer |
| Embeddings | Numerical representations of text used by models | Long explainer |
| Context window | The amount of text a model can process at once | Long explainer |
| Acronyms | Glossary of AI-related acronyms | Reference page |
| Abbreviations | Short forms common in AI writing | Reference page |
A reader new to the field can finish Artificial intelligence, Large Language Model, and Transformer in a single sitting and come away with enough vocabulary to follow most other pages on the wiki. The Tokenization and Embeddings entries are useful next reads because almost every later topic, from prompt design to retrieval, depends on understanding how text becomes numbers.
Prompt engineering is the practice of writing inputs to language models so the output matches what the user wants. The wiki covers the core techniques in long-form pages and supplements them with shorter rule-of-thumb articles.
| Article | Topic | Approximate length |
|---|---|---|
| Prompt Engineering | The full survey: zero-shot, few-shot, role prompting, system prompts, more | Long explainer |
| 26 Principles of Good Prompts | Concrete guidelines for prompt phrasing, drawn from the 2024 paper by Bsharat, Myrzakhan and Shen | Short reference |
| Chain-of-Thought Prompting | Asking the model to reason step by step | Long explainer |
| Tree of Thoughts | Search over branching reasoning paths | Long explainer |
| ReAct (prompting) | Interleaving reasoning steps with tool calls | Long explainer |
| Zero shot, one shot and few shot learning | The three basic prompting regimes | Long explainer |
| Reasoning (artificial intelligence) | How current models approach logical and mathematical problems | Long explainer |
| Programming with ChatGPT | Using ChatGPT to write and debug code | Short reference |
| Fine-tune ChatGPT with Perplexity, Burstiness, Professionalism, Randomness and Sentimentality Guide | A short procedure for adjusting ChatGPT output style | Short reference |
| How to Pressure LLMs for Better Output | Techniques for forcing more careful answers from a model | Short reference |
| How to Pressure LLMs | Companion notes on demanding higher-quality responses | Short reference |
Readers who want to start with one article should pick Prompt Engineering. It introduces every term used in the more specialised pages. The 26 Principles of Good Prompts page is a good follow-up because it converts the survey into 26 concrete instructions that can be applied immediately. After that, Chain-of-Thought Prompting and ReAct (prompting) cover the two patterns that most agentic systems rely on.
Once the basics are in place, the next step is wiring a language model into an application. The wiki covers the API mechanics, the agent patterns, and the protocols that have appeared between 2023 and 2026.
| Article | Topic | Approximate length |
|---|---|---|
| GPT API | OpenAI API endpoints, request structure, parameters, code samples | Long explainer |
| Function calling | Letting an LLM call developer-defined functions | Long explainer |
| Tool use (artificial intelligence) | The general pattern of giving a model external tools | Long explainer |
| Model Context Protocol | The 2024 Anthropic protocol for connecting LLMs to data sources | Long explainer |
| Retrieval-augmented generation | Combining a retriever with an LLM | Long explainer |
| Vector database | Storage systems used in retrieval pipelines | Long explainer |
| Agent | What an AI agent is, types of agents | Long explainer |
| AI agents | A broader survey of agent systems and frameworks | Long explainer |
| Agentic AI | Patterns for autonomous, goal-driven LLM systems | Long explainer |
For a developer building a first prototype, a sensible reading order is: GPT API, then Function calling, then Tool use (artificial intelligence), then Retrieval-augmented generation. The Model Context Protocol page covers the standard introduced by Anthropic in 2024 that has since been adopted across multiple model providers and IDEs. The Vector database page lists the storage systems most often used to back retrieval pipelines, including pgvector, Pinecone, Weaviate, Milvus, Qdrant, and Chroma.
Fine-tuning takes a pre-trained model and adapts it to a narrower task or style. The wiki covers both classical full-parameter fine-tuning and the parameter-efficient methods that have become standard since 2021.
| Article | Topic | Approximate length |
|---|---|---|
| Fine Tuning | Full-parameter and parameter-efficient fine-tuning | Long explainer |
| PEFT | Parameter-efficient fine-tuning methods, including adapters and prefix tuning | Long explainer |
| LoRA (Low-Rank Adaptation) | The 2021 method that adapts large models with small low-rank updates | Long explainer |
| QLoRA | LoRA combined with 4-bit quantization, introduced in 2023 | Long explainer |
| Quantization | Reducing model precision to save memory and increase speed | Long explainer |
| Knowledge Distillation | Training a smaller student model to imitate a larger teacher | Long explainer |
| Mixture of Experts (MoE) | Architectures that route tokens through specialist subnetworks | Long explainer |
| Instruction Tuning | Training a model on instruction-response pairs to improve alignment | Long explainer |
| Reinforcement Learning from Human Feedback (RLHF) | The training stage that uses human preferences to shape outputs | Long explainer |
| DPO | Direct Preference Optimization, an alternative to RLHF introduced in 2023 | Long explainer |
| Pre-training | The first large-scale unsupervised training stage | Long explainer |
| Training | Overview of the full training pipeline | Long explainer |
A reader who wants to fine-tune an open-weight model on a single GPU should start with LoRA (Low-Rank Adaptation) and QLoRA. Both pages walk through the rank parameter, the alpha parameter, and the trade-offs between LoRA, QLoRA, and full fine-tuning. The Quantization page is a useful companion because QLoRA is built on 4-bit NF4 quantization. For preference training, the wiki covers both RLHF, the original approach used in InstructGPT and ChatGPT, and DPO, the simpler alternative published by Rafailov and colleagues in 2023.
Safety and alignment are covered as separate but linked topics. Safety articles describe the risks and mitigations; alignment articles describe the technical problem of getting a model to pursue the goals its developers actually intend.
| Article | Topic | Approximate length |
|---|---|---|
| AI safety | The full survey: misuse, accidents, governance, evaluation | Long explainer |
| AI Alignment | The technical alignment problem and approaches to it | Long explainer |
| Reinforcement Learning from Human Feedback (RLHF) | The most widely deployed alignment technique | Long explainer |
| DPO | A simpler preference-tuning method | Long explainer |
| How to Steal ChatGPT-4, GPT-4 and other Proprietary LLMs | Notes on model extraction risks and defenses | Short reference |
| How to Prevent OpenAI and Google From Training Their LLMs on Your Website's Data | Practical steps for site owners using robots.txt and headers | Short reference |
The AI safety and AI Alignment pages are the two most important entry points. They link out to specific topics like reward hacking, deceptive alignment, evaluations, red-teaming, and interpretability. The two short "how to" pages cover concrete defensive measures: one for model owners worried about extraction, and one for website operators who do not want their content scraped for training data.
See also: ChatGPT Guides
This section collects pages aimed specifically at users of ChatGPT and the underlying GPT API.
| Article | Topic |
|---|---|
| ChatGPT Uses | Common applications of ChatGPT |
| Programming with ChatGPT | Using ChatGPT for code |
| Fine-tune ChatGPT with Perplexity, Burstiness, Professionalism, Randomness and Sentimentality Guide | Style adjustment with five named axes |
| GPT API | The developer reference for OpenAI's API |
| 26 Principles of Good Prompts | Prompt rules that apply to ChatGPT and similar models |
The wiki maintains separate index pages for books, podcasts, blog posts, videos, and films. These are linked below with a short note on what each one contains.
| Index page | What it lists |
|---|---|
| Books | Non-fiction and textbook titles, including Deep Learning (Goodfellow, Bengio, Courville, 2016), Superintelligence (Bostrom, 2014), and The Singularity Is Near (Kurzweil, 2005) |
| Podcasts | Episodes from Lex Fridman, Greymatter, Machine Learning Street Talk, and others, with guests including Sam Altman, Demis Hassabis, and George Hotz |
| Blog posts | Long-form essays such as Tim Urban's The AI Revolution, Andrej Karpathy's Software 2.0, and Gwern's The Scaling Hypothesis |
| Videos | Talks and recorded interviews on alignment, scaling, and model interpretability |
| Movies | Documentaries and feature films, including AlphaGo (2017), Her (2013), and 2001: A Space Odyssey (1968) |
| Presentations | Slide decks and conference talks |
The Books entry has been the most stable of these lists. Deep Learning by Ian Goodfellow, Yoshua Bengio and Aaron Courville is still the most widely cited textbook on neural networks. Superintelligence by Nick Bostrom is the standard reference for long-term safety arguments. The Podcasts entry favours episodes that pair a serious technical guest with a long-form interviewer, so the conversations stay substantive across an hour or more.
The wiki indexes the major academic conferences and a small number of industry events. Conference pages cover history, paper acceptance rates, notable papers, and venue information. Talk summary pages cover individual sessions in detail.
| Page | Type | Description |
|---|---|---|
| NeurIPS | Conference | The Conference on Neural Information Processing Systems, the largest annual machine learning research conference |
| ICML | Conference | The International Conference on Machine Learning, founded in 1980 |
| International Conference on Learning Representations | Conference | ICLR, founded in 2013, focused on representation learning and deep learning |
| CVPR (Conference on Computer Vision and Pattern Recognition) | Conference | The leading computer vision venue |
| International Conference on Computer Vision | Conference | ICCV, the second flagship vision conference, held in odd-numbered years |
| European Conference on Computer Vision | Conference | ECCV, held in even-numbered years |
| Dartmouth Conference | Historical | The 1956 workshop where the term "artificial intelligence" was first proposed |
| World Robot Conference | Conference | An annual robotics event held in Beijing |
| A Survey of Techniques for Maximizing LLM Performance (OpenAI Dev Day 2023) | Talk summary | A walkthrough of prompt engineering, RAG, and fine-tuning trade-offs from OpenAI Dev Day 2023 |
| The New Stack and Ops for AI (OpenAI Dev Day 2023) | Talk summary | An operations-focused talk on running LLM applications in production |
The NeurIPS page is the best starting point for a reader new to academic machine learning research. It explains the review process, lists notable past papers, and gives a sense of what kinds of work get accepted. The Dartmouth Conference page is more historical: the 1956 workshop organised by John McCarthy, Marvin Minsky, Nathaniel Rochester, and Claude Shannon is generally treated as the founding event of the field.
The two OpenAI Dev Day 2023 talk pages are kept because both sessions remain useful even years later. A Survey of Techniques for Maximizing LLM Performance covers a four-quadrant model that puts prompt engineering, retrieval, and fine-tuning on the same axes, and it is still widely cited as a quick way to think about which technique fits which problem.
The table below groups guides into three learning paths. Readers do not have to follow them in order, but each row works as a self-contained sequence.
The table below shows the longest guide-style explainer pages currently on the wiki, grouped by the topic they cover. Word counts are taken from the database as of 2026-05-09.
| Topic | Article | Word count |
|---|---|---|
| Retrieval | Vector database | 9,869 |
| Training | Knowledge Distillation | 9,031 |
| Architecture | Mixture of Experts (MoE) | 8,966 |
| API | GPT API | 8,651 |
| Protocol | Model Context Protocol | 8,573 |
| Optimization | Quantization | 8,568 |
| Fine-tuning | LoRA (Low-Rank Adaptation) | 8,308 |
| Tokens | Tokenization | 8,098 |
| Safety | AI safety | 8,034 |
| Alignment | RLHF | 8,021 |
| Fine-tuning | PEFT | 7,906 |
| Agents | AI agents | 7,850 |
| Tools | Tool use (artificial intelligence) | 7,839 |
| Prompting | Prompt Engineering | 7,636 |
| Fine-tuning | Fine Tuning | 7,634 |
| Embeddings | Embeddings | 7,530 |
| Alignment | DPO | 7,442 |
| Agents | Agent | 7,261 |
| Models | LLMs | 6,957 |
| Models | Large Language Model | 6,862 |
| Architecture | Transformer | 6,745 |
| Fine-tuning | QLoRA | 6,309 |
| Prompting | Chain-of-Thought Prompting | 5,891 |
| Alignment | AI Alignment | 5,805 |
| Tools | Function calling | 5,675 |
Ranking the pages this way is useful for two reasons. First, the longest articles tend to be the most thoroughly cross-linked, so they double as hubs into smaller related pages. Second, when several articles in a topic appear in the table, that topic is well covered in the wiki and a reader can rely on it for serious study.
This page is rebuilt whenever new guide-style articles are added to the wiki, when existing articles cross significant length or quality thresholds, or when a new technique becomes prominent enough to deserve its own row in the tables. The lists are derived from queries against the wiki's PostgreSQL database. Word counts and titles are pulled directly from the pages table, so they reflect what is actually published rather than what was planned.
If an article is missing from this index but exists on the wiki, that usually means one of three things: the article does not yet have the Guides category, the article is shorter than the cut-off used in the table, or the article belongs to a more specialised gateway page such as Models, Applications, or Terms. Readers who notice gaps can check those gateways before assuming a topic is uncovered.
| Page | What it indexes |
|---|---|
| Terms | Definitions and short reference entries |
| Models | Specific AI models such as GPT-4, Claude, Gemini, Llama, and others |
| Applications | Products and use cases built on AI |
| ChatGPT Guides | A narrower index focused on ChatGPT |