Code for the book that walks you through building a GPT-style model in PyTorch

Most people use LLMs as black boxes. This repository is the code companion to the book Build a Large Language Model (From Scratch). It lets you open the box by coding each stage yourself, chapter by chapter, in Jupyter notebooks.

You start with text data and tokenizing, then write attention mechanisms, then assemble a GPT model. Next you pretrain it on unlabeled text and finetune it for classification and for following instructions. The model is small and meant for learning, but the README says the method mirrors how large foundation models are made. The code also shows how to load weights from larger pretrained models for finetuning.

It uses plain PyTorch and no external LLM libraries, so nothing hides behind an API. The main chapters are designed to run on an ordinary laptop and use a GPU automatically if one is present. Bonus folders go further, with topics such as KV cache, grouped-query attention, mixture-of-experts and LoRA. A video course and a free quiz PDF are linked as companions.

Read more on interesting repos · GitHub

Reasoning models like DeepSeek R1 are easy to misread if you only read about them in theory. This repository is the official code for the book Build a Reasoning Model (From Scratch). You start with a pre-trained open base LLM, Qwen3, and add reasoning skills yourself in PyTorch.

The chapters go in order. You load and run the model, then build an evaluation setup on MATH-500. Next come inference-time methods such as chain-of-thought prompting, self-consistency, best-of-N and self-refinement. After that you train with reinforcement learning (GRPO), improve it, and distill reasoning from a larger teacher model.

The code is for learning, so the model is small and the main chapters are meant to run on consumer hardware. It stands alone, though the author's earlier book covers building a base LLM. The repo follows a printed book, so changes to the main chapter code are not accepted as contributions.

Read more on interesting repos · GitHub

Making a good lesson takes time: outlining, building slides, writing quizzes, and finding ways to practise. OpenMAIC (Open Multi-Agent Interactive Classroom) automates that. You give it a topic or attach materials, and it builds a lesson you can play through.

Several AI agents take part in the lesson. Some act as teachers and some as classmates. They speak aloud, draw on a whiteboard, and discuss the material with you in real time. Lessons can include slides, quizzes, interactive HTML simulations, and project-based learning. Version 1.0 added an agent workbench, a chat workspace where an agent plans a course, builds each page, and revises it using your uploaded documents, audio or video.

It is provider-neutral: you bring your own language models, media, search and storage. It is a Next.js and TypeScript app, and the README lists LangGraph among its tools. It also ships a SKILL.md skill, so agent tools such as OpenClaw and Codex can generate classrooms for you. There is a live demo and a Vercel deploy button.

Read more on interesting repos · GitHub

Claude Code has a growing ecosystem of skills, hooks, plugins, guides and side projects, spread across GitHub and blogs. This repo collects the ones the maintainer judged worth your time, with a short note on each.

The README is organized as a table of contents. Sections include Start Here, From Anthropic, Documentation and Learning, Skills, Agent Orchestration, Memory and Context Persistence, Status Lines, Security, Testing and Linting. Most entries give the author, a one or two sentence description, and badges for creation date, last commit, license and stars.

The README says the list is mid-transition. The current version highlights resources that were not on the earlier one. Older entries are being migrated back over time and are kept, without updates, in a README_ALTERNATIVES folder. It says it favors code quality, security and originality, but those are the maintainer's own claims.

Read more on interesting repos · GitHub

AI coding assistants often write code that ignores project patterns, skips error handling or adds technical debt. This playbook treats the AI like a junior engineer: it must propose a plan, wait for approval, and then have its output checked.

The kit is mostly documents. You copy a Cursor rules file and a folder of prompt templates into your project, then add a shell script that runs type checks, linting, circular-dependency detection and tests. Prompts use XML tags to separate the objective, context, constraints, deliverables and an approval gate.

Review has three layers: automated checks, AI-assisted checks of logic and edge cases, and a human look at architecture. The examples lean toward TypeScript, Next.js, tRPC and Prisma. The README's figures, such as 2-3x faster delivery and 50% fewer bugs, come with no supporting data, so treat them as claims.

Read more on interesting repos · GitHub

AI answer engines like ChatGPT, Google AI Overviews, Gemini and Perplexity decide which pages to cite. This list gathers writing and research on how they make that choice, so you can optimize content for them as well as for classic search.

It is a plain list of links, each with a long note on what the source found. The first sections cover how AI search works: reverse-engineered retrieval pipelines, grounding comparisons across vendors, and ad infrastructure. The next section covers research papers, including large citation experiments and brand-visibility audits.

The contents also list sections on content engineering, technical SEO, measurement, Reddit and off-page signals, agentic SEO, open source tools, people to follow and related lists. The README text provided cuts off partway through the Research section, so those later sections are not described here. Many early entries come from the maintainer's own company blog, so read them with that in mind.

Read more on interesting repos · GitHub

This is a self-paced course, not a software library. It runs 12 weeks across 24 lessons. Each lesson has written material, and many have a notebook and a lab to practise on.

It starts with the older symbolic approach, where knowledge is written down as rules and ontologies. It then moves to neural networks and deep learning, with code in TensorFlow and PyTorch. Later lessons cover architectures for images and text, plus less common topics such as genetic algorithms and multi-agent systems.

The README is clear about what it leaves out. Classic machine learning is covered in Microsoft's separate Machine Learning for Beginners course. Business use of AI, Azure cloud services, chatbots and the deep maths are pointed to other resources. It also says its coverage of recent models may lag the state of the art.

Read more on interesting repos · GitHub

Every chatbot has a hidden set of instructions, called a system prompt, that is sent before your first message. It sets tone, rules and how the model uses tools. Vendors rarely publish them. This repo collects the full text as files you can read in your browser.

Files are sorted by company: Anthropic, OpenAI, xAI and others. The Anthropic section covers Claude on web and mobile, many Claude Code versions, Claude Design, Cowork, and the Office and Chrome integrations. The OpenAI section covers ChatGPT models and Codex. A dated table at the top lists the newest additions.

It is a reference library, not software you run. The prompts are extracted from live products rather than published by the vendors, so treat each file as the author's capture at that time. The author invites pull requests from anyone who got a different result.

Read more on interesting repos · GitHub

Many AI tutorials teach in scattered pieces: a paper, a fine-tuning post, an agent demo. This repo lays out one path through all of it. It has 20 phases that start with setup and math and end with agents, production infrastructure and capstone projects.

Each lesson sits in its own folder with runnable code, a written narrative (docs/en.md) and an outputs folder. The README describes a Build It / Use It split. You first write the algorithm from raw math, such as backprop, a tokenizer, attention or an agent loop. Then you run the same thing through a production library.

It also ships as agent skills. With npx and a skill-capable coding agent such as Claude Code or Codex, you can install a tutor that runs a placement quiz, saves a study plan to LEARNING.md and teaches one lesson per session. There are focused routes for MCP and Agent Skills, and the lessons are also on a website.

Read more on interesting repos · GitHub

This is a guide, not a library. Its author, Dex of HumanLayer, tried many agent frameworks and talked to over 100 founders building AI features. They found that many products sold as agents are mostly ordinary code with LLM steps placed where they help.

The guide starts from the common agent loop. The LLM picks the next step as structured JSON, normal code runs it, the result is added to the context window, and the loop repeats until the LLM says it is done. The author says this pattern alone tends to stall at around 70-80% quality, and that getting further means reverse-engineering the framework.

The answer is twelve principles in the spirit of the 12 Factor App. They cover owning your prompts, context window and control flow, treating tools as structured output, pausing and resuming, contacting humans, and keeping agents small and focused. Each factor has its own page. Examples are mostly TypeScript, and the README says the ideas apply in Python or any other language.

Read more on interesting repos · GitHub

Easy-Vibe is a course, not a software library. It starts from the idea that programming now begins with describing what you want, and it teaches you how to turn that description into a real product.

The course has three stages. Stage 1 covers finding a problem, talking to users, building a prototype and adding AI. Stage 2 covers full-stack work: frontend, databases, backend APIs, deployment and payments. Stage 3 covers advanced Claude Code topics such as MCP, Skills and Agent Teams, plus cross-platform apps.

Lessons are web pages with animations and interactive pieces, such as a simulated IDE and a click-through RAG data flow. An appendix works as a reference library of 80+ topics. The content is under CC BY-NC-SA 4.0, so it is free to read but not for commercial reuse.

Read more on interesting repos · GitHub

Claude Skills are small packages of instructions and scripts, usually built around a SKILL.md file, that teach Claude how to do a specific job. Good ones are scattered across many GitHub repos. This repo gathers them in one README.

The list is split into sections such as Document Skills, Development & Code Tools, Data & Analysis, Scientific & Research Tools, Media & Content and Security & Web Testing. There are also sections for articles and for larger collections. Each entry is a link with a one-line description. Many point to Anthropic's official skills repo (docx, pdf, pptx, xlsx). Others are community projects, such as read-only database query skills, test-design tools and linters for agent configs.

It is a link list, not a framework or installer, and it has no code of its own. You judge each skill by its own repo. The README also shows sponsor banners for two API platforms, which are ads. The README text I was given cuts off partway through the Data & Analysis section, so I can't describe the later sections in detail.

Read more on interesting repos · GitHub

This repo is a course-style set of hands-on tutorials for large language models. It grew out of lecture notes from two Shanghai Jiao Tong University courses, one on frontier NLP and one on AI security. The authors say it is free and meant to help students start quickly with course projects or research.

Each chapter has three parts: a PDF slide deck, a written tutorial in Markdown, and a Jupyter notebook with the code. Topics run from fine-tuning and deploying a model, and prompting with chain of thought, to newer ones like math reasoning, multimodal models, GUI agents and RLHF.

Many chapters focus on safety and security: jailbreak attacks, text watermarking, steganography, agent safety and safety alignment. Beginner LLM courses rarely cover these. The README is in Chinese. It also links to a separate Huawei Ascend course on the full LLM development process. It says the project is still being built and that its tips are not guaranteed to be correct.

Read more on interesting repos · GitHub

Web pages are built for people, with navigation, ads and JavaScript around the content. Turning that back into clean text is messy, and context windows are still too small for most whole sites. Agents do better with a short, curated entry point.

The proposal asks sites to publish an llms.txt markdown file. It has an H1 title, an optional summary quote, and H2 sections of links with short notes. An agent reads the small file, then fetches only the linked pages it needs. It also suggests serving a markdown copy of each page at the same URL with .md added.

It is not a sitemap, which lists every page and often lacks LLM-readable versions. It is also not robots.txt, which sets crawl permissions. The README explains why it uses a path-based file instead of /.well-known/: authors on shared hosts often control only a subpath. The repo is the spec and its discussion, not a tool.

Read more on interesting repos · GitHub