A curated collection of must-use, actively maintained resources for building and shipping AI systems.
Focus: AI engineering (RAG, agents, evals, guardrails, deploy) plus the best books, guides, papers, and a carefully selected set of tools.
Deep, durable knowledge — still valuable five years from now.
Modern & Practical
- Designing Machine Learning Systems — Scalable, maintainable ML pipelines (Chip Huyen).
- AI Engineering — End-to-end AI product building (Chip Huyen).
- Build a Large Language Model from Scratch — Transformers in raw PyTorch, layer by layer (Sebastian Raschka).
- Hands-On Large Language Models — Visual + practical guide to LLM applications (Jay Alammar, Maarten Grootendorst).
- LLM Engineer's Handbook — Production LLMOps: fine-tuning, quantization, serving (Labonne, Iusztin).
- The 100-Page Language Models Book — Concise, math-grounded path from n-grams to transformers (Andriy Burkov).
- Generative Deep Learning (2nd Edition) — GANs, VAEs, diffusion models (David Foster).
Foundational
- Artificial Intelligence: A Modern Approach — The canonical AI theory text (Russell, Norvig).
- Deep Learning — Mathematical foundations of neural networks (Goodfellow, Bengio, Courville).
- Deep Learning: Foundations and Concepts — Bishop's 2024 update; probability-grounded modern DL (Bishop & Bishop).
- Understanding Deep Learning — Math + intuition + Python notebooks (Simon Prince).
- Speech and Language Processing (3rd Edition) — The NLP reference, kept current through the deep learning era (Jurafsky, Martin).
- Reinforcement Learning: An Introduction (2nd Edition) — RL foundations (Sutton, Barto).
Beginner
Intermediate / Advanced
Focused
- DeepLearning.AI Short Courses
- Google DeepMind — Introduction to Reinforcement Learning
- Karpathy — Neural Networks: Zero to Hero
Research that shaped modern AI — worth reading to understand the "why" behind today's architectures.
- Attention Is All You Need — Transformer architecture.
- Scaling Laws for Neural Language Models — Model/data/compute scaling.
- Language Models are Few-Shot Learners — GPT-3 capabilities.
- Constitutional AI — Safer model alignment.
The toolchain for building with AI. Personal note: you don't need tons of frameworks — start with simple LLM calls and work up.
- Building Effective Agents (Anthropic) — ⭐ Patterns, pitfalls, and tradeoffs for designing AI agents.
- OpenAI Agents Guide — Practical guide on building agents.
- Google AI Agents Paper — Practical guide to building AI agents from Google.
- Google Agents Companion Paper — Companion guide from Google.
- OpenAI Cookbook — Example code, recipes, and best practices for working with OpenAI APIs.
- LLM Engineer Handbook — A goldmine of useful links for AI engineers.
- PocketFlow — Extremely minimalist AI agent framework in just 100 lines of code. Fantastic way to learn.
- Google ADK — Google's Agent Development Kit (Python, Java). Great local development experience + A2A + MCP.
- Pydantic-AI — Typed, structured LLM orchestration framework built on Pydantic models for safe, predictable outputs.
- LangGraph — Build multi-agent workflows with stateful graphs on top of LangChain.
- CrewAI — Agent orchestration with structured tasks and human-in-the-loop controls.
- AutoGen — Microsoft's framework for multi-agent conversation and collaboration.
- LlamaIndex — Data framework for ingesting, indexing, and querying private data with LLMs.
- Haystack — Open-source search/RAG framework with modular pipelines.
- Docling — Great library for ingesting any kind of document for RAG ⭐
- OpenAI Evals — OpenAI's framework for writing evals.
- Cursor — LLM-powered IDE for multi-file edits and codebase-aware chat.
- GitHub Copilot — In-IDE code completion, chat, and refactors.
Harnesses that turn LLMs into autonomous workers. The model is swappable; the harness is the product.
For live capability comparison, see Terminal-Bench and SWE-bench.
- Claude Code — Anthropic's CLI agent; multi-file codebase refactoring with long context.
- Codex CLI — OpenAI's Rust-based local terminal agent; lightweight and fast.
- Gemini CLI — Google's official open-source terminal agent; long-context repo exploration.
- Cursor CLI — Cursor's terminal-native agent with sandboxed permissions.
- Aider — Git-integrated pair programming with surgical edits and undo.
- OpenCode — Provider-agnostic terminal harness with a strong TUI.
- OpenHands — Open-source autonomous SWE platform; browser + shell + editor loop.
- Cline — Open-source agentic IDE extension with strong multi-provider support.
- Continue — Open-source IDE + CLI assistant with source-controlled rules.
- Goose — Block's extensible, MCP-driven local agent.
- Factory Droid — Benchmark-leading multi-model harness with BYOK local execution.
- Amp — Sourcegraph's commercial agentic coding tool with strong product UX.
- Mistral Vibe — Mistral's agentic coding CLI, powered by Devstral.
- Qwen Code — Alibaba's terminal coding agent, optimized for Qwen models.
- Pi — Highly customizable terminal harness; minimal base prompt, extension-driven.
- Nanocoder — Private, local-first agent for Ollama and LM Studio.
- Kilo CLI — Multi-mode agent with a unified gateway to 500+ models.
State-of-the-art models by modality.
The major frontier labs.
- ChatGPT — Best for general reasoning, tool use, and the broadest ecosystem.
- Claude — Best for long-context analysis, coding, and structured thinking.
- Gemini — Best for multimodal tasks and Google ecosystem integration.
- Grok — Best for real-time information via X and very long context.
- Llama — Best open-weight family for self-hosting and fine-tuning.
- Mistral — Best for lightweight, high-performance open-weight models.
- DeepSeek — Best for cost-efficient reasoning with open weights.
- Qwen — Best for multilingual and Chinese-first applications.
- Kimi — Best for long-context instruction following.
- GLM — Frontier-tier Chinese model with open weights.
- Cohere — Best for enterprise LLMs with strong retrieval-augmented generation APIs.
- GPT Image — OpenAI's integrated image generation with near-perfect text rendering.
- Midjourney — Artistic and photorealistic images.
- Adobe Firefly — Integrated into Creative Cloud; commercial-safe.
- Ideogram — Precise, legible text in generated images.
- Flux — High-res, prompt-editable, open-weight images.
- Google Veo — High-quality video with synchronized audio.
- Runway — Video editing + generation with granular creative control.
- Kling — Cinematic, realistic video generation.
- ElevenLabs — High-quality text-to-speech and voice cloning.
- Suno — AI music from text prompts.
Live benchmarks, pricing, and the latest model versions.
- OpenRouter — Unified API + live pricing across ~300 models.
- LMArena — Human-preference Elo rankings for text, image, and video.
- Artificial Analysis — Speed, price, and quality benchmarks across providers.
Stay current without drowning in noise.