A curated library of the best AI learning resources—courses, tutorials, docs, and talks on agentic AI, MCP, Agent Skills, RAG, machine learning, fine-tuning, models, and frameworks, for AI engineers and technical teams.
The standard PyTorch toolkit for image, video and audio generation — inference, LoRA adapters, quantization and ten maintained training scripts.
Ten worked kernels — from vector add to fused attention — that teach GPU programming at block level in Python instead of CUDA.
The reference for ZeRO, offloading and 3D parallelism — how to train or fine-tune a model that does not fit on your GPUs.
One storage layer for vector search, full-text search, versioning and training-data loading — instead of a vector DB bolted beside your data lake.
The open spec for letting an agent spend money for a user — and cryptographically prove it was authorized.
One decorator to trace an LLM app, and the same scorers for offline evals, live monitors and guardrails.
The resources learners opened most this week.
The full training account of a 229.9B-parameter MoE with 9.8B active — built end-to-end for agentic work.
Build your own coding agent — sandboxed execution, custom tools and multi-LLM routing, all documented.
Trace it, score it, gate it in CI — the whole LLM evaluation loop, self-hostable under Apache-2.0.
The post-training stack behind Nemotron — one codebase from a single GPU to thousand-GPU GRPO runs.
Stand up a self-hosted, citation-backed document QA system over messy enterprise PDFs in an afternoon.
The missing protocol layer between your agent backend and the screen a human is actually looking at.
Score your terminal agent against the benchmark everyone quotes, and learn how the harness decides pass or fail.
The GPU-accelerated pipeline NVIDIA uses to clean, dedupe and synthesize training data at trillion-token scale.
Turn prompt tweaking into a test suite, then attack your own agent before someone else does.
Derive flow matching and diffusion from the SDEs up, then build a latent diffusion model from scratch.
One API for distributed data, training, tuning and serving — the compute layer under most in-house AI platforms.
The official recipe book for quantizing an LLM into a format vLLM serves — FP8, INT4, NVFP4 and MXFP4.
Song Han's MIT course on making models small and fast — pruning, quantization and LLM deployment, with five labs.
Run RAG retrieval inside the Postgres you already operate — HNSW indexes, metadata filters and hybrid search, in SQL.
The gap between 'my notebook works' and 'this serves traffic' — GPU deployment as decorated Python, no Dockerfile or Kubernetes.
Microsoft's own AI red team framework — run repeatable multi-turn jailbreak campaigns against your endpoint and score the results.
Make an LLM drive a real browser without selectors that break — and keep deterministic control where the page is stable.
Carnegie Mellon's graduate LLM curriculum with slides and lecture code in the open — build LLaMa from scratch, then everything that comes after it.