← The lab

Learn to train the thing yourself.

Everything below was written or taught by someone else — universities, research labs, and the people who wrote the libraries you will import. None of it costs money, and none of it is summarised here: each row is a link to the original, with one line saying what it is and how long it takes. Sorted so you can start anywhere and know what you are skipping.

Links: 254
Sections: 20
Behind a paywall: 0
Last checked: 2026-09-12

Start here

Three honest routes in. Pick the one that describes you now, not the one that describes who you want to be.

You have never trained anything

Start at the bottom and do not skip the maths. Three to six months, honestly.

Foundations Python & tooling Classical ML Deep learning

You can code and want to fine-tune this week

Skip the theory for now. Run one fine-tune end to end, then come back for the parts that broke.

Fine-tuning Data & datasets Evaluation Free compute

You want to understand a model, not just use one

Write one. Everything after that reads differently.

Build an LLM from scratch NLP & transformers Pretraining & scale Interpretability

254 shown

Foundations

The linear algebra, calculus and probability every later section quietly assumes. Skip it if you already read a Jacobian without flinching.

VideoEssence of Linear Algebra3Blue1Brown15 short films · the geometric picture behind matrices VideoEssence of Calculus3Blue1Brown12 films · derivatives and the chain rule, visually VideoNeural Networks3Blue1Brownbackprop and, since 2024, attention and transformers BookMathematics for Machine LearningDeisenroth, Faisal & Ongfree PDF · covers exactly what ML needs, no more CourseMIT 18.06 Linear AlgebraGilbert Strang · MIT OCW35 lectures · CC BY-NC-SA · the classic CourseMIT 18.065 Matrix Methods in Data AnalysisGilbert Strang · MIT OCWStrang again, aimed squarely at learning problems InteractiveSeeing TheoryBrown Universityprobability and statistics you can drag around BookIntroduction to Probability for Data ScienceStanley Chan · Purduefree PDF released by the author ArticleThe Matrix Calculus You Need For Deep LearningParr & Howardone long page that replaces a semester of confusion CourseComputational Linear Algebrafast.ainotebooks plus video lectures, the numerical side CourseStatistics & ProbabilityKhan Academyif the ground floor is missing BookImmersive Linear AlgebraStrom, Astrom & Akenine-Mollerthe first linear algebra book with fully interactive figures

Python & the tooling

You will spend more hours in NumPy, pandas and a terminal than in any model architecture.

CourseCS50's Introduction to Programming with PythonHarvardfree · CC BY-NC-SA · from zero BookPython Data Science HandbookJake VanderPlasfull text online · NumPy, pandas, matplotlib, scikit-learn DocsNumPy: the absolute basics for beginnersNumPyofficial, short, correct Docs10 minutes to pandaspandasthe fastest honest introduction BookScientific Visualization: Python + MatplotlibNicolas Rougierfree PDF · how to make a figure that is not embarrassing CourseThe Missing Semester of Your CS EducationMITshell, git, tmux, debugging · CC BY-NC-SA BookPro GitChacon & Straubfree · CC BY-NC-SA · the entire book online InteractiveLearn Git BranchingPeter Cottlegit as a visual puzzle game DocsThe Python TutorialPython Software Foundationthe official tour of the language, still the most accurate one Docsuv — Python packaging that is not painfulAstralthe modern, fast way to manage environments DocsJupyter Notebook documentationProject Jupyterthe tool every course below hands you

Classical machine learning

Gradient boosting still wins most tabular problems. Learn this before reaching for a transformer.

CourseMachine Learning Crash CourseGooglefree · ~15 h · interactive, rebuilt in 2024 CourseMachine Learning SpecializationAndrew Ng · DeepLearning.AIfree to audit · the canonical entry point CourseStanford CS229: Machine LearningStanfordlecture notes and problem sets, openly posted VideoCS229 lecture videosStanford Onlinethe full lecture series, Autumn 2018 BookAn Introduction to Statistical LearningJames, Witten, Hastie & Tibshiranifree PDF · Python and R editions · plus a free video course BookThe Elements of Statistical LearningHastie, Tibshirani & Friedmanfree PDF · the harder sequel CourseKaggle LearnKaggle~20 micro-courses, a few hours each, all in-browser Docsscikit-learn User Guidescikit-learna textbook disguised as documentation CourseMachine Learning for BeginnersMicrosoft12 weeks, 26 lessons · MIT licence CourseData Science for BeginnersMicrosoft10 weeks, 20 lessons · MIT licence Coursemlcourse.aiYury Kashnitsky · OpenDataSciencefree · notebooks, assignments and Kaggle contests DocsXGBoost tutorialsXGBoostthe algorithm that quietly wins competitions BookProbabilistic Machine LearningKevin Murphyfree draft PDFs of both volumes DocsGoogle Machine Learning GuidesGooglethe rules-of-ML and tuning playbooks used internally

Deep learning

Backpropagation, optimisers, regularisation, convolutions — the layer of knowledge that does not go stale.

CoursePractical Deep Learning for Codersfast.ai · Jeremy Howardfree · top-down, code first · the best-loved course in the field CourseMIT 6.S191: Introduction to Deep LearningMITfree · one week of lectures and labs, re-recorded every year BookDive into Deep LearningZhang, Lipton, Li & Smolafree · maths, code and discussion on the same page BookDeep LearningGoodfellow, Bengio & Courvillefree HTML · the reference text BookUnderstanding Deep LearningSimon J.D. Princefree PDF, slides and notebooks · modern and beautifully drawn BookNeural Networks and Deep LearningMichael Nielsenfree · derives backprop by hand, gently CourseNYU Deep Learning (DS-GA 1008)Yann LeCun & Alfredo Canzianifree · lectures, notes and notebooks CourseStanford CS231n: Deep Learning for Computer VisionStanfordnotes and assignments open · the CNN classic DocsPyTorch TutorialsPyTorchofficial · start with 'Learn the Basics' CourseUvA Deep Learning TutorialsUniversity of Amsterdamnotebooks in both PyTorch and JAX CourseDeep Learning SpecializationAndrew Ng · DeepLearning.AIfree to audit · five courses CodemicrogradAndrej KarpathyMIT · ~150 lines · autograd small enough to read in one sitting BookDeep Learning Tuning PlaybookGoogle Research & Harvardhow to actually choose hyperparameters, from people who do it daily CodeAnnotated Paper Implementationslabml.aiMIT · 60+ architectures implemented in PyTorch with the paper alongside ArticleA Recipe for Training Neural NetworksAndrej Karpathythe debugging discipline nobody teaches you

NLP & transformers

Tokenisation, embeddings, attention — where language models actually begin.

CourseStanford CS224n: NLP with Deep LearningChristopher Manning · Stanfordslides, notes and assignments open every year VideoCS224n lecture videosStanford Onlinethe full lecture series on YouTube CourseHugging Face NLP CourseHugging Facefree · Apache-2.0 · transformers, datasets and tokenizers, hands-on ArticleThe Illustrated TransformerJay Alammarthe diagram everyone has seen, and for good reason CodeThe Annotated TransformerHarvard NLPthe 2017 paper, line by line, as runnable code BookSpeech and Language Processing (3rd ed. draft)Jurafsky & Martinfree draft chapters · the field's standard textbook CourseCMU CS11-711: Advanced NLPGraham Neubig · CMUslides, videos and assignments, all public ArticleThe Illustrated Word2vecJay Alammarembeddings from first principles ArticleTransformers from ScratchBrandon Rohrera very patient derivation, matrices included ArticleThe Transformer Family v2.0Lilian Wengevery architectural variant, catalogued ArticleAttention? Attention!Lilian Wengwhere attention came from PaperAttention Is All You NeedVaswani et al.the 2017 paper itself · read it after the illustrated version CourseAdvanced NLP with spaCyInes Montanifree · interactive · the practical pipeline side CodeNLP with Transformers — notebooksTunstall, von Werra & Wolfthe book's code, Apache-2.0, free to run

Build an LLM from scratch

The fastest way to stop treating a model as a black box is to write one. Everything here ends in code you typed yourself.

CourseNeural Networks: Zero to HeroAndrej Karpathyfree · ~10 video lectures · micrograd, then makemore, then GPT VideoLet's build GPT: from scratch, in code, spelled outAndrej Karpathy2 h · a working GPT in one sitting VideoLet's build the GPT TokenizerAndrej Karpathy2 h 13 · BPE, and why tokenisation causes half of all LLM weirdness VideoDeep Dive into LLMs like ChatGPTAndrej Karpathy3 h 31 · pretraining, SFT and RLHF end to end, little maths required CodenanoGPTAndrej KarpathyMIT · ~300 lines that reproduce GPT-2 CodeminGPTAndrej KarpathyMIT · the readable, minimal ancestor of nanoGPT Codellm.cAndrej KarpathyMIT · GPT-2 training in plain C and CUDA, no framework CodenanochatAndrej Karpathythe full ChatGPT pipeline — pretrain, SFT, RL, web UI — in one clean repo CourseStanford CS336: Language Modeling from ScratchStanfordthe whole pipeline as coursework: tokenizer, model, training, evaluation VideoCS336 lecture videosStanford Onlineall lectures, free CodeBuild a Large Language Model (From Scratch) — codeSebastian RaschkaApache-2.0 · every chapter's notebooks are free; the book is optional CourseLLM CourseMaxime LabonneApache-2.0 · roadmaps plus dozens of free Colab notebooks CourseHugging Face LLM CourseHugging Facethe NLP course's successor: pretraining, fine-tuning, inference Codemodded-nanogptKeller Jordan et al.the community speedrun — read the diffs to see what actually helps CodelitgptLightning AIApache-2.0 · 20+ models as readable single-file implementations ArticleThe Annotated GPT-2Aman AroraGPT-2 explained alongside its own code

Pretraining & scale

What changes when the run stops fitting on one GPU: parallelism, schedules, scaling laws, and the failure modes nobody warns you about.

BookThe Ultra-Scale Playbook: Training LLMs on GPU ClustersHugging Facefree · 5D parallelism, ZeRO, kernels · drawn from 4000+ experiments BookThe Smol Training PlaybookHugging Facehow a small model actually gets trained, decision by decision CodeEleutherAI CookbookEleutherAIMIT · practical recipes and calculators from people who train at scale ArticleTransformer Math 101EleutherAIthe arithmetic of memory, FLOPs and cost, before you rent a GPU BookHow to Scale Your ModelGoogle DeepMindfree · a systems view of TPUs, parallelism and rooflines PaperScaling Laws for Neural Language ModelsKaplan et al.the original scaling-law paper PaperTraining Compute-Optimal LLMs (Chinchilla)Hoffmann et al.the correction that reset the field's data budgets CodeOLMo — a fully open language modelAllen Institute for AIweights, data, code and training logs, all released CodenanotronHugging FaceApache-2.0 · minimalist 3D-parallel pretraining CodeMegatron-LMNVIDIAthe reference implementation for large-scale training DocsDeepSpeed tutorialsMicrosoftZeRO, offload and pipeline parallelism, explained by its authors DocsGetting Started with FSDPPyTorchthe sharding you will actually reach for first DocsHugging Face AccelerateHugging Facethe same training loop on one GPU, eight GPUs or a TPU PaperOPT-175B logbookMeta AIthe unvarnished day-by-day diary of a large run going wrong

Fine-tuning

Taking a trained model and bending it to your task. The largest section here, because it is the part most people actually need and the part with the most folklore around it.

DocsFine-tune a pretrained modelHugging Face Transformersofficial · the honest starting point, Trainer and a plain PyTorch loop DocsPEFT documentationHugging FaceLoRA, prefix tuning, IA3, adapters · conceptual guides plus API DocsTRL — Transformer Reinforcement LearningHugging Facethe library behind most open SFT, DPO and GRPO recipes DocsSFTTrainer guideHugging Face TRLsupervised fine-tuning, including packing and completion-only loss DocsUnsloth documentationUnslothfree · fine-tuning that fits in a free Colab GPU, with the tricks explained NotebookUnsloth notebooksUnsloth100+ ready Colab notebooks — Llama, Qwen, Gemma, Mistral, vision, GRPO DocsAxolotl documentationAxolotl AIApache-2.0 · config-file fine-tuning; the YAML examples are the real course CodeLLaMA-Factoryhiyouga et al.Apache-2.0 · 100+ models, a web UI, and every method in one place DocstorchtunePyTorchnative PyTorch recipes you can read top to bottom PaperLoRA: Low-Rank Adaptation of Large Language ModelsHu et al.the paper that made fine-tuning affordable PaperQLoRA: Efficient Finetuning of Quantized LLMsDettmers et al.65B on a single 48 GB card · read alongside the bitsandbytes docs ArticleMaking LLMs even more accessible with bitsandbytes and QLoRAHugging Facethe practical write-up of 4-bit fine-tuning ArticleParameter-Efficient Fine-Tuning using PEFTHugging Facethe short version, with working code CodeAlignment HandbookHugging FaceApache-2.0 · full, reproducible recipes: SFT then DPO, configs included Coursesmol-courseHugging Facea hands-on course on aligning small models on modest hardware CourseFinetuning Large Language ModelsDeepLearning.AI & Laminifree short course · ~1 h · when to fine-tune at all ArticlePractical Tips for Finetuning LLMs Using LoRASebastian Raschkahundreds of ablations distilled into rules of thumb ArticleFinetuning LLMs Efficiently with AdaptersSebastian Raschkawhat each PEFT method actually changes in the weights DocsChat templatesHugging Face Transformersthe single most common cause of a fine-tune that trains but never answers ArticleFine-tune Llama 3 with ORPOMaxime LabonneSFT and preference alignment collapsed into one step NotebookLlama CookbookMetaofficial fine-tuning and deployment recipes NotebookGemma CookbookGooglethe same idea for Gemma, including Keras and JAX paths PaperFinetuned Language Models Are Zero-Shot Learners (FLAN)Wei et al.the paper that introduced instruction tuning PaperSelf-InstructWang et al.how to generate the instruction data you do not have PaperLIMA: Less Is More for AlignmentZhou et al.1000 carefully chosen examples beating far larger sets CodemergekitArcee AImerging fine-tuned checkpoints instead of retraining NotebookHugging Face Open-Source AI CookbookHugging Facedozens of task-shaped recipes, all runnable NotebookPyTorch Lightning / Fabric fine-tuning studiosLightning AIfree-tier GPU studios with published fine-tuning templates

Preference tuning & RLHF

Turning a model that completes text into one that answers. RLHF, DPO, GRPO, and the reasoning-model recipes that followed.

BookThe RLHF BookNathan Lambertfree online · the only full-length treatment of RLHF, written openly ArticleIllustrating Reinforcement Learning from Human FeedbackHugging Facethe diagrams most people learned RLHF from PaperTraining language models to follow instructions (InstructGPT)Ouyang et al.the paper behind the chat models you use PaperDirect Preference OptimizationRafailov et al.RLHF's result without the reinforcement learning DocsDPOTrainerHugging Face TRLDPO in practice, with the dataset format spelled out DocsGRPOTrainerHugging Face TRLgroup-relative policy optimisation, the R1-style recipe PaperDeepSeek-R1: incentivising reasoning via RLDeepSeek-AIreasoning emerging from reinforcement learning alone Codeopen-r1Hugging Facea fully open reproduction of R1 · read the issues as much as the code PaperConstitutional AI: Harmlessness from AI FeedbackAnthropicreplacing part of the human feedback with a written constitution ArticleReward Hacking in Reinforcement LearningLilian Wengevery way a reward model gets gamed BlogInterconnectsNathan Lambertthe running commentary on post-training, from inside the field CourseProximal Policy Optimization explainedHugging Face Deep RL Coursethe algorithm underneath classic RLHF ReferenceAnthropic's Claude constitutionAnthropicwhat an actual alignment specification looks like in writing PaperTulu 3: pushing frontiers in open post-trainingAllen Institute for AIa complete, documented, reproducible post-training pipeline

Reinforcement learning

The background the section above assumes. Worth a detour if policy gradients are still a rumour to you.

CourseSpinning Up in Deep RLOpenAIfree · MIT licence · theory, key papers and clean implementations CourseDeep RL CourseHugging Facefree · Apache-2.0 · hands-on, with a leaderboard to train against BookReinforcement Learning: An IntroductionSutton & Bartofree PDF · the field's textbook, from its authors VideoRL Lecture SeriesDavid Silver · DeepMind & UCLten lectures, still the clearest introduction on video CourseCS285: Deep Reinforcement LearningSergey Levine · UC Berkeleylectures, homework and video, all public DocsGymnasium documentationFarama Foundationthe environment API everything else is built on DocsStable-Baselines3DLR-RMreliable reference implementations to compare yours against CodeCleanRLCosta Huang et al.single-file implementations with published training curves

Evaluation

The part people skip, and the reason so many fine-tunes look great in a demo and fail in use.

BookEvaluation GuidebookHugging Facefree · when to use which evaluation and how each one lies to you Codelm-evaluation-harnessEleutherAIMIT · the de-facto standard harness · 60+ benchmarks DocsInspectUK AI Security InstituteMIT · a serious framework for writing your own evaluations ReferenceHELMStanford CRFMholistic evaluation, with all raw results published DocsEvaluateHugging Facemetrics, with their limitations documented next to them ReferenceLMArenaLMSYSblind pairwise human comparison at scale ReferenceOpen LLM Leaderboard archiveHugging Faceand, more usefully, the write-ups on why it kept being gamed DocsRagasExploding Gradientsevaluation aimed specifically at retrieval pipelines PaperJudging LLM-as-a-Judge (MT-Bench)Zheng et al.how far a model-as-grader can be trusted, measured ArticleYour AI Product Needs EvalsHamel Husainthe practitioner's argument for building evals before features

Data & datasets

Model quality is mostly data quality. This is the least glamorous section and the highest-leverage one.

DocsDatasets documentationHugging Faceloading, streaming and processing corpora that do not fit in RAM ArticleFineWeb: decanting the web for the finest text dataHugging Facean unusually complete account of how a 15T-token corpus was filtered CodedatatroveHugging FaceApache-2.0 · the pipeline library FineWeb was built with CodeDolmaAllen Institute for AI3T-token open corpus plus the toolkit that produced it ReferenceThe PileEleutherAIthe 825 GiB corpus, with a paper documenting every component DocsdistilabelArgillasynthetic data generation with the provenance kept intact DocsArgillaArgilla · Hugging Facehuman annotation and dataset curation, self-hostable CourseIntroduction to Data-Centric AIMITfree · lectures and labs on improving data rather than models ArticleCosmopedia: how to create large synthetic datasetsHugging Face25B synthetic tokens, and the prompt design behind them ReferenceHugging Face Hub datasetsHugging Facewhere to look first before building anything yourself DocsData Cards / Dataset documentationHugging Facehow to document a corpus so that it stays usable

Quantization, inference & speed

Getting a model to fit, and then to be fast. GPU arithmetic, kernels, quantization formats and serving.

DocsQuantizationHugging Face Transformersevery supported format compared in one table Docsbitsandbytesbitsandbytes8-bit and 4-bit, the backend behind QLoRA Codellama.cppggml.orgMIT · GGUF, CPU inference, and the quant formats everyone ships DocsvLLM documentationvLLMApache-2.0 · paged attention and continuous batching, explained by the implementers ArticleMaking Deep Learning Go Brrrr From First PrinciplesHorace Hecompute-bound, memory-bound, overhead-bound — the mental model ArticleOptimizing LLMs for speed and memoryHugging Facequantization, Flash Attention and KV caching, with the arithmetic shown PaperFlashAttentionDao et al.the IO-aware attention kernel now in everything PaperGPTQFrantar et al.post-training quantization down to 3-4 bits PaperAWQ: Activation-aware Weight QuantizationLin et al.the other format you will meet on the Hub CourseGPU MODE lecturesGPU MODEfree · a whole community-run course on CUDA and kernel writing DocsTriton tutorialsOpenAIwriting GPU kernels in Python, starting from vector add DocsCUDA C++ Programming GuideNVIDIAthe primary source, when the tutorials run out DocsOptimumHugging FaceONNX, TensorRT and hardware-specific export paths CodeOllamaOllamaMIT · the shortest path to running your fine-tune locally DocsText Generation InferenceHugging Faceproduction serving, with the trade-offs documented

RAG, embeddings & retrieval

Putting the knowledge in the context instead of in the weights. Usually cheaper and more correctable than fine-tuning.

DocsSentence TransformersUKP Lab · Hugging FaceApache-2.0 · embeddings, and how to fine-tune your own NotebookAdvanced RAG on Hugging Face documentationHugging Face Cookbooka complete pipeline in one runnable notebook ReferenceMTEB leaderboardHugging Facehow to choose an embedding model on evidence DocsLangChain tutorialsLangChainthe most common orchestration layer, taught by example DocsLlamaIndex documentationLlamaIndexindexing and query engines, the other standard option DocsFAISS wikiMeta AIvector search from the library that defined it CoursePinecone LearnPineconefree · vendor-made but unusually honest on the fundamentals DocsChroma documentationChromathe lightest vector store to start with PaperRetrieval-Augmented Generation (original paper)Lewis et al.where the term came from, in 2020 ArticleContextual RetrievalAnthropica measured fix for the chunking problem, with numbers

Prompting & agents

The cheapest lever, and the one to exhaust before you fine-tune anything.

DocsPrompt engineering overviewAnthropicofficial · specific techniques, each with worked examples CoursePrompt Engineering Interactive TutorialAnthropicfree · nine chapters with exercises you actually run NotebookAnthropic CookbookAnthropicclassification, RAG, tool use, evaluation — as runnable notebooks ArticleBuilding effective agentsAnthropicthe essay that cut through most agent-framework hype NotebookOpenAI CookbookOpenAIMIT · a large, maintained recipe collection CoursePrompt Engineering GuideDAIR.AIfree · MIT licence · techniques with the papers they came from CourseAI Agents CourseHugging Facefree · from first principles to a deployed agent, with a certificate DocssmolagentsHugging FaceApache-2.0 · agents in about a thousand lines you can read DocsModel Context ProtocolMCPthe open standard for giving a model tools and data PaperChain-of-Thought PromptingWei et al.the prompt change that unlocked reasoning behaviour PaperReAct: Synergizing Reasoning and ActingYao et al.the loop nearly every agent framework re-implements ArticleEffective context engineering for AI agentsAnthropicwhat to put in the window, and what to leave out

Vision, audio & diffusion

Everything that is not text: image generation, vision-language models, speech.

CourseDiffusion Models CourseHugging Facefree · Apache-2.0 · theory and Diffusers, four units CourseComputer Vision CourseHugging Facefree · community-written, from convolutions to ViTs CourseAudio CourseHugging Facefree · ASR, TTS and audio classification, hands-on CourseDeep Learning Foundations to Stable Diffusionfast.aifree · 30 hours · rebuilds Stable Diffusion from the ground up ArticleWhat are Diffusion Models?Lilian Wengthe maths, derived carefully, in one post ArticleThe Illustrated Stable DiffusionJay Alammarthe components and how they fit together DocsDiffusers documentationHugging FaceApache-2.0 · including LoRA and DreamBooth fine-tuning PaperDenoising Diffusion Probabilistic ModelsHo, Jain & Abbeelthe 2020 paper that started the current wave PaperLearning Transferable Visual Models (CLIP)Radford et al.the joint image-text space nearly everything builds on PaperAn Image is Worth 16x16 Words (ViT)Dosovitskiy et al.transformers taking over vision CodeWhisperOpenAIMIT · weights and code for speech recognition, fine-tunable ArticleVision Language Models ExplainedHugging Facehow VLMs are built, trained and evaluated

Interpretability & safety

What is actually happening inside the weights, and what to do about the parts you would rather were not.

ReferenceTransformer Circuits ThreadAnthropicfree · the central body of mechanistic interpretability work PaperA Mathematical Framework for Transformer CircuitsElhage et al.the paper the rest of the field cites PaperToy Models of SuperpositionElhage et al.why single neurons rarely mean one thing CourseHow to become a mechanistic interpretability researcherNeel Nandaa concrete, opinionated study path with exercises DocsTransformerLensNeel Nanda et al.MIT · the library most interpretability work runs on CourseARENAARENAfree curriculum · transformers, interpretability and RL, as exercises CourseAI Safety FundamentalsBlueDot Impactfree · curricula and readings, alignment and governance tracks ReferenceDistillDistillarchived but unmatched · interactive explanations of what networks learn PaperConcrete Problems in AI SafetyAmodei et al.the 2016 paper that framed the practical questions DocsResponsible Scaling / model cardsHugging Facehow to document a model honestly, including what it fails at

Shipping it: MLOps & systems

The distance between a notebook that works and a service that keeps working.

CourseMade With MLGoku Mohandasfree · MIT licence · design, develop, deploy, iterate CourseFull Stack Deep LearningFSDLfree · lectures and labs on building ML-powered products CourseCS329S: Machine Learning Systems DesignChip Huyen · Stanfordslides, notes and readings, all public BookML Engineering Open BookStas Bekmanfree · hard-won operational detail on large training runs CourseMLOps ZoomcampDataTalks.Clubfree · nine weeks, run publicly every year CourseLLM ZoomcampDataTalks.Clubfree · building an LLM application end to end CourseW&B CoursesWeights & Biasesfree · experiment tracking, evaluation and LLM-app courses ArticleMLOps: continuous delivery and automation pipelinesGoogle Cloudthe maturity-model write-up everyone quotes DocsGradio documentationHugging FaceApache-2.0 · the fastest way to put a model in front of a person DocsHugging Face Hub documentationHugging Facepublishing models, datasets and Spaces properly

Staying current

The field moves fast enough that a course from last year has gaps. These are the feeds worth keeping.

ReferenceDaily PapersHugging Facea curated, discussed selection rather than the full arXiv firehose ReferencearXiv cs.CL — recentarXivthe firehose itself, when you want it BlogLil'LogLilian Wenglong, careful surveys · several are the best sources on their topic BlogAhead of AISebastian Raschkaresearch summaries with working code, free archive ReferenceConnected PapersConnected Papersfree tier · the citation graph around a paper, as a map ReferencealphaXivalphaXivarXiv with a comment layer, useful for hard papers BlogThe GradientThe Gradientlonger-form essays, editorially reviewed BlogImport AIJack Clarkweekly · research plus policy, from someone who reads everything ReferenceAI Index ReportStanford HAIfree · the annual numbers, with sources ArticleHow to Read a PaperS. Keshavthree pages · the three-pass method, and it works

Free compute & practice

Nothing on this page sticks until you run it. All of these give you a GPU without a credit card.

ToolGoogle ColabGooglefree tier · a T4 GPU, enough for LoRA fine-tuning of small models ToolKaggle NotebooksKagglefree · ~30 GPU hours a week, plus TPU v3-8 and v5e quotas ToolKaggle CompetitionsKagglethe only reliable way to find out whether you actually know this ToolHugging Face SpacesHugging Facefree CPU hosting for demos, and the source of thousands of examples CourseHugging Face LearnHugging Facethe index of every free course they publish, in one place ToolLightning StudiosLightning AIfree monthly GPU credits, persistent environments ReferenceKaggle DatasetsKagglea few hundred thousand public datasets to practise on ReferenceOpen LLM resources on the HubHugging Facestart from a small open base model; the 0.5B ones train in minutes

Nothing on this page is hosted here and nothing has been copied from it: every row links to its source, and the one-line notes are ours. Licences belong to the original authors — several items are Apache-2.0, MIT or CC BY-NC-SA, and that is stated where it matters.

Found something dead, or something missing that should be here? Tell us.
Sane Labs · 2026 · the lab · Synth-2
↑ Back to top