Every doc carries a controlled tags: list (layered on top of the original free-form topics: field, which is left untouched). Five namespaces:
| Tag | Docs | Covers |
|---|---|---|
topic/neural-network-foundations |
84 | MLPs, the forward pass, backpropagation, and the training machinery — optimizers (SGD/Adam), normalization, dropout, initialization. The entry point. |
topic/classical-ml |
35 | Non-neural methods: regression, SVMs/kernels, trees & ensembles, naive Bayes, EM, PCA, and learning theory. |
topic/computer-vision |
89 | CNNs (ResNet, VGG, DenseNet), image classification, object detection & segmentation, and vision transformers (ViT, DETR, MAE). |
topic/sequence-models-rnn |
76 | Word embeddings, RNN/LSTM/GRU, seq2seq with attention, neural machine translation, and classic NLP tasks. The bridge to transformers. |
topic/transformers-attention |
140 | The transformer architecture and attention itself: self-/multi-head attention, positional encodings, and architecture explainers. |
topic/language-models |
377 | Building LLMs end-to-end: pretraining, the BERT/GPT/T5/LLaMA model reports, scaling laws, emergent abilities, and mixture-of-experts. |
topic/efficient-architectures |
205 | Alternatives to dense quadratic attention: sparse/linear attention, FlashAttention, long-context methods, and state-space models (Mamba, RWKV). |
topic/generative-models |
144 | Deep generative modeling: VAEs, GANs, normalizing flows, autoregressive models, and diffusion incl. latent/stable diffusion and text-to-image. |
topic/multimodal |
69 | Models bridging modalities: contrastive vision-language (CLIP), text-to-image, audio/speech transformers, and general multimodal work. |
topic/reinforcement-learning |
45 | RL as a subfield: MDPs, value/policy iteration, Q-learning, policy gradients, deep RL, imitation, and game-playing agents. |
topic/alignment-rlhf |
93 | Post-training LLMs to human preferences: instruction tuning/SFT, RLHF/InstructGPT, reward modeling, DPO/ORPO/GRPO, and safety alignment. |
topic/reasoning-agents |
166 | Eliciting and structuring reasoning, and tool-using agents: chain-of-thought, tree-of-thoughts, test-time compute, ReAct/Toolformer, and RAG. |
topic/efficiency-systems |
44 | Making models cheap to adapt, train, and serve: quantization, LoRA/PEFT, CUDA/Triton kernels, parallelism, and paged-attention serving (vLLM). |
topic/interpretability |
29 | Reverse-engineering trained models: mechanistic interpretability (circuits, superposition, induction heads), probing, and knowledge editing. |
topic/evaluation-trust |
82 | Measuring and stress-testing models: benchmarks, evaluation methodology, calibration/uncertainty, robustness, bias/fairness, and hallucination. |
topic/ml-engineering |
109 | Hands-on build & ship: production agent engineering, app/agent frameworks, dev tooling (FastHTML, nbdev, CUDA-for-Python), and prompt engineering. |
topic/ai-industry-news |
71 | News, commentary, and the broader ecosystem: model-release roundups, policy, AGI debate, conference coverage, and applied 'AI for X' talks. |
level/intro · level/intermediate · level/advanced · level/frontier
medium/paper · medium/lecture · medium/article
task/vision · task/language · task/speech-audio · task/multimodal · task/graph · task/rl-control · task/tabular-classical · task/general
technique/mlp · technique/cnn · technique/rnn-lstm · technique/transformer · technique/attention · technique/diffusion · technique/gan · technique/vae · technique/normalizing-flow · technique/ssm · technique/moe · technique/lora-peft · technique/quantization · technique/rlhf · technique/dpo · technique/ppo · technique/cot · technique/rag · technique/flashattention · technique/embeddings
Tags are auto-assigned (keyword rules + a content-reading pass); they're great for narrowing, but full-text search is the ground truth when a filter looks sparse.