Skip to content

Cheat sheets

High-density, printable 2-column reference sheets that condense foundational math, system architectures, runnable PyTorch implementations, napkin scaling, debugging playbooks, and interview cards for revision.

File Specs Description
llm-genai-cheatsheet.pdf 24 pages, 2 columns LLM & Generative AI Engineer: Math, GPU hardware, pretraining, post-training (RLHF/GRPO), inference serving, RAG, agents, from-scratch code, debugging, formulas, and papers.
vision-cheatsheet.pdf 25 pages, 2 columns Vision & Multimodal AI Engineer: Classical geometry, CNNs, ViTs, Detection (YOLOv1--v11, DETR), SAM 1/2, SSL (MAE/DINOv2), CLIP/SigLIP, VLMs (LLaVA/Qwen2-VL), Diffusion (DDPM/EDM), DiT, Flow Matching (Flux), Video, 3DGS, Edge NPUs, and training loops.
life-of-a-token.pdf 5 pages, 2 columns Narrative recap: The life of a token told as one continuous story across 9 stages.

Build Instructions

Each document is compiled using TeX Live / TinyTeX (pdflatex) with mathpazo, tcolorbox, listings, pgfplots, microtype, and booktabs. Run pdflatex twice to populate the Table of Contents and cross-references.

1. Build Vision Cheatsheet (25 Pages)

cd cheatsheet/vision/src
pdflatex main.tex
pdflatex main.tex

2. Build LLM & GenAI Cheatsheet (24 Pages)

cd cheatsheet/src
pdflatex main.tex
pdflatex main.tex

3. Build Story PDF

cd cheatsheet/story
pdflatex life-of-a-token.tex
pdflatex life-of-a-token.tex

Vision Cheatsheet Section Directory (vision/src/sec/)

  • 00-front.tex --- Title, study method, 12 numbers everyone asks, Table of Contents
  • 01-geometry-filtering.tex --- Pinhole camera, epipolar geometry, Fundamental/Essential matrix, Canny, Harris, SIFT, Optical Flow
  • 01b-classical-descriptors.tex --- Keypoint detectors (FAST/ORB), Lowe's ratio test, RANSAC iteration math, stereo depth \(Z=fB/d\), Zhang calibration and PnP, RAFT optical flow
  • 02-cnn-foundations.tex --- Convolution arithmetic, backprop derivation, BatchNorm vs LayerNorm, ResNet skip connection, Kaiming vs Xavier init
  • 02b-modern-convnets.tex --- Depthwise separable convolutions, MobileNetV2 inverted bottlenecks, ShuffleNet MAC, ConvNeXt-V2 GRN
  • 03-vision-transformers.tex --- ViT architecture, configurations (Base to ViT-22B), DeiT distillation, Attention Distance, NaViT patch packing
  • 04-hierarchical-vit.tex --- Swin Transformer, W-MSA/SW-MSA cyclic shift masking, PVT Spatial Reduction, CoAtNet, FastViT
  • 05-object-detection.tex --- Two-stage detection (Faster R-CNN, FPN), RoIPool vs RoIAlign, CIoU loss, RPN anchors, FPN level assignment, COCO mAP computation
  • 05b-single-stage-yolo.tex --- SSD, RetinaNet, YOLOv1--v11 architectural lineage, focal loss derivation and gradient with plot, Task-Aligned Assigner (TAL), DFL, YOLOv10 NMS-free
  • 06-detr-query-detection.tex --- DETR Hungarian bipartite set loss, Deformable Attention, DINO-DETR contrastive denoising, RT-DETR
  • 07-segmentation-sam.tex --- Semantic (U-Net, DeepLabv3+ ASPP), Instance (Mask R-CNN mask head), Panoptic (Mask2Former), SAM foundation segmentation
  • 07b-sam-foundation-seg.tex --- SAM ViT-H windowed backbone, prompt encoder, ambiguity handling, SAM 2 streaming video memory bank
  • 08-self-supervised-vision.tex --- Contrastive learning (SimCLR InfoNCE MI lower bound, MoCo), non-contrastive (BYOL non-collapse), SSL eval protocols
  • 08b-ssl-mae-dino.tex --- MAE asymmetric 75% patch masking, DINO centering and sharpening, DINOv2 KoLeo regularizer, Register tokens
  • 09-vlm-clip-siglip.tex --- CLIP dual-encoder loss and gradients, distributed All-Gather memory issue, OpenCLIP, SigLIP sigmoid pairwise loss
  • 10-multimodal-llms.tex --- LLaVA architecture, AnyRes dynamic grid slicing, InternVL-2, LLaVA training stages, Qwen2-VL naive dynamic resolution and M-RoPE
  • 10b-vlm-benchmarks-hallucination.tex --- POPE random/popular/adversarial splits, Visual Contrastive Decoding (VCD), MMMU, DocVQA OCR
  • 11-diffusion-foundations.tex --- DDPM forward/reverse process, Score-based SDE/ODE, Tweedie's formula, EDM continuous \(\sigma\), CFG guidance, linear vs cosine schedule plot
  • 12-latent-diffusion-dit.tex --- Latent Diffusion Models (LDM), VAE \(f=8\) perceptual compression, SDXL conditioning, DiT adaLN-Zero scaling
  • 13-flow-matching-flux.tex --- Flow Matching straight trajectories, velocity field objective, OT-CFM, Rectified Flow reflowing, Flux.1 12B
  • 14-video-models.tex --- 3D Conv, TimeSformer factorized space-time attention, 3D Causal VAE, Sora spacetime patch grid, VBench
  • 15-3d-nerf-3dgs.tex --- NeRF volumetric quadrature rendering, positional encoding, hierarchical sampling, NeRF vs Instant-NGP vs 3DGS table, Instant-NGP spatial hashing, 3DGS covariance projection, tile rasterization
  • 15b-3d-splatting-details.tex --- 3D covariance parameterization, EWA Jacobian projection, Spherical Harmonics (degrees 0--3), 64-bit Radix sort
  • 16-benchmarks-eval.tex --- ImageNet-1K, COCO mAP, Panoptic Quality (SQ/RQ), FID 2-Wasserstein derivation, PSNR vs SSIM vs LPIPS
  • 17-production-edge.tex --- Conv-BN fusion derivation, INT8 PTQ KL divergence calibration, TensorRT plugins, ByteTrack Kalman tracking
  • 18-system-design.tex --- BEVFormer autonomous driving perception pipeline, ColPali late-interaction visual document retrieval RAG
  • 19-code-blind.tex --- Minimal PyTorch stubs: Conv2D im2col, ResNet Bottleneck, ViT PatchEmbed, 2D RoPE, InfoNCE, Vectorized NMS
  • 19b-training-loops.tex --- Execution loops: SimCLR step, DDPM step, DDIM sampler, AnyRes dynamic image slicer
  • 20-napkin-math.tex --- Conv FLOPs, ViT quadratic attention scaling, receptive field recurrence, VLM AnyRes tokens, 3DGS VRAM
  • 20b-napkin-scaling.tex --- Factorized video attention complexity, ViT patch size shrinkage laws, learning rate scaling rules, NPU TOPS
  • 21-debugging.tex --- NaN box regression losses, ViT small data failure, Diffusion gray mush, 3DGS floater needles, VLM visual token loops
  • 22-rapid-glossary.tex --- 20 rapid-fire questions, 20 essential glossary definitions
  • 22b-rapid-fire-drill.tex --- 20 rapid-fire technical questions covering entire vision stack, flashcards, and edge case traps
  • 23-formulas.tex --- Master vision formula sheet covering all 24 architectural mechanisms
  • 24-seminal-papers.tex --- 30 seminal vision papers card (AlexNet to Flux.1 and Sora)
  • 25-canon-references.tex --- Canonical vision bibliography arranged by section with arXiv IDs
  • 26-study-plan.tex --- 21-day revision protocol, phase breakdown, and interview red flags