01
Transformer
Self-attention, KV caching, MoE, LoRA, RLHF, speculative decoding — the architecture that powers GPT, Claude, and Gemini.
12 Architecture Overviews
Modern AI Architecture Atlas
A bird's-eye map of every major AI architecture — what it is, how it works, how it's trained, and how it's used. Interactive diagrams, intuitive explanations, and the research context you need to navigate the field. Not a textbook. Not a tutorial. A compass.
12
Architectures
~15
Min each
48+
Interactive diagrams
The Atlas
02
State Space Models
Selective scan, linear recurrence, and hybrid SSM-attention architectures like Jamba and Zamba. The O(n) alternative.
03
Diffusion Models
DDPM, DDIM, latent diffusion, classifier-free guidance, ControlNet, consistency models — the dominant generative paradigm.
04
Flow Matching
Continuous normalizing flows, optimal transport, conditional flow matching — straighter paths, fewer steps. Behind SD3 and Flux.
05
VAE / VQ-VAE / Tokenizers
Variational inference, ELBO, codebook learning, FSQ — the secret plumbing behind every generative system.
06
GAN
Adversarial training, StyleGAN, spectral normalization, PatchGAN — the OG generative model, still alive at the edges.
07
Contrastive / Representation Learning
CLIP, SimCLR, DINO, MAE, SigLIP — the glue that makes multimodal AI possible.
08
Vision-Language Models
Visual encoder + LLM fusion, instruction tuning, grounding, OCR-free document understanding — GPT-4V and beyond.
09
Vision-Language-Action
Action heads, diffusion policies, action chunking, cross-embodiment training — foundation models that physically act.
10
World Models
JEPA, Dreamer, Genie, UniSim — learned dynamics, latent imagination, and video prediction as world simulation.
11
NeRF / 3D Gaussian Splatting
Neural radiance fields, 3DGS, generative 3D, SDF representations — 3D understanding from 2D images.
12
Reward Models / Alignment
RLHF, DPO, KTO, Constitutional AI, process reward models — making AI systems do what we actually want.