Generative Models & Frontier Research

Each step builds on the one before. Open any card for the paper and resources.

1. Generative models

   7 algorithms

   Variational Autoencoder (VAE) Encodes data as a distribution in latent space and decodes samples into new data. 2013core Denoising Diffusion Probabilistic Models (DDPM) Learns to turn noise into images step by step. 2020core Latent Diffusion Models Runs diffusion in a VAE's compressed latent space to cut cost. 2021core Classifier-Free Guidance (CFG) Extrapolates between conditional and unconditional predictions to follow prompts. 2022core Diffusion Transformer (DiT) Replaces the U-Net denoiser with a Transformer. 2022core Flow Matching Learns a velocity field that flows noise into data. 2022core Rectified Flow Straightens generative paths so few or one step suffices. 2022ext

2. Multimodality

   5 algorithms

   Vision Transformer (ViT) Treats 16×16 image patches as words for a standard Transformer. 2020core CLIP (Contrastive Language–Image Pretraining) Aligns images and text in one embedding space from 400M pairs. 2021core Visual Instruction Tuning (LLaVA) Connects a vision encoder to an LLM via a projector, tuned on visual instructions. 2023core Whisper General speech recognition trained on 680K hours of weakly labeled web audio. 2022core Segment Anything (SAM) A promptable foundation model that segments any object from a click or box. 2023core

3. New architectures

   6 algorithms

   Mamba (Selective State Spaces) Input-dependent state spaces with linear-time cost and Transformer-like quality. 2023core Linear Attention Approximates softmax with kernels to make attention linear in length. 2020core Gated Delta Networks Linear attention combining delta-rule memory updates with gated forgetting. 2024core Hybrid Architectures (Jamba) Interleaves attention and Mamba layers (plus MoE) in one model. 2024core Diffusion Language Models (LLaDA) Generates text by iteratively unmasking a fully masked sequence. 2025core Native Sparse Attention (NSA) Trainable, hardware-aligned dynamic sparse attention. 2025core

4. World models

   4 algorithms

   World Models Learn a dreamable environment model, then train the policy inside it. 2018core Dreamer (Latent Imagination) Trains behavior by imagining trajectories in a learned latent world model. 2019core Joint-Embedding Predictive Architecture (JEPA) Predicts masked content in representation space rather than pixels. 2023core Genie (Generative Interactive Environments) Learns playable, controllable worlds from unlabeled videos. 2024core

5. Opening the black box

   4 algorithms

   Superposition Models pack more features than dimensions by superimposing them. 2022core Sparse Autoencoders (SAE) Decomposes activations into many sparse, interpretable features. 2023core Circuit Tracing & Attribution Graphs Maps causal feature pathways behind a specific response. 2025core Activation Steering Adds a direction vector to activations at inference to change behavior. 2023core

Goals, projects and self-tests that go with these algorithms.

• 1Zero to Deep Learning · From linear regression all the way to ResNet and the precursors of the Transformer — build intuition for how models are trained.24
• 2Understanding an LLM End to End · From tokenisation to the Transformer, modern architectures, pre-training and inference — every part of an LLM, in nanoGPT order.19
• 3Post-training & Reasoning Models · From instruction tuning and LoRA to RLHF and DPO, then GRPO, RLVR and DeepSeek-R1 — how ChatGPT and o1/R1 are made.19
• 4AI Application Engineer · For product engineers: retrieval, RAG, agents, inference serving and security — turn AI into reliable applications fast.20