Curiosity
Learn mathematics, physics, and computer science through interactive visualizations. Build deep intuition by playing with the concepts.
Recently Added
Latest research paper breakdowns
Mask R-CNN
Instance segmentation redefined. Extending Faster R-CNN to pixel-level prediction with a parallel mask branch.
EfficientNet
Scaling models with compound scaling—optimally scaling depth, width, and resolution for better accuracy with fewer parameters.
SENet
Squeeze-and-Excitation networks: channel-wise attention that lets models focus on what's important.
SimCLR
Simple framework for contrastive learning. The breakthrough that made self-supervised vision competitive with supervised methods.
BYOL
Bootstrap Your Own Latent. Contrastive learning without negative pairs—using two views and asymmetric networks.
CLIP
Connecting vision and language. Learning visual concepts from natural language supervision for zero-shot transfer.
CycleGAN
Unpaired image-to-image translation. Learning bidirectional mappings without aligned training pairs.
StyleGAN / StyleGAN2
Generator architecture for high-resolution synthesis. Style-based generation, AdaIN, and controlling image attributes.
DALL-E / Latent Diffusion
From text to image. Understanding VAEs, diffusion models, and the LDM architecture powering generative AI.
GPT-2
Language models are unsupervised multitask learners. The architecture and scaling that started the LLM revolution.
PaLM
Pathways Language Model. Scaling to 540B parameters with mixture-of-experts and the training breakthroughs.
Chain-of-Thought
Prompting for reasoning. How structured reasoning chains dramatically improve LLM performance on complex tasks.
Neural ODEs
Neural Ordinary Differential Equations. Continuous-depth networks and adaptive computation for memory-efficient modeling.
NeRF
Neural Radiance Fields. Representing scenes as continuous 5D functions for photorealistic novel view synthesis.
Perceiver / Perceiver IO
General-purpose perception with transformers. Fixed latent size, cross-attention, and handling arbitrary input modalities.
AlphaGo Zero
Mastering Go without human data. Self-play reinforcement learning with MCTS and value/policy networks.
Kimi Linear
The first linear attention to outperform full attention: 6x faster decoding, 75% memory reduction at million-token contexts.
Engram
Conditional Memory via Scalable Lookup: a new sparsity axis for LLMs using O(1) n-gram retrieval with 2.8% offload overhead for 100B params.
DeepSeek-R1
Incentivizing Reasoning in LLMs: how pure RL with simple rewards produces emergent reasoning, self-reflection, and 'aha moments'.
DeepSeek-OCR 2
Visual Causal Flow: how to make AI read images the way humans do—by semantic meaning, not pixel order.
DeepSeek-OCR
Contexts Optical Compression: achieving 97% accuracy with 16× token compression for production-scale document AI.
Segment Anything (SAM)
The foundation model that changed computer vision: segment any object with a single click, zero-shot transfer to any domain.
Muon Optimizer
Matrix orthogonalization meets LLM training: 2× compute efficiency over AdamW, explained from first principles.
Attention Residuals
Replacing identity residuals with learned depth-attention: selective information routing across transformer layers.
Browse by Category
22 categories · 291 topics