AI research and projects
The research, projects, and explanations behind the people. Read the source, inspect verified code and models, and find a collaborator.
A source check confirms that a source supports the summary. It is not an independent replication or endorsement. Our methodology ↗
Scalable Diffusion Models with Transformers (DiT)
Dr. Saining Xie · New York University (NYU)Authored the landmark ICCV 2023 paper proposing Diffusion Transformers (DiT), replacing traditional U-Nets with Transformer backbones operating on latent image patches, establishing power-law compute scaling for generative diffusion.
Full quadratic attention scaling over high-resolution token patches requires patchification or windowing to avoid compute explosion.
Context: Latent Diffusion, Vision Transformer backbones, adaptive layer norm (adaLN-Zero) conditioning.
View missionSmall Language Model (SLM) Fine-Tuning and Inference Optimization
Tuhin Sharma · Bionic / Independent AI AdvisoryEngineered domain-specific distillation pipelines compressing 70B parameter models into 3B-7B parameter edge models for on-premise enterprise deployment with zero data leakage.
SLMs have lower zero-shot generalization on tasks outside the fine-tuning training distribution.
Context: Phi-3, LLaMA-3-8B, LoRA / QLoRA, vLLM, TensorRT-LLM.
View missionSuperhuman AI for Multiplayer Poker: Pluribus
Dr. Noam Brown · OpenAIDeveloped Pluribus, the first AI to defeat elite human professionals in six-player no-limit Texas hold'em, demonstrating that depth-limited search with real-time belief distribution solving can scale to high-dimensional imperfect information.
Imperfect information equilibrium computation is computationally expensive for state spaces exceeding billions of action sequences.
Context: Monte Carlo CFR (Counterfactual Regret Minimization), depth-limited subgame solving.
View missionThe Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks
Dr. Jonathan Frankle · Databricks Mosaic AI / HarvardAuthored the seminal ICLR 2019 Best Paper showing dense, randomly-initialized feed-forward networks contain subnetworks ('winning tickets') that—when trained in isolation from early initialization—reach test accuracy comparable to the original network in similar iterations.
Finding winning tickets requires iterative prune-and-retrain cycles, which is computationally heavier than standard one-pass training.
Context: Iterative Magnitude Pruning (IMP), ResNet, VGG, MNIST, CIFAR-10.
View missionUnpaired Image-to-Image Translation using Cycle-Consistent Adversarial Networks
Dr. Alexei Efros · UC BerkeleyCo-authored the landmark ICCV 2017 CycleGAN paper, introducing cycle consistency loss to enable high-quality visual translation between two unpaired image collections (horses to zebras, summer to winter) with 23,000+ citations.
Struggles with geometric transformations (e.g. cat to dog) because cycle consistency enforces structural topology preservation.
Context: Cycle consistency loss ||G(F(y)) - y||, dual generator/discriminator pairs.
View missionVery Deep Convolutional Networks for Large-Scale Image Recognition (VGG)
Dr. Karen Simonyan · Google DeepMind / Inflection AIAuthored the monumental ICLR 2015 paper demonstrating that pushing network depth to 16-19 layers using homogenous 3x3 convolution filters establishes state-of-the-art vision representations, with over 120,000 academic citations.
Fully connected classifier heads in original VGG architectures contain large parameter counts (138M+) prone to memory bottlenecks.
Context: VGG-16, VGG-19, ImageNet challenge winning architecture, perceptual loss standard.
View missionAccurate Structure Prediction of Biomolecular Interactions with AlphaFold 3
Dr. John Jumper · Google DeepMindCo-led AlphaFold 3 research in Nature 2024, replacing Evoformer representations with a Diffusion Module to predict joint structures of proteins, nucleic acids (DNA/RNA), post-translational modifications, and small-molecule drugs.
Drug binding pocket predictions require experimental crystallization or binding assays for clinical drug validation.
Context: Pairformer architecture coupled with a 3D coordinate diffusion denoising module.
View missionAWQ: Activation-aware Weight Quantization for On-Device LLM Compression and Acceleration
Dr. Song Han · MIT / HAN LabPublished the MLSys Best Paper proving that protecting the top 1% of salient weight channels based on activation magnitudes enables hardware-friendly 4-bit weight-only quantization without costly backpropagation or retraining.
Requires representative calibration datasets to accurately identify activation outliers across specialized domains.
Context: 4-bit integer weights, FP16 activations, evaluated across LLaMA-1/2, Vicuna, StarCoder.
View missionAya: An Open Multilingual Foundation Model and Evaluation Dataset for 101 Languages
Dr. Sarah Hooker · Cohere For AIDirected the global Aya research collaboration involving 3,000+ researchers across 119 countries to produce instruction datasets and state-of-the-art open-weights multilingual foundation models for underrepresented languages.
Extremely low-resource languages with minimal digital footprint require specialized human annotation to avoid hallucination.
Context: Cohere Command base weights, 101 languages, 513M human-curated multilingual instruction tokens.
View missionClassifier-Free Diffusion Guidance (CFG)
Dr. Jonathan Ho · Independent / ex-OpenAIInvented Classifier-Free Guidance (CFG), a mathematical formulation that jointly trains conditional and unconditional diffusion models, dramatically improving prompt fidelity without requiring separate classifier gradients.
High guidance scales cause over-saturation and high-frequency pixel artifacts without dynamic thresholding.
Context: Conditioning dropout during training, weighted score extrapolation during inference sampling.
View missionDeep Residual Learning for Image Recognition (ResNet)
Dr. Kaiming He · MIT EECSFirst author of the CVPR 2016 Best Paper introducing identity skip connections (Residual Networks), overcoming vanishing gradients to train 100+ layer networks, which became standard across all modern deep learning and transformers.
Pure convolutional ResNets possess localized receptive fields, necessitating hybrid attention blocks for global context modeling.
Context: Skip connections (x + F(x)), Batch Normalization, ResNet-50/101/152 architectures.
View missionDenoising Diffusion Probabilistic Models (DDPM)
Dr. Jonathan Ho · Independent / ex-OpenAIFirst author of the groundbreaking NeurIPS 2020 paper that demonstrated diffusion models could surpass GANs and autoregressive models in generative image fidelity via variational bound simplification.
Standard DDPM required hundreds of sequential sampling steps, requiring modern ODE solvers or flow matching for real-time speed.
Context: Gaussian forward diffusion schedule, reverse Markov chain denoising score matching, U-Net architecture.
View missionDense Passage Retrieval for Open-Domain Question Answering (DPR)
Dr. Danqi Chen · Princeton UniversityPublished the landmark EMNLP 2020 paper demonstrating that dual-encoder BERT architectures trained on contrastive negative passages dramatically outperform classic BM25 and TF-IDF sparse retrieval in question answering.
Pure dense retrieval can struggle on out-of-domain lexical keywords and acronyms compared to hybrid dense-sparse search.
Context: Dual BERT encoders, dot-product similarity search over FAISS vector indices.
View missionEfficient Memory Management for Large Language Model Serving with PagedAttention
Woosuk Kwon · vLLM / UC BerkeleyAuthored the seminal SOSP 2023 paper introducing PagedAttention, inspired by operating system virtual memory paging, allowing non-contiguous KV-cache storage and near-zero memory waste during high-concurrency LLM inference.
Paged cache lookup introduces minor pointer indirection overhead in memory-bandwidth-bound token generation loops.
Context: Evaluated across LLaMA, OPT, and Falcon architectures with continuous batching and multi-query attention.
View missionEnd-to-End Training of Deep Visuomotor Policies
Dr. Sergey Levine · UC Berkeley / Physical IntelligencePublished the foundational Journal of Machine Learning Research paper proving that robotic controllers can map raw camera pixels directly to continuous motor torque voltages using guided policy search.
Direct pixel-to-torque policies are vulnerable to novel environmental lighting and visual background shifts.
Context: Convolutional neural network visuomotor policies operating on physical PR2 robotic systems.
View missionExploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer (T5)
Dr. Colin Raffel · University of Toronto / Vector InstituteFirst author of the seminal T5 paper establishing the unified text-to-text paradigm across all NLP tasks, introducing the C4 (Colossal Clean Crawled Corpus) dataset with over 25,000 academic citations.
Encoder-decoder architecture has higher KV-cache memory requirements during generation than pure autoregressive decoders.
Context: Encoder-decoder transformer architecture up to 11B parameters trained on C4 web corpus.
View mission