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 ↗
Open X-Embodiment: Robot Learning Datasets and Generalist Policies (RT-X)
Dr. Chelsea Finn · Stanford UniversityCo-led the global Open X-Embodiment consortium uniting 33 academic laboratories across 22 robot types to train RT-X, the world's largest open multi-embodiment robotic policy.
Cross-robot embodiment generalization degrades when sensor modalities or physical gripper geometries differ substantially.
Context: 1M+ robotic trajectory demonstrations across Franka, UR5, Aloha, and mobile manipulators.
View missionParameter-Efficient Transfer Learning for NLP (Foundational PEFT Adapters)
Dr. Colin Raffel · University of Toronto / Vector InstituteCo-authored the landmark ICML 2019 paper introducing lightweight adapter modules injected into transformer layers, proving models could learn new tasks with less than 3% added parameters.
Serial adapter bottlenecks add minor inference latency compared to re-parameterization techniques like LoRA.
Context: Bottleneck projection adapters inserted after multi-head attention and feedforward layers.
View missionProximal Policy Optimization Algorithms (PPO)
Dr. John Schulman · Anthropic / Thinking MachinesInvented PPO, a clipped objective reinforcement learning algorithm that maintains policy stability during stochastic gradient descent, which became the global algorithmic standard for aligning LLMs via RLHF.
Hyperparameter sensitivity to KL penalty coefficients and reward model reward hacking.
Context: Clipped surrogate objective, generalized advantage estimation (GAE), actor-critic networks.
View missionRetrieval-Augmented Generation for Knowledge-Intensive NLP Tasks
Dr. Patrick Lewis · CohereAuthored the foundational NeurIPS 2020 paper defining Retrieval-Augmented Generation (RAG), introducing end-to-end differentiable architectures that ground generative models against external non-parametric neural indices.
Early RAG formulated single-step retrieval; complex multi-hop reasoning requires iterative or agentic retrieval loops.
Context: DPR (Dense Passage Retriever) coupled with BART sequence-to-sequence generation over Wikipedia index.
View missionRewardBench: Evaluating Reward Models for Language Model Alignment
Nathan Lambert, Ph.D. · Allen Institute for AI (Ai2)Created RewardBench, the industry's primary benchmark for evaluating reward models and preference classifiers across chat capabilities, reasoning, and adversarial safety edge cases.
Static evaluation test sets face saturation as training data mixtures incorporate benchmark distributions.
Context: Evaluation suite assessing over 100 reward models across Chat, Reasoning, and Safety axes.
View missionSimCSE: Simple Contrastive Learning of Sentence Embeddings
Dr. Danqi Chen · Princeton UniversityCreated SimCSE, an unsupervised contrastive learning framework that uses standard dropout as data augmentation to learn state-of-the-art sentence embeddings, advancing semantic retrieval without labeled pairs.
Sentence embeddings compress entire paragraphs into single fixed-size vectors, losing granular token-level context.
Context: BERT and RoBERTa backends with contrastive temperature scaling.
View missionSmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models
Dr. Song Han · MIT / HAN LabInvented SmoothQuant, an algorithm that mathematically splits quantization difficulty between activations and weights via per-channel scaling transformations, unlocking 8-bit weight and activation (W8A8) tensor-core acceleration.
Requires support for per-token dynamic scaling or fixed calibration scale tensors in runtime engines.
Context: INT8 matrix multiplications on NVIDIA Ampere/Ada Lovelace tensor cores.
View missionThe Annotated Transformer: A Line-by-Line PyTorch Implementation
Dr. Sasha Rush · Cornell Tech / Hugging FaceAuthored the universally referenced educational codebase providing an exact, executable line-by-line PyTorch implementation of 'Attention Is All You Need', bridging theoretical tensor math with production code.
Presents vanilla quadratic multi-head attention without later FlashAttention or grouped-query optimizations.
Context: PyTorch tensor operations, multi-head attention, positional encoding, and beam search decoding.
View missionThe Hardware Lottery: Algorithmic Stagnation and Hardware Alignment
Dr. Sarah Hooker · Cohere For AIAuthored the CACM article formalizing the 'hardware lottery', demonstrating how deep learning architectures win not on fundamental superiority but on whether they fit contemporary GPU and accelerator architectures.
Qualitative and historical systems critique; guiding architectural paradigms rather than releasing a standalone model.
Context: Historical analysis of neural networks, specialized ASICs, and modern transformer tensor architectures.
View missionThe Mojo Programming Language for AI Hardware Kernel Development
Dr. Chris Lattner · ModularDesigned and built Mojo, a high-performance language that combines Python syntax and developer ergonomics with systems programming primitives, SIMD vectorization, and hardware-accelerated memory ownership.
Ecosystem packages and third-party library bindings are actively expanding relative to decades-old C++/CUDA tooling.
Context: Modular MAX Engine, Mojo compiler, heterogenous CPU/GPU execution pipelines.
View missionTraining Language Models to Follow Instructions with Human Feedback
Dr. John Schulman · Anthropic / Thinking MachinesLed the landmark InstructGPT research demonstrating that applying RLHF with fine-grained human preference data to a 1.3B parameter model produces superior instruction compliance compared to a 175B raw model.
Reward models can over-index on output verbosity and confident tone rather than factual accuracy.
Context: Supervised fine-tuning (SFT), Bradley-Terry reward modeling, PPO reinforcement learning over GPT-3.
View missionTülu 3: Pushing Frontiers in Open Language Model Post-Training
Nathan Lambert, Ph.D. · Allen Institute for AI (Ai2)Co-led the Tülu 3 open post-training initiative, releasing complete training datasets, recipes, and checkpoint weights demonstrating open-source parity with proprietary frontier models through RLVR and DPO.
Verifiable rewards require deterministic verifiers (math/code compilers), limiting RLVR application in subjective writing.
Context: Llama 3.1 base models, Direct Preference Optimization, Reinforcement Learning from Verifiable Rewards (RLVR).
View missionUnsloth: Fast and Memory-Efficient Open-Source LLM Fine-Tuning
Daniel Han · Unsloth AICreated an open-source library that rewrites PyTorch forward and backward passes into handwritten Triton kernels, supporting LoRA, QLoRA, and full fine-tuning with 0% accuracy loss while preventing out-of-memory errors.
Heavily optimized for NVIDIA Ampere, Ada Lovelace, and Hopper architectures; AMD ROCm support is emerging.
Context: Llama 3.1/3.2, Mistral NeMo, DeepSeek-R1-Distill, Gemma 2; 4-bit bitsandbytes and 16-bit mixed precision.
View missionvLLM: High-Throughput and Memory-Efficient LLM Serving Engine
Woosuk Kwon · vLLM / UC BerkeleyEngineered and open-sourced vLLM, an enterprise serving system providing continuous batching, chunked prefill, speculative decoding, and tensor parallelism, serving billions of production inference tokens daily worldwide.
Requires dedicated GPU clusters and NVLink interconnects for multi-node tensor-parallel deployments.
Context: CUDA/ROCm kernels, Triton backends, Ray distributed orchestration, vLLM OpenAI-compatible server.
View missionwhisper.cpp: High-Performance Automatic Speech Recognition in Pure C/C++
Georgi Gerganov · ggml.aiDeveloped a pure C/C++ port of OpenAI Whisper speech recognition with zero external dependencies, streaming audio support, WebAssembly compilation, and hardware acceleration on Apple Silicon and mobile devices.
Real-time streaming requires audio buffering and speech activity detection (VAD) to prevent hallucinated audio loops.
Context: OpenAI Whisper tiny/base/small/medium/large models, 16-bit float and 4-bit/5-bit/8-bit integer quantization.
View missionπ₀ (pi_0): Generalist Physical Intelligence Robot Foundation Model
Dr. Sergey Levine · UC Berkeley / Physical IntelligenceCo-developed π₀, a generalist physical intelligence foundation model trained across diverse robotic embodiments and tasks, demonstrating zero-shot physical manipulation, folding laundry, and assembly.
Physical real-world manipulation requires sub-millisecond control loops and collision avoidance safeguards.
Context: Flow-matching vision-action transformer architecture controlling multi-degree-of-freedom robotic arms.
View mission