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 ↗
A Style-Based Generator Architecture for Generative Adversarial Networks (StyleGAN)
Tero Karras · NVIDIA ResearchInvented the style-based generator architecture for GANs, introducing adaptive instance normalization (AdaIN) to separate high-level facial attributes (pose, identity) from stochastic variations (hair, freckles).
Susceptible to characteristic water-droplet artifacts in feature maps before the StyleGAN2 weight demodulation redesign.
Context: FFHQ dataset, NVIDIA DGX-1 clusters, custom CUDA kernels.
View missionEmergent Social Learning via Multi-Agent Reinforcement Learning
Natasha Jaques · University of Washington / Google DeepMindDemonstrated that multi-agent reinforcement learning with intrinsic social curiosity incentives enables agents to spontaneously acquire social learning, imitation, and cultural transmission without human expert demonstrations.
Social imitation policies can converge to suboptimal collective traps if demonstrator agents experience local reward plateaus.
Context: Deep RL, multi-agent gridworld environments, social reward functions.
View missionHughes Hallucination Evaluation Model (HHEM)
Amr Awadallah · VectaraEngineered an open-source factual consistency classifier calculating the precise probability that an LLM-generated summary hallucinated information not present in the reference context document.
Focuses on text summarization fidelity; does not evaluate arithmetic reasoning or multi-hop logical deductions.
Context: RoBERTa-based factual entailing model, Hugging Face leaderboards, benchmark datasets.
View missionInstant Neural Graphics Primitives with a Multiresolution Hash Encoding (Instant NGP)
Tero Karras · NVIDIA ResearchIntroduced multiresolution hash encoding for neural graphics primitives, accelerating NeRF and signed distance field training from hours to seconds on a single GPU.
Hash collisions can induce high-frequency reconstruction noise in complex out-of-distribution scenes without adequate spatial bounding.
Context: Custom CUDA spatial hash tables, tiny-cuda-nn, real-time NeRF rendering.
View missionNeural Architecture Search with Reinforcement Learning
Barret Zoph · OpenAI / Former Google BrainIntroduced Neural Architecture Search (NAS), employing a recurrent neural network controller trained via reinforcement learning to generate model descriptions and automate neural network architecture design, winning seminal citations.
Original NAS required massive compute expenditures (thousands of GPU hours) before differentiable architecture search (DARTS) was invented.
Context: CIFAR-10, Penn Treebank, distributed TPU clusters, policy gradients.
View missionPAIRED: Emergent Complexity and Zero-shot Transfer via Adversarial Environment Generation
Natasha Jaques · University of Washington / Google DeepMindInvented Protagonist Antagonist Induced Regret Environment Design (PAIRED), using an adversarial teacher agent to procedurally generate training curriculums that optimize agent regret and zero-shot out-of-distribution transfer.
Adversarial curriculum generators can encounter unstable training dynamics when reward estimation variances are high.
Context: NeurIPS 2020 spotlight, adversarial RL, procedural generation.
View missionST-MoE: Designing Stable and Transferable Sparse Mixture-of-Experts Models
Barret Zoph · OpenAI / Former Google BrainAuthored the foundational engineering playbook for training multi-hundred-billion parameter Sparse Mixture-of-Experts (MoE) models, introducing the router z-loss to eliminate numerical instability and routing collapse.
Sparse routing creates communication-heavy all-to-all cross-accelerator collective operations that require high-bandwidth interconnects (NVLink/InfiniBand).
Context: TPU v4 pods, Megatron/Jax, multi-billion parameter foundation language models.
View missionA Neural Probabilistic Language Model
Yoshua Bengio · Mila - Quebec AI Institute / Université de MontréalIntroduced the seminal concept of learning continuous distributed word feature vectors alongside statistical language modeling, defeating the curse of dimensionality and establishing modern neural language modeling.
Computational bottleneck in softmax normalization over large vocabularies prior to hierarchical softmax and negative sampling.
Context: Feed-forward neural network sharing distributed representation layers across context words.
View missionAdversarial Attacks on Neural Networks in Physical and Digital Environments
Dawn Song · UC BerkeleyDemonstrated that adversarial perturbations crafted via gradient-based optimization survive physical-world transformations (such as camera capture, lighting variations, and printing), exposing critical safety vulnerabilities in autonomous perception systems.
Defending against physical attacks often degrades clean-sample classification accuracy on subtle visual nuances.
Context: Inception and ResNet vision models subjected to physical adversarial stickers and perturbations.
View missionAI Feynman: A Physics-Inspired Method for Symbolic Regression
Max Tegmark · MIT / Future of Life InstituteDeveloped AI Feynman, a recursive neural network algorithm combining dimensional analysis, symmetry detection, and polynomial decomposition to discover exact symbolic mathematical laws from raw scientific experimental data.
Exponential complexity scaling in high-dimensional systems without clear separability or symmetries.
Context: Neural networks trained to identify translational, rotational, and scaling invariants across 100 equations from the Feynman Lectures on Physics.
View missionChain-of-Thought Prompting Elicits Reasoning in Large Language Models
Hyung Won Chung · OpenAICo-authored the foundational paper discovering that generating step-by-step reasoning tokens before the final answer dramatically unlocks multi-step arithmetic, commonsense, and symbolic reasoning in foundation models.
Chain-of-thought benefits appear as an emergent property only in models with sufficient scale (~100B+ parameters or instruction-tuned compact models).
Context: PaLM 540B, GPT-3 175B tested on GSM8K math benchmark and SVAMP.
View missionConstitutional AI: Harmlessness from AI Feedback (RLAIF)
Dario Amodei · AnthropicIntroduced the Constitutional AI paradigm enabling models to critique, revise, and align their own behavioral outputs according to written principles without human-labeled toxicity feedback. Established the foundational technique powering Claude's red-teaming resilience.
Constitution design must be continuously audited for latent political biases, conflicting clauses, and nuanced real-world legal edge cases.
Context: Frontier autoregressive transformer checkpoints trained via self-critique, Chain of Thought, and Reinforcement Learning from AI Feedback (RLAIF).
View missionDeep Reinforcement Learning from Human Preferences (RLHF)
Paul Christiano · Alignment Research Center (ARC)Authored the foundational paper establishing that training deep neural policies via pairwise human preference reward models allows agents to master complex tasks (Atari games and robotic backflips) without hard-coded programmatic reward functions.
Reward models can be gamified by agents via reward hacking and sycophantic responses.
Context: PPO actor-critic architectures and learned Bradley-Terry reward models.
View missionDota 2 with Large Scale Deep Reinforcement Learning (OpenAI Five)
Greg Brockman · OpenAILed the engineering and systems team that built OpenAI Five, scaling Proximal Policy Optimization (PPO) across 128,000 CPU cores and 256 GPUs playing 180 years of game simulation daily, defeating reigning world champion esports team OG.
Required heavy reward engineering and hero draft restrictions to converge on winning policies in complex game spaces.
Context: 4,096-unit single-layer LSTM processing continuous state-action spaces with 20,000-step time horizons.
View missionEureka: Human-Level Reward Design via Coding Large Language Models
Anima Anandkumar · California Institute of Technology (Caltech)Created Eureka, an autonomous algorithm leveraging frontier coding LLMs to synthesize, evaluate, and iteratively refine complex reward algorithms for reinforcement learning, outperforming human expert-designed rewards across 29 dexterous robot manipulation tasks.
Requires fast GPU simulation environments to provide high-throughput performance feedback to the outer-loop LLM.
Context: GPT-4 reward coding loops interacting with NVIDIA Isaac Gym GPU physics simulations.
View missionFourier Neural Operator for Parametric Partial Differential Equations (FNO)
Anima Anandkumar · California Institute of Technology (Caltech)Pioneered the Fourier Neural Operator (FNO), formulating a deep learning architecture that directly learns mappings between infinite-dimensional function spaces via Fast Fourier Transforms, achieving mesh-independent zero-shot super-resolution on Navier-Stokes fluid turbulence 1,000x faster than standard numerical solvers.
Periodic boundary condition assumptions require specialized padding/coordinate embeddings for complex arbitrary 3D boundaries.
Context: Fourier spectral convolution layers evaluated on 2D/3D Navier-Stokes, Darcy Flow, and global atmospheric weather forecasting (FourCastNet).
View mission