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 ↗
Grokking Deep Learning
Andrew Trask · OpenMined / Google DeepMind / University of OxfordAuthored the seminal deep learning curriculum teaching backpropagation, matrix calculus, convolution, and recurrent architectures using only standard Python and NumPy without frameworks.
Pedagogical pure NumPy implementations lack GPU-accelerated cuDNN kernels for enterprise-scale foundation model pretraining.
Context: Pedagogical algorithmic synthesis, foundational vector mathematics.
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 missionHybrid Search: Merging Sparse and Dense Retrieval for Generative Grounding
Bob van Luijt · WeaviateAuthored the technical blueprint explaining Relative Score Fusion (RSF) and reciprocal rank fusion algorithms for combining sparse lexical matching with dense vector embeddings to maximize RAG precision.
Weighting hyperparameter tuning between keyword and vector dense scores varies by industry domain and query semantics.
Context: BM25 scoring, dense vector similarity, reciprocal rank fusion.
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 missionMinion AI: Autonomous Software Automation Agent
Alex Graveley · Minion AI / Former GitHubEngineered autonomous web and desktop task execution agents utilizing vision-language models, DOM inspection trees, and self-correcting action planning loops.
Dynamic client-side hydration, CAPTCHAs, and non-deterministic visual UI state transitions require resilient fallback recovery mechanisms.
Context: Multimodal LLMs, headless browser drivers, reactive task graphs.
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 missionOptimizing Latent Diffusion Models for Edge & Mobile Deployment
Sayak Paul · Hugging FaceAuthored comprehensive engineering benchmarks and guides demonstrating INT8 and FP16 quantization, TensorRT compilation, and CoreML conversion for running diffusion pipelines on consumer hardware.
Quantization below 8-bit precision requires calibration against representative prompt distributions to prevent latent vector drift.
Context: TensorRT, Core ML, ONNX, mobile inference constraints.
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 missionPySyft: Library for Secure & Private Machine Learning
Andrew Trask · OpenMined / Google DeepMind / University of OxfordEngineered an open-source decentralized privacy framework extending PyTorch and NumPy with Federated Learning, Differential Privacy, and Multi-Party Computation (SMPC) across untrusted nodes.
Encrypted tensor computation induces substantial network latency and memory overhead compared to plain-text distributed SGD.
Context: Python, PyTorch, cryptographic SMPC protocols.
View missionsmol-developer: Autonomous AI Agent First-Principles Scaffold
Shawn Wang · Smol AI / Latent SpaceArchitected a minimalist autonomous junior developer agent implementing hierarchical prompt chaining, markdown codeblock parsing, and iterative multi-file synthesis from human specification prompts.
Relies on single-pass top-down dependency planning without dynamic AST execution sandboxing or test-driven self-healing loops.
Context: Python, Anthropic Claude & OpenAI API agent loops.
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 missionThe Rise of the AI Engineer
Shawn Wang · Smol AI / Latent SpaceAuthored the foundational essay formalizing the emergence of AI Engineering as a distinct discipline bridging fundamental AI research with real-world product engineering, evaluated by thousands of software organizations.
Descriptive industry synthesis rather than empirical algorithmic benchmark.
Context: Industry analysis, historical tech shifts, software engineering economics.
View missionv0.dev: Natural Language Generative User Interface Engine
Guillermo Rauch · VercelConceived and deployed v0, a generative UI system that interprets natural language specifications to stream, validate, and preview accessible React, Tailwind CSS, and Lucide component hierarchies.
Generative UI component trees require human audit for state synchronization, accessibility landmarks, and backend API integration.
Context: React Server Components, Next.js, Anthropic Claude & OpenAI foundation models.
View missionVectara: End-to-End GenAI & Hybrid Search Platform
Amr Awadallah · VectaraArchitected an enterprise GenAI platform integrating hybrid retrieval (BM25 keyword + dense semantic embeddings), cross-attentional reranking, and citation-grounded RAG with automated hallucination scoring.
Cross-encoder reranking adds computational latency to large top-k query passes compared to pure bi-encoder dot-product search.
Context: Distributed vector storage, hybrid neural retrieval, SOC2 Type II compliance.
View missionVercel AI SDK: Unified TypeScript Toolkit for AI Applications
Guillermo Rauch · VercelLed the architecture and release of the Vercel AI SDK (ai), providing provider-agnostic streaming abstractions (streamText, streamObject), unified tool calling, and React hooks (useChat, useCompletion).
Streaming protocols must account for varying token delimiter conventions and rate limits across heterogeneous LLM inference providers.
Context: TypeScript, Node.js, Web Streams API, Server-Sent Events (SSE).
View missionWeaviate: Open-Source AI Vector Search Engine
Bob van Luijt · WeaviateEngineered a low-latency, cloud-native vector database written in Go featuring custom HNSW implementations, product quantization (PQ), and native GraphQL/REST/gRPC interfaces.
High-dimensional HNSW vector graphs require substantial resident memory footprints without product quantization.
Context: Go, vector math SIMD intrinsics, distributed Raft clustering.
View mission