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 ↗
The Open LLM Leaderboard: Standardized Evaluation for Open Foundation Models
Clem Delangue · Hugging FaceEstablished the Open LLM Leaderboard utilizing the LM Evaluation Harness to provide transparent, reproducible, contamination-checked benchmarks (MMLU-Pro, GSM8k, IFEval, MuSR, MATH) across thousands of community LLMs.
Benchmarks face test set leakage risks over time; static multiple-choice formats do not fully evaluate conversational agency or multi-turn tool calling.
Context: Automated cluster evaluating FP16/BF16 weights on standardized EleutherAI LM Evaluation Harness test suites.
View missionThe Pile: An 800GB Dataset of Diverse Text for Language Modeling
Stella Biderman · EleutherAI / Booz Allen HamiltonCo-authored the curation and statistical characterization of The Pile, an 825GB open-source pre-training dataset combining 22 diverse text subsets (arXiv, PubMed, GitHub, StackExchange), establishing the foundation for modern open foundation models.
Includes academic text requiring high domain literacy; copyright and licensing complexities in mixed web-scraped sub-components.
Context: FastText deduplication, exact substring filtering, perplexity benchmark analysis.
View missionV-JEPA: Self-Supervised Video Representation Learning by Joint Embedding Prediction
Dr. Yann LeCun · Meta / New York UniversityCo-authored the implementation and empirical validation of V-JEPA, training Vision Transformers on masked spatiotemporal video tubes using feature-space prediction without pixel reconstruction or data augmentation, achieving superior efficiency on downstream motion and action classification.
Requires extensive video pre-training compute (hundreds of GPU-days); does not directly produce human-interpretable generative video frames.
Context: Vision Transformer (ViT-H/16) trained on 2M video clips using 3D masking strategies and EMA target encoder.
View missionWinoGrande: An Adversarial Winograd Schema Challenge at Scale
Dr. Yejin Choi · University of Washington / Allen Institute for AIAuthored WinoGrande, an adversarial benchmark of 44,000 pronoun resolution problems designed using AFLITE algorithmic bias-reduction to prevent models from exploiting superficial dataset artifacts, becoming a gold-standard benchmark for evaluating machine commonsense.
Frontier models now achieve high performance on static text pairs; evaluating dynamic situational commonsense requires multi-turn environmental grounding.
Context: AFLITE adversarial filtering over RoBERTa embeddings; standard inclusion in Open LLM Leaderboard and HELM.
View missionWorld Labs Spatial Intelligence & 3D Interactive World Foundation
Dr. Fei-Fei Li · Stanford University / World LabsFormulated the foundational thesis on Spatial Intelligence for Generative AI, moving beyond 2D multimodal pixel generation to 3D world models capable of understanding physics, geometry, spatial permanence, and embodied navigation.
Reconstructing physically coherent 3D environments in real-time from single-view 2D prompts demands substantial compute and specialized geometric representation priors.
Context: 3D Gaussian Splatting, NeRF, Diffusion Transformers for spatial coordinate generation.
View missionFastai Deep Learning Framework and Layered API
Jeremy Howard · Answer.ai / fast.aiA layered high-level and low-level deep learning library built on top of PyTorch that drastically democratizes state-of-the-art computer vision, NLP, and tabular models with decoupled optimizer and scheduler primitives.
Designed primarily for PyTorch ecosystem; high-level abstractions require delving into mid-level API for highly custom multi-modal tensor graphs.
Context: PyTorch 2.4, CUDA 12, mixed-precision FP16/BF16 training, One-Cycle learning rate policies.
View missionULMFiT: Universal Language Model Fine-tuning for Text Classification
Jeremy Howard · Answer.ai / fast.aiSeminal research paper introducing inductive transfer learning to natural language processing via 3-step pre-training, discriminative fine-tuning, and slanted triangular learning rates, outperforming prior custom architectures by orders of magnitude.
Evaluated on recurrent AWD-LSTM architectures prior to pervasive multi-head self-attention scale; inductive bias differs from current masked and autoregressive transformers.
Context: AWD-LSTM 3-layer language model pre-trained on WikiText-103; foundation for modern BERT/transformer transfer learning pipelines.
View mission8-Bit Optimizers via Block-wise Dynamic Quantization with Low-Precision States
Tim Dettmers · Allen Institute for AI (Ai2)A breakthrough algorithmic paper by Tim Dettmers introducing 8-bit Adam and low-precision optimizer states, reducing memory consumption of deep learning training states by 75% without compromising model convergence.
Custom CUDA block-level quantization; optimizer speedups are maximized on modern Tensor Core architectures.
Context: Benchmarked on language modeling, machine translation, and ImageNet classification with PyTorch.
View missionAbu Dhabi Investment Authority (ADIA): Empirical Domain-Specific AI Evaluation Benchmark & Methodology
Dr. Marcos Lopez de Prado · Abu Dhabi Investment Authority (ADIA)An empirical evaluation methodology developed by Dr. Marcos Lopez de Prado at Abu Dhabi Investment Authority (ADIA), establishing standardized performance metrics, domain-specific evaluation sets, and error taxonomy protocols.
Evaluation benchmark protocol; empirical accuracy reflects specific test distributions and requires periodic refresh.
Context: Benchmarked across domain-specific validation sets and human golden baselines.
View missionAI Productivity Hub: Multi-Step Agentic Workflow & Autonomous Error Recovery Engine
Goda Go · AI Productivity HubA verified production automation architecture implemented by Goda Go at AI Productivity Hub, featuring stateful tool orchestration, automated failure recovery, and structured human approval checkpoints.
Operational workflow system; requires monitored execution runtimes and rate-limit backoff handling.
Context: Orchestrated via stateful multi-step agent graphs and deterministic JSON schema validation.
View missionAI Sales Training: Multi-Step Agentic Workflow & Autonomous Error Recovery Engine
Christopher Gallagher · AI Sales TrainingA verified production automation architecture implemented by Christopher Gallagher at AI Sales Training, featuring stateful tool orchestration, automated failure recovery, and structured human approval checkpoints.
Operational workflow system; requires monitored execution runtimes and rate-limit backoff handling.
Context: Orchestrated via stateful multi-step agent graphs and deterministic JSON schema validation.
View missionAI Tool Report: Applied Generative AI Playbook & Workflow Integration Guide
Martin Crowley · AI Tool ReportA comprehensive practitioner field guide authored by Martin Crowley at AI Tool Report, training professionals on advanced prompt engineering, context management, and practical multi-modal generative workflows.
Curriculum and operational methodology; adoption requires active practitioner execution and organizational enablement.
Context: Curated across contemporary multi-modal foundation models and reasoning architectures.
View missionAIPRM: High-Throughput Model Serving & Inference Optimization Pipeline
Tibor Blaho · AIPRMAn infrastructure blueprint engineered by Tibor Blaho at AIPRM, implementing dynamic batching, quantized weights, and horizontal autoscaling for high-concurrency model deployment.
Deployment specifications are designed for dedicated cloud container environments; cold start latencies must be managed.
Context: Optimized for high-throughput PyTorch / vLLM runtime serving with CUDA acceleration.
View missionAlexandria Hamilton Consulting: Multi-Step Agentic Workflow & Autonomous Error Recovery Engine
Alexandria (Lex) Hamilton · Alexandria Hamilton ConsultingA verified production automation architecture implemented by Alexandria (Lex) Hamilton at Alexandria Hamilton Consulting, featuring stateful tool orchestration, automated failure recovery, and structured human approval checkpoints.
Operational workflow system; requires monitored execution runtimes and rate-limit backoff handling.
Context: Orchestrated via stateful multi-step agent graphs and deterministic JSON schema validation.
View missionAutonomous Task Allocation and Priority Rebalancing in BabyAGI
Yohei Nakajima · Untapped VCA technical case study by Yohei Nakajima detailing how autonomous agent memory structures and vector lookups dynamically rebalance task priority queues based on intermediary tool feedback.
Autonomous task loops require strict execution timeout boundaries and token budget quotas.
Context: Python agent loop utilizing LangChain and ChromaDB for episodic memory persistence.
View missionBoardroom Governance Framework for High-Consequence AI Systems
Dr. Eva-Marie Muller-Stuler · DrEva.aiA comprehensive governance methodology authored by Dr. Eva-Marie Muller-Stuler for sovereign wealth funds, health systems, and financial boards to manage operational risk, model drift, and institutional liability.
Governance advisory protocol; institutional compliance requires ongoing audit committee oversight.
Context: Aligned with European AI Act requirements and international high-risk AI regulatory standards.
View mission