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 ↗
Hands-on Fine-Tuning Open Source LLMs for Production Tasks
Santiago Valdarrama · UnderfittedCreated open-source walkthroughs showing practitioners how to prepare structured JSONL datasets, apply LoRA adapters via Hugging Face PEFT, and quantize models using bitsandbytes on a single GPU.
Single-GPU fine-tuning restricts effective batch sizes without gradient accumulation and activation checkpointing.
Context: Llama 3 8B, Mistral 7B, PyTorch, Hugging Face TRL and PEFT.
View missionInflection-1 & Pi: Empathetic High-Fidelity Conversational Agents
Mustafa Suleyman · Microsoft AISpearheaded the training and deployment of Inflection-1 and Inflection-2, powering the Pi conversational assistant with natural voice synthesis, low-latency conversational turn-taking, and empathetic persona tuning.
Empathetic conversational tone requires strict guardrails to prevent emotional over-dependence in vulnerable users.
Context: Proprietary foundation model trained on a 22,000 NVIDIA H100 GPU cluster with FP8 mixed precision.
View missionSafeTensors: Fast, Secure Tensor Storage for Deep Learning
Julien Chaumond · Hugging FaceSpearheaded the creation and adoption of SafeTensors, replacing insecure Python pickle files with zero-copy memory-mapped deserialization that prevents arbitrary code execution and boosts model loading speeds 2-10x.
Requires strict contiguous memory layout incompatible with raw arbitrary python object serialization.
Context: Rust-based tensor serialization format adopted universally across Hugging Face, PyTorch, and vLLM.
View missionScale AI Data Engine for Supervised Fine-Tuning & RLHF at Scale
Alexandr Wang · Scale AIEngineered human-in-the-loop expert annotation pipelines generating multi-million token pairs of complex instruction following, code review, and preference datasets powering the training of leading open and commercial models.
Human expert verification costs scale super-linearly with problem complexity in post-doctoral mathematics and legal reasoning.
Context: Distributed annotation platforms with automated inter-annotator agreement metrics and synthetic validation loops.
View missionSEAL Leaderboards: Standardized Private Evals for Frontier LLM Benchmarking
Alexandr Wang · Scale AICreated the SEAL evaluation platform providing contamination-free, expert-curated benchmark suites testing advanced programming (Coding-GSM), mathematical rigor, and adversarial safety to combat public leaderboard gaming.
Requires ongoing prompt refresh cycles to prevent indirect test data leakage into web-crawled pretraining corpora.
Context: Private held-out dynamic evaluation testbeds continuously updated by human domain experts.
View missionTenstorrent Wormhole & Blackhole RISC-V AI Accelerator Architecture
Jim Keller · TenstorrentArchitected the Wormhole and Blackhole processor cards integrating Tensix array cores with native RISC-V scalar processors and on-chip Ethernet switches, enabling direct chip-to-chip 2D mesh networking for LLM tensor parallel training without external PCIe bridges.
Software compiler ecosystem (TT-Metalium) requires continuous software developer adoption to compete with NVIDIA CUDA.
Context: Scalable RISC-V and Tensix matrix cores running Llama 3 and DeepSeek MoE workloads over native Ethernet topology.
View missionTesla Full Self-Driving (FSD) Dual-Chip Computer Architecture
Jim Keller · TenstorrentEngineered Tesla's custom in-car FSD silicon featuring dual independent neural network accelerators processing 2,300 frames per second at under 72 watts, providing redundant lockstep execution for safety-critical real-time automotive vision.
Thermal dissipation in vehicle environments imposes strict wattage ceilings on sustained inference throughput.
Context: Custom dual NPU silicon operating on 8 camera streams at 36-teraops per chip.
View missionThe Covariant Brain: Universal AI for Industrial Robotic Manipulation
Pieter Abbeel · UC Berkeley / CovariantArchitected the commercial robotic perception and control system deployed across North American and European logistics warehouses, autonomously picking, placing, and bagging millions of previously unseen SKUs at 99.9% reliability.
Deformable objects (e.g. polybags, transparent shrink-wrap) require specialized tactile and multi-view sensor fusion.
Context: Multimodal vision-action transformer policies trained on real-world industrial sensor telemetry.
View missionThe Hugging Face Hub: Decentralized Open-Weight ML Infrastructure
Julien Chaumond · Hugging FaceDesigned and scaled the open machine learning repository architecture supporting version-controlled model weights (SafeTensors), streaming dataset loaders, model cards, and automated evaluation leaderboards.
High bandwidth demands require aggressive regional edge caching and chunked tensor streaming.
Context: Global CDN serving petabytes of model checkpoints daily with SafeTensors memory-mapped loading.
View missionThe Machine Learning Reproducibility Checklist & Code Submission Mandates
Joelle Pineau · Meta FAIR / McGill UniversityInstituted the comprehensive standard adopted across NeurIPS, ICML, and ICLR requiring verified disclosures of random seeds, hyperparameter search spaces, computing infrastructure specs, and open code repositories to resolve the ML reproducibility crisis.
Checking hardware parity across multi-thousand GPU clusters remains difficult for smaller independent academic labs.
Context: Adopted across tens of thousands of peer-reviewed machine learning papers globally.
View missionThe Vesuvius Challenge: Machine Learning on Carbonized Papyrus
Nat Friedman · AI Grant / Andromeda ClusterCo-founded and funded the global open scientific prize that mobilized computer vision and 3D deep learning teams to detect hidden ink patterns in non-invasively scanned 3D CT volumes of Herculaneum scrolls sealed in 79 AD.
Papyrus segmentation requires substantial manual inspection and geometric mesh flattening prior to virtual unrolling.
Context: 3D U-Nets and Vision Transformers processing gigavoxel synchrotron X-ray tomographic volumes.
View missionTransformerLens: Open Source Mechanistic Interpretability Framework
Neel Nanda · Google DeepMindCreated and maintains the premier open-source library allowing researchers to hook into, intervene upon, and cache intermediate activations across 50+ autoregressive transformer architectures with PyTorch integration.
Activation patching on models larger than 70B parameters requires substantial distributed multi-GPU memory.
Context: Hook point engine supporting GPT-2, Llama, Pythia, Mistral, and custom transformer variants.
View missionYou.com Multimodal Conversational Search Engine & Real-Time Web RAG
Richard Socher · You.comEngineered one of the first production conversational search engines combining real-time web crawling, zero-click answer synthesis, source URL grounding citations, and interactive code execution in a single chat interface.
Real-time web retrieval latency introduces latency overhead compared to static in-memory parametric generation.
Context: Hybrid BM25 and dense neural retrieval pipeline integrated with frontier LLMs and live web indexing.
View missionAI Due Diligence Framework for Growth-Stage Corporate Acquisitions
Matt Higgins · RSE VenturesDeveloped an institutional investment audit standard assessing target companies' defensibility against generative disruption, proprietary data moat validity, and internal AI workflow integration maturity.
Rapid foundation model capability leaps can render specific vertical software moats obsolete quickly.
Context: Private equity due diligence protocols, proprietary dataset evaluation, tech stack vulnerability audits.
View missionAI Video Production & Voice Cloning Pipeline for Multilingual Media
Shelby Church · Shelby Church MediaDocumented an automated media localization pipeline integrating ElevenLabs, HeyGen, and Descript to translate long-form technical video into 5 languages while preserving voice inflection and lip synchronization.
Complex idioms and regional humor still require human translator cultural inspection.
Context: Voice synthesis, automated transcript timing, AI generative video inpainting.
View missionAI-Augmented Account-Based Marketing (ABM) Personalization Matrix
Nicole Leffer · Nicole Leffer ConsultingBuilt automated research and briefing workflows enabling enterprise sales development reps to tailor outreach to enterprise accounts using contextual AI data extraction from 10-K filings and executive interviews.
Public filing data can be up to 90 days lagging real-time corporate organizational shifts.
Context: SEC EDGAR API, PDF extraction, conversational synthesis, CRM integration.
View mission