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 ↗
ResNet Strikes Back: An Improved Training Procedure in PyTorch
Ross Wightman · Hugging FaceCo-authored the NeurIPS 2021 benchmark study demonstrating that modern training recipes (BCE loss, RandAugment, AdamW) lift standard ResNet-50 accuracy from 76.5% to 80.4% without any architectural changes.
Heavy data augmentations significantly increase training epochs and compute overhead compared to vanilla SGD.
Context: ResNet-50, LamB/AdamW optimizers, repeated augmentations.
View missionRobust Speech Recognition via Large-Scale Weak Supervision (Whisper)
Alec Radford · OpenAIArchitected Whisper, an open-weights sequence-to-sequence Transformer trained on 680,000 hours of multilingual audio, establishing zero-shot robustness across accents, background noise, and automated timestamp generation.
Long audio files can experience hallucination loops or timestamp drift during extended periods of ambient silence.
Context: Seq2Seq Transformer, 16kHz log-mel spectrogram audio input, 99 language recognition.
View missionSandboxed Code Execution Protocols for Autonomous LLM Agents
Sulaiman Ghori · CognosysDeveloped security isolation protocols utilizing gVisor and WebAssembly to safely execute arbitrary agent-generated code with network boundary controls.
Container cold-start latency introduces 1-2 second delays on initial tool execution invocations.
Context: gVisor, WASM runtime, Linux cgroups, LLM tool dispatch.
View missionScalable Diffusion Models with Transformers (DiT)
Dr. Saining Xie · New York University (NYU)Authored the landmark ICCV 2023 paper proposing Diffusion Transformers (DiT), replacing traditional U-Nets with Transformer backbones operating on latent image patches, establishing power-law compute scaling for generative diffusion.
Full quadratic attention scaling over high-resolution token patches requires patchification or windowing to avoid compute explosion.
Context: Latent Diffusion, Vision Transformer backbones, adaptive layer norm (adaLN-Zero) conditioning.
View missionSGLang: Fast Serving Framework for Complex Language Model Programs
Lianmin Zheng · LMSYS Org / UC BerkeleyEngineered SGLang featuring RadixAttention, which maintains a radix tree over KV caches to enable automatic multi-turn and prefix cache reuse, delivering up to 5x higher throughput for complex multi-call LLM workflows.
Radix tree traversal introduces slight metadata lookup overhead for workloads with completely random, unshared prefixes.
Context: DeepSeek-V3/R1, LLaMA-3, Qwen-2.5; FlashInfer integration; Triton custom kernels.
View missionSmall Language Model (SLM) Fine-Tuning and Inference Optimization
Tuhin Sharma · Bionic / Independent AI AdvisoryEngineered domain-specific distillation pipelines compressing 70B parameter models into 3B-7B parameter edge models for on-premise enterprise deployment with zero data leakage.
SLMs have lower zero-shot generalization on tasks outside the fine-tuning training distribution.
Context: Phi-3, LLaMA-3-8B, LoRA / QLoRA, vLLM, TensorRT-LLM.
View missionSpiral: Natural Language Creative Writing and Synthesis Engine
Dan Shipper · Every.toArchitected Spiral, an AI writing tool designed to capture a writer's idiosyncratic voice and transform rough brain-dumps into polished essays using recursive prompt chaining.
Capturing idiosyncratic writing voice requires extensive fine-grained few-shot style examples.
Context: Claude 3.5 Sonnet, recursive LLM prompt chains, personal style embedding.
View missionStatistical Testing and Evaluation Metrics for Stochastic AI Systems
Cassie Kozyrkov · Data ScientificAuthored widely cited methodologies on designing rigorous out-of-sample test splits, avoiding p-hacking in model benchmarks, and setting risk tolerance thresholds for generative AI.
Generative tasks with open-ended creative outputs require human pairwise rubrics alongside automated statistics.
Context: Hypothesis testing, out-of-sample validation, statistical power analysis.
View missionSuperhuman AI for Multiplayer Poker: Pluribus
Dr. Noam Brown · OpenAIDeveloped Pluribus, the first AI to defeat elite human professionals in six-player no-limit Texas hold'em, demonstrating that depth-limited search with real-time belief distribution solving can scale to high-dimensional imperfect information.
Imperfect information equilibrium computation is computationally expensive for state spaces exceeding billions of action sequences.
Context: Monte Carlo CFR (Counterfactual Regret Minimization), depth-limited subgame solving.
View missionThe AI Exchange Community Library of Automated Workflows
Rachel Woods · The AI ExchangeCurated and validated hundreds of production business automation templates integrating LLMs with Zapier, Make, and Airtable for lead scoring and data extraction.
Connector reliability depends on third-party webhook uptime and API schema stability.
Context: Make.com, Zapier, Airtable, OpenAI API, Anthropic Claude.
View missionThe AI Solopreneur System: Operational Workflows for Solo Founders
Bardia (Barsee) · The AI Solopreneur / Independent AgencyDeveloped repeatable prompt workflows and automation templates enabling solo founders to manage customer research, marketing analysis, and content synthesis with 90% reduced manual effort.
Requires ongoing prompt maintenance as underlying frontier model system instructions update.
Context: Claude 3.5 Sonnet, GPT-4o, Perplexity AI, Zapier automation webhooks.
View missionThe AI Use Case Assessment Framework: Scoping and ROI Measurement
Rachel Woods · The AI ExchangeCreated the widely adopted enterprise framework for vetting generative AI ideas, scoring feasibility, calculating token and labor cost delta, and predicting time-to-value.
Intangible productivity gains (e.g. employee creativity) are harder to capture than direct headcount hours.
Context: AI ROI calculation, risk vs feasibility matrices, operational workflow mapping.
View missionThe Architecture of Agentic Go-to-Market Workflows
Paul Yacoubian · Copy.aiPublished case studies and enterprise deployment playbooks illustrating 80% reduction in customer acquisition operational overhead by replacing manual SDR tasks with validated agentic pipelines.
Requires strong human governance on external messaging to prevent automated brand reputation mishaps.
Context: LLM synthesis, structured output validation, HubSpot and Salesforce bidirectional sync.
View missionThe Architecture of Conversational Programming and AI-Native Workflows
Dan Shipper · Every.toAuthored landmark essays dissecting how solo founders build complete production web applications via conversational dialogue with Claude and Cursor.
Conversational programming requires structured git commits to roll back erroneous multi-file generations.
Context: Cursor, Claude 3.5 Sonnet, Next.js, prompt-driven refactoring.
View missionThe Architecture of Low-Latency Voice-to-Voice Foundation Systems
Kanjun Qiu · Outspeed / Generally IntelligentPublished foundational engineering breakdowns on chunked audio tokenization, speculative speech generation, and interruption handling in spoken human-AI dialogue.
Interruption handling latency is constrained by acoustic round-trip echo feedback.
Context: Voice activity detection (VAD), speculative decoding, speech token streaming.
View missionThe Chatbot Era Is Ending. Your AI Team Is Next.
Igor Pogany · The AI AdvantageDesigned structured prompt frameworks and context-setting blueprints adopted by over 200,000 professionals to automate operational drafting, customer segmentation, and market research.
Natural language prompts cannot enforce deterministic relational constraints without middleware validation.
Context: GPT-4o, Claude 3.5 Sonnet, Custom GPT knowledge files, structured system prompts.
View mission