PUBLISHED AI RESEARCH AND PROJECTS

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 ↗

Showing 353–368 of 567 proofs of work · Newest first
Page 23 of 36
implementationChecked Sep 20, 2026

AutoResearch: Autonomous AI Research Agent Conducting Iterative ML Experiments

Andrej Karpathy · Eureka Labs / ex-Tesla, OpenAI

An autonomous multi-step research agent framework designed to conduct end-to-end machine learning experimentation without human intervention. The agent autonomously formulates algorithmic hypotheses, modifies PyTorch model architectures via AST rewriting, schedules multi-GPU training jobs, parses validation loss curves, and generates structured empirical research reports and pull requests.

Scope & limitations

Requires sandboxed execution environments to prevent arbitrary code execution risks during dynamic model definition; experiments can diverge if loss curves suffer from exploding gradients without bounded learning rate search spaces.

Context: PyTorch 2.4, LLM function calling, Automated Evaluation Loop, Python AST manipulation, TensorBoard metric parsing.

View mission
implementationChecked Sep 20, 2026

llm.c: Large Language Model Training in Pure C/CUDA Without Dependencies

Andrej Karpathy · Eureka Labs / ex-Tesla, OpenAI

A zero-dependency, pure C and CUDA implementation of GPT-2 and modern transformer pre-training. Eliminates the 245MB+ PyTorch/Python runtime overhead by directly implementing forward passes, backward passes, AdamW optimizers, and fused CUDA kernels in a single clean compilation unit, achieving near-parity with PyTorch cuDNN speed.

Scope & limitations

Lower level of architectural expressiveness compared to dynamic Python autograd graphs; adding novel attention variations requires manual mathematical derivation and custom CUDA kernel writing.

Context: Pure C99, CUDA 12, FP32 and BF16 mixed-precision kernels, fused attention, multi-GPU MPI execution.

View mission
explanationChecked Sep 20, 2026

micrograd: Minimal Scalar-Valued Autograd Engine with Dynamic DAG Visualization

Andrej Karpathy · Eureka Labs / ex-Tesla, OpenAI

A foundational, pedagogical autograd engine implementing reverse-mode autodiff over dynamically constructed Directed Acyclic Graphs (DAGs) in approximately 100 lines of pure Python, alongside a multi-layer perceptron (MLP) library that replicates PyTorch's core tensor mechanics from first principles.

Scope & limitations

Operates on individual scalar values rather than vectorized multidimensional tensors, making it unsuitable for training production models beyond small educational datasets.

Context: Pure Python 3.10+, scalar-valued autodiff DAG, Graphviz topological computation graph visualization.

View mission
implementationChecked Sep 20, 2026

Automated Code Review and PR Triage AI Assistant

Michael Ballard · Slashdev

Built a GitHub Action and webhook bot that analyzes incoming pull requests, identifies security vulnerabilities and formatting anti-patterns, and generates unit test coverage proposals automatically.

Scope & limitations

Large pull requests exceeding 2,000 lines require chunking across logical Git commits to fit within prompt context ceilings.

Context: Claude 3.5 Sonnet, GitHub Octokit API, AST tree-sitter parsers.

View mission
implementationChecked Sep 20, 2026

Automated Content Repurposing Engine for Long-Form Media

Paul Macko · Deplatformable / Macko Media

Built an end-to-end automated workflow that ingests audio podcasts, generates precise Whisper transcripts, identifies viral hooks using LLM evaluation, and renders styled short-form video clips automatically.

Scope & limitations

Multi-speaker audio with heavy background music can lead to transcript misalignment without voice separation preprocessing.

Context: Whisper API, Claude 3.5 Sonnet, FFmpeg command automation, Python script runner.

View mission
implementationChecked Sep 20, 2026

High-Speed Optical Character Recognition and Table Extraction Engine

Daniil Komov · SentiSight.ai / Neurotechnology

Developed an end-to-end OCR and table structure recognition engine capable of processing warped or skewed scans at under 100ms per page with 98.7% character accuracy across Latin and Cyrillic scripts.

Scope & limitations

Heavily degraded handwritten documents require specialized handwriting transformer (HTR) checkpoints.

Context: Convolutional recurrent neural network (CRNN) with CTC loss, line segmentation heuristics.

View mission
implementationChecked Sep 20, 2026

Self-Healing Kubernetes CI/CD Pipeline with AI Log Anomaly Detection

Michael Ballard · Slashdev

Engineered an automated observability pipeline that ingests Kubernetes container logs, flags runtime memory spikes or deadlocks, and automatically recommends or executes configuration rollbacks.

Scope & limitations

High false-alarm rates during deliberate rolling updates require suppression windows during planned maintenance deployments.

Context: Prometheus, Grafana Loki, Mistral-7B-Instruct fine-tune for root cause analysis.

View mission
explanationChecked Sep 20, 2026

The Deplatformable Guide to Local LLMs and Self-Hosted Data Sovereignty

Paul Macko · Deplatformable / Macko Media

Published an authoritative guide detailing the hardware specifications, quantization trade-offs (GGUF, EXL2), and network sandboxing needed to run private foundation models offline on consumer Apple Silicon and NVIDIA workstations.

Scope & limitations

Consumer hardware memory limits restrict inference speeds when running 70B parameter models without substantial unified RAM.

Context: Ollama, LM Studio, llama.cpp, Llama 3 8B/70B quantized models.

View mission
implementationChecked Sep 20, 2026

Web-Based Deep Learning Model Training and Annotation Platform

Daniil Komov · SentiSight.ai / Neurotechnology

Architected a cloud-based platform enabling non-technical users to annotate images with bounding boxes, polygons, and keypoints, and train custom classification, object detection, and OCR models with one click.

Scope & limitations

Training speed depends on GPU cluster queue availability during peak multi-user training workloads.

Context: PyTorch, ONNX Runtime, WebSockets for live training metrics, automated augmentations.

View mission
researchChecked Sep 20, 2026

A Path Towards Autonomous Machine Intelligence (JEPA Architecture Framework)

Dr. Yann LeCun · Meta / New York University

Published the comprehensive theoretical blueprint proposing Joint Embedding Predictive Architecture (JEPA) as an alternative to generative pixel-level or token-level prediction, learning representations in abstract embedding space to enable hierarchical planning.

Scope & limitations

Action-conditional latent dynamics models remain active research topics; translating abstract latent representations back into fine-grained execution paths requires auxiliary decoding heads.

Context: Self-supervised non-contrastive energy-based models; architectural basis for I-JEPA (vision) and V-JEPA (video).

View mission
explanationChecked Sep 20, 2026

AI Snake Oil: What Computers Can, Can't, and Shouldn't Do

Dr. Sayash Kapoor · Princeton University

Developed the analytical distinction between perception AI (generative synthesis, speech recognition) and predictive AI (predicting human recidivism, job success), providing rigorous frameworks for evaluating claims of AI accuracy in business and governance.

Scope & limitations

Practical applications in enterprise require navigating existing regulatory guidelines that may lag behind empirical machine learning consensus.

Context: Statistical risk assessment, cross-entropy evaluation, algorithmic bias metrics in high-stakes human systems.

View mission
researchChecked Sep 20, 2026

Attention Is All You Need: The Seminal Transformer Architecture Paper

Aidan Gomez · Cohere

Co-authored the groundbreaking NeurIPS 2017 research paper introducing the Transformer, an architecture dispensing entirely with recurrence and convolutions in favor of multi-head self-attention mechanisms, founding modern LLM technology.

Scope & limitations

Original architecture scaled quadratically O(N^2) with sequence length, prompting subsequent research into grouped-query attention, sparse attention, and linear recurrence hybrids.

Context: 8-head multi-head self-attention, sinusoidal positional encodings, trained on 8 NVIDIA P100 GPUs.

View mission
explanationChecked Sep 20, 2026

Auditing Multimodal Generative Systems for Representational and Allocative Harm

Dr. Margaret Mitchell · Hugging Face

Formulated structured auditing protocols for text-to-image and multimodal foundation models, measuring bias propagation across CLIP embeddings, latent diffusion prompt expansions, and synthetic face generation.

Scope & limitations

Balancing automated quantitative metric auditing with contextual qualitative evaluation requires domain-specific sociological input.

Context: CLIP, Stable Diffusion, demographic diversity benchmarking over fine-grained prompt variations.

View mission
researchChecked Sep 20, 2026

Bottleneck Transformers for Visual Recognition (BoTNet)

Aravind Srinivas · Perplexity AI

Co-authored BoTNet, an architecture that replaces the final spatial convolutions in ResNet with Multi-Head Self-Attention, achieving 84.7% Top-1 accuracy on ImageNet while substantially improving instance segmentation and object detection on COCO.

Scope & limitations

Quadratic memory complexity in early spatial layers restricts attention blocks to deeper, downsampled feature maps in the network.

Context: ResNet backbone with Multi-Head Self-Attention bottleneck blocks, TPU v3 accelerators, 2D relative position encodings.

View mission
researchChecked Sep 20, 2026

Chain-of-Thought Prompting Elicits Reasoning in Large Language Models

Dr. Jason Wei · OpenAI

Authored the seminal NeurIPS 2022 research paper demonstrating that generating intermediate reasoning steps dramatically boosts LLM performance on arithmetic, symbolic, and multi-hop reasoning tasks.

Scope & limitations

Chain-of-thought gains exhibit emergence only at significant model scale (~50B+ parameters); smaller models can generate superficially coherent rationales that lead to incorrect arithmetic conclusions.

Context: PaLM 540B, GPT-3 175B, GSM8K, SVAMP, BIG-bench reasoning tasks.

View mission
implementationChecked Sep 20, 2026

Cohere Command R+ Enterprise Agentic Retrieval & Tool Calling Engine

Aidan Gomez · Cohere

Directed the development of Command R+, an open-weights 104B parameter model optimized specifically for multi-step agentic workflows, multi-lingual enterprise Retrieval-Augmented Generation (RAG), and cited tool calling across complex relational schemas.

Scope & limitations

Heavy memory footprint requiring 4x NVIDIA A100 (80GB) or quantized 8-bit / 4-bit deployments in production environments.

Context: 104B parameters, 128k context window, native citation grounding, multi-turn tool use API schema.

View mission
Showing 353–368 of 567 proofs of work