Andrej Karpathy
VERIFIED TECHNICAL DOSSIERSources checked

Andrej Karpathy

Founder, Eureka Labs | Former Director of AI at Tesla & OpenAI Founding Member

5 Verified ArtifactsSource Checked & Attributed

Verified Proof of Work Artifacts

5 items cataloged

Each artifact below represents an authenticated research publication, production code repository, or technical architectural framework directly authored or co-created by Andrej Karpathy. Every entry undergoes editorial source verification.

#1
IMPLEMENTATION Checked Sep 20, 2026

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

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.

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

#2
EXPLANATION Checked Sep 20, 2026

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

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.

Model & Execution Context:Pure Python 3.10+, scalar-valued autodiff DAG, Graphviz topological computation graph visualization.
Scope & Limitations

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

#3
IMPLEMENTATION Checked Sep 20, 2026

AutoResearch: Autonomous AI Research Agent Conducting Iterative ML Experiments

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.

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

#4
EXPLANATION Checked Sep 20, 2026

Neural Networks: Zero to Hero Video Lecture Curriculum & Repository

A masterclass educational curriculum and open-source codebase tracing neural networks from basic micrograd autograd engines through multi-layer perceptrons, WaveNet, and full Transformer language models. Authored by Andrej Karpathy to provide deeply intuitive, first-principles deep learning literacy.

Model & Execution Context:Pure PyTorch and custom Python micrograd engine; runnable on consumer GPUs or Google Colab.
Scope & Limitations

Pedagogical and reference implementation. Optimized for code clarity, educational inspection, and conceptual clarity rather than maximum distributed hardware throughput.

#5
IMPLEMENTATION Checked Sep 20, 2026

nanoGPT: Minimal, Readable, and Fast PyTorch Framework for Training Medium GPTs

A pure, clean, and hackable 300-line PyTorch implementation for training, fine-tuning, and evaluating transformer language models with support for FlashAttention and torch.compile.

Model & Execution Context:PyTorch 2.0, CUDA, Python, FlashAttention, and OpenWebText.
Scope & Limitations

Engineered for educational transparency and small-to-medium parameter models; multi-node distributed trillion-token training requires complex pipeline parallel frameworks like Megatron.