# Andrej Karpathy

> 1986– · Computer Scientist, AI Researcher
>
> **Recorded contribution:** Deep learning educator; Tesla Autopilot AI director; minGPT; char-rnn

## How to use this dossier

Read for a causal chain, not a hero story: inherited problem → contribution → mechanism → downstream capability → limit. Then close the page and complete the reconstruction exercise from memory.

## 1. Historical orientation

Slovak-Canadian computer scientist Andrej Karpathy became known for Stanford's CS231n course, work on image captioning and recurrent character models, research at OpenAI, leadership of Tesla's vision-based Autopilot team, and later open educational implementations such as minGPT and micrograd. His durable contribution is as much pedagogical as technical: he repeatedly exposes the smallest working mechanism behind deep-learning systems. Karpathy repeatedly turned opaque neural systems into inspectable programs. CS231n connected convolution mathematics to implementation; char-rnn exposed sequence generation; micrograd reduced reverse-mode automatic differentiation to a small computation graph; minGPT reconstructed the Transformer training loop without production complexity.

## 2. The problem inherited

Deep learning became powerful but opaque to learners and product teams, while deploying perception models in vehicles required linking training data, model behavior, onboard compute, and safety evidence.

## 3. The central contribution

Karpathy advanced visual-language and sequence modeling, led applied vision teams, and created influential from-scratch educational resources for neural networks and Transformers.

## 4. Reconstruct the mechanism

1. Reduce a model to inspectable tensors, losses, gradients, and update steps.
2. Train on a clearly defined dataset and visualize errors rather than only aggregate metrics.
3. Compress and deploy the model under latency and hardware constraints.
4. Collect new edge cases from operation and feed them into a versioned evaluation and retraining loop.

## 5. What changed downstream

- CS231n and open implementations educated a generation of deep-learning practitioners.
- His applied work popularized data-engine-centric approaches to large vision systems.
- These materials trained a generation of practitioners and made educational reimplementation a legitimate form of technical scrutiny rather than merely a simplified tutorial exercise.

## 6. Attribution, limits, and uncertainty

- Research, Tesla Autopilot, and OpenAI systems are team efforts; education visibility must not absorb collaborators' credit.
- From-scratch demos omit production safety, data labor, privacy, compute, and governance, while driver-assistance claims require independent crash and operational evidence.
- Minimal code omits distributed training, numerical stability, data governance, evaluation, security, and production failure handling; conceptual clarity must not be confused with deployment completeness.

## 7. Reconstruction lab

Implement scalar automatic differentiation, then train a two-layer network. Write a model card identifying every production requirement the educational artifact does not satisfy. Add gradient checks and one deliberately unstable operation to micrograd, then explain which protections a production autodiff system needs beyond the pedagogical core.

## 8. Evidence trail

- [Andrej Karpathy](https://karpathy.ai/) — Andrej Karpathy
- [CS231n: Deep Learning for Computer Vision](https://cs231n.stanford.edu/) — Stanford University
- [micrograd](https://github.com/karpathy/micrograd) — Andrej Karpathy on GitHub

---

*Research checked 2026-08-09. Dates, roles, and claims about living people are historical snapshots. Linked sources remain the authority; this dossier is original instructional synthesis.*