Andrej recounts his first encounter with deep learning in a University of Toronto class taught by Geoff Hinton, where he was fascinated by the idea of networks learning like a mind. Later, during his master’s at UBC, a course with Nando de Freitas deepened his interest, and he found neural networks far more satisfying than traditional AI techniques. He describes how he became the “human benchmark” for ImageNet by manually classifying thousands of images (building a JavaScript interface to handle the 1 000‑class problem) and discovering that networks often outperform humans, especially on fine‑grained statistics and texture, while still struggling with tasks that require reading text. Andrej highlights the transformative impact of teaching a hands‑on deep‑learning class (CS231n), where implementing everything from scratch gave students a true understanding of the stack and made the course a highlight of his PhD. Reflecting on the field’s evolution, he notes the surprising generality and scalability of deep learning, the power of transfer learning/fine‑tuning, and the limited progress of unsupervised learning despite early optimism. Looking ahead, he sees AI splitting into applied engineering (improved supervised/unsupervised models) and an AGI track focused on building a single, end‑to‑end neural agent whose objectives yield intelligent behavior. His advice to newcomers: learn by building from scratch—write your own library, understand every layer, and only then use high‑level frameworks—so you truly grasp what’s happening under the hood.
1. Andrej Karpathy was an undergraduate at the University of Toronto.
2. Geoff Hinton taught a deep learning class at the University of Toronto during Karpathy's undergraduate studies.
3. The class covered restricted Boltzmann machines trained on MNIST digits.
4. Karpathy completed a master's degree at the University of British Columbia.
5. During his master's, he took a machine learning class with Nando de Freitas.
6. Karpathy conducted a self‑classification experiment on the CIFAR‑10 dataset.
7. In that experiment he achieved an error rate of about 6% on CIFAR‑10.
8. He created a JavaScript interface to measure human performance on the ImageNet classification task.
9. ImageNet contains 1,000 object categories.
10. For the ImageNet human baseline task he listed all categories and provided example images for each.
11. He spent approximately one to two weeks on the ImageNet human baseline experiment.
12. He attempted to recruit additional lab members to repeat the experiment, obtaining at least some approximate performance data.
13. Other researchers began referring to him jokingly as the “reference human” for ImageNet.
14. Deep neural networks later surpassed his personal ImageNet classification performance.
15. Karpathy taught a deep learning course (CS231N) and made the lecture materials available online.
16. He devoted essentially all of his research time to teaching that course for about four months during his PhD.
17. The course included discussion of recent research papers from the preceding week or even the same day.
18. Students appreciated that the course covered low‑level details and required them to implement concepts from scratch.
19. Karpathy wrote his own JavaScript library, ConvNetJS, to implement convolutional neural networks.
20. The deep learning‑AI specialization begins with several weeks of Python programming instruction.
21. Karpathy spent roughly the last year and a half working at OpenAI.