Machine learning framework and tool creators on GitHub
The developers behind the frameworks and minimal implementations that shaped how deep learning research code actually gets written.
À propos de cette page
L’interface du site s’affiche dans votre langue. Le texte long de cette page a été rédigé en anglais et n’a pas encore été traduit.
| N° | Développeur | Connu pour |
|---|---|---|
| 1 |
Andrej Karpathy
@karpathy
|
Founding member of OpenAI and former Tesla AI director, known for minimal, readable implementations like micrograd and nanoGPT. |
| 2 |
François Chollet
@fchollet
|
Created Keras, which made deep learning approachable years before most competing frameworks caught up. |
| 3 |
Soumith Chintala
@soumith
|
Co-created PyTorch, now the dominant framework for machine learning research. |
| 4 |
George Hotz
@geohot
|
Built tinygrad, a deliberately minimal deep-learning framework, after founding comma.ai. |
| 5 |
Jeremy Howard
@jph00
|
Co-created fast.ai, a course and library built around making deep learning usable without a research background. |
Deep learning’s tooling split fairly early into two philosophies: frameworks built to scale to enormous production and research clusters, and smaller projects built to be read end-to-end by one person in an afternoon. Both philosophies are represented here, sometimes by the same author at different points in their career.
Scale and readability pull in opposite directions
PyTorch and Keras exist because research teams needed frameworks that could handle large models across real hardware without every researcher reimplementing gradient computation from scratch. Andrej Karpathy’s micrograd and George Hotz’s tinygrad exist for close to the opposite reason - stripping a framework down until the entire mechanism fits in a file a newcomer can read in one sitting.
Neither approach is more legitimate. A profile built around minimal, readable code is optimizing for a different reader than one built around production scale, and it’s worth being clear about which one you’re aiming for.
Teaching shaped the tooling as much as research did
Jeremy Howard built fast.ai explicitly to lower the barrier to entry into a field that had become intimidating to approach from the outside. That educational framing shows up throughout this list - readable reference implementations tend to double as teaching material, whether or not that was the original intent.
What this means for your own profile
If you’re building or studying machine learning tooling, a small, well-commented implementation that another developer can actually read start to finish is a legitimate and often more visible contribution than a marginal benchmark improvement buried in a much larger codebase.
You can see how that kind of project reads on a showcase profile, or generate the same breakdown for your own account.