Tensorflow wins

Github profiles Leaf and Tensorflow

We started with the development of Leaf briefly before Google released Tensorflow. For two weeks Leaf’s hypothesis seemed unique.

Developers will build machine learning applications, not scientists — let’s give developers a tool to do that.

Although Leaf has top-notch performance and an uniquely simple yet expressive API, it will lose against Tensorflow. Leaf’s current theoretical benefits[1] are less significant, than the benefits that Tensorflow provides[2] over the early, scientific frameworks like Torch, Caffe, Theano.

The next generation of tools, that help developers to build machine learning applications will build on Tensorflow, or more specifically on higher-level frameworks, like Keras, who abstract over multiple AI Engines[3]. Back in November, when we started, this trend was less obvious to us[4].

Now that good-enough tools, to build maintainable machine-learning applications, like Tensorflow and Keras, exist, venture capital flows more and more into companies who try to create immense value in verticals with new AI-driven applications, instead of infrastructure providers.

We wanted Leaf to become the #1 machine learning framework for developers, but it became apparent, that Leaf will not receive substantial traction outside the Rust community.

Which is why Max and I will suspend the development of Leaf and focus on new ventures.

I am staying in the space of startups and AI, working with a VC to explore a new type of early stage investment fund.

Thank you so much everyone who supported us on the way to more than 4.000 Github stars and almost into Github’s Top 1000. We are very grateful.

We will continue to give back to the community with our future projects.

[1]: Significantly easier/slimmer abstractions over computation and scheduling, first-class citizen support for CUDA/OpenCL/Rust and co., clean foundation for auto diff via dual numbers and differentiable programming, compression of neural network models.

[2]: Tensorflow and its ecosystem (incl. GCloud) provides a proven solution/process for testing, deploying and maintaining models.

[3]: E.g. Facebook is working on Flow, AutoML and Asimo, which are tools to make the creation of machine learning models even easier. If they build on Tensorflow or Torch or another framework needs to be seen.

[4]: This is why the the other frameworks will vanish or take niche positions. This includes Microsoft’s CNTK.