Home

It's 2023, and the AI market has exploded. AI is no longer confined to a single sector; it's permeating various industries, from healthcare and finance to manufacturing and retail and is creating unprecedented opportunities for businesses and individuals who embrace this transformative technology and provide the solutions and services needed to harness its potential.

However, the initial hype is beginning to wane. The AI landscape is maturing, and data is becoming more proprietary. In this evolving landscape, the role of the data scientist is shifting from a pure analyst to an engineer role, requiring less expertise in data analysis and more in programming, data engineering and software development, for integrating existing trained models instead of training them from scratch.

In this changing environment, a C++ deep learning library can now become more relevant than a Python library. The significance lies not only in C++'s speed and resource efficiency but also in its compatibility with industries that heavily rely on C++ codebases.

Industries like robotics or finance have long-established codebases predominantly written in C++. As AI continues to weave itself into the fabric of these sectors, seamless integration with existing systems becomes paramount. C++ libraries offer a natural bridge between AI and these industries, enabling the development of AI-powered solutions that seamlessly integrate with their existing C++ code. ( I don’t see Tesla’s people writting the software from their self driving cars in python, they do it in C++ )

While libraries like PyTorch and TensorFlow offer powerful Python APIs for constructing intricate machine learning models, the transition to a C++ production environment presents distinct challenges. These challenges primarily revolve around the need to seamlessly integrate the models built in Python into a C++ ecosystem tailored for production, since the C++ front ends they provide are poorly designed and bloated with unwanted python dependencies.

PyTorch team recognizes this shifting of paradigms in the field and the limitations of Python, prompting their efforts to rewrite their code base in C++. However, they face the challenge of maintaining backward compatibility during this transition. TensorFlow by the other hand, faced a lot of criticism for its bad design, which has led to a decline in its popularity. Meanwhile, Google's JAX, in contrast, is again another pure Python library. And those are the most popular libraries out there.

In this context, a simple lightweight C++ deep learning library can become very relevant.