¿Qué opinas sobre que D tenga sus propias librerías de IA y compita con Python?

Hola a todos 👋,

Últimamente he estado desarrollando un proyecto de inteligencia artificial utilizando el lenguaje D, y me surgió una reflexión que quiero compartir con la comunidad:

Sabemos que Python domina el área de la inteligencia artificial por su enorme ecosistema (TensorFlow, PyTorch, scikit-learn, etc.), pero también presenta desventajas: rendimiento limitado, dependencia de extensiones en C/C++, y en muchos casos una menor eficiencia en entornos con pocos recursos.

Por otro lado, D es un lenguaje potente, de alto rendimiento, con capacidades de metaprogramación, manejo eficiente de memoria, y sintaxis moderna. ¿Por qué no aprovechar estas ventajas para construir herramientas y librerías nativas de IA?

Algunas ideas que se podrían trabajar:

• Librerías de redes neuronales en D (como ya existen prototipos de LSTM, perceptrones, etc.)
• Implementación de algoritmos de aprendizaje reforzado puro en D
• Utilidades de manejo de datos, entrenamiento y visualización
• Interfaz mínima compatible con formatos de datasets estándar (CSV, JSON, etc.)

Ventajas de usar D para IA:

• Compilación nativa y alto rendimiento.
• Control preciso del uso de memoria (útil en sistemas embebidos o sin GPU).
• Código más expresivo que C++ y más rápido que Python.

Mi pregunta para ustedes:

¿Creen que el lenguaje D tiene el potencial para convertirse en una alternativa real a Python en el desarrollo de IA? ¿Qué opinan de crear una comunidad o iniciativa abierta para impulsar librerías de inteligencia artificial hechas 100% en D?

Los leo con interés. 🚀

Saludos,
Daniel N.

What do you think about D having its own AI libraries and competing with Python?

Hi everyone 👋,

Recently, I’ve been working on an artificial intelligence project using the D programming language, and it sparked a thought I’d like to share with the community:

We all know Python dominates the AI space due to its vast ecosystem (TensorFlow, PyTorch, scikit-learn, etc.), but it also comes with downsides: slower performance, dependency on C/C++ extensions, and inefficiency on low-resource systems.

On the other hand, D is a powerful systems-level language, offering high performance, modern syntax, meta-programming, and precise memory control. So, why not leverage these strengths to build native AI tools?

Some ideas worth exploring:

• Neural network libraries written in pure D (e.g., LSTM, perceptron, CNNs)
• Pure D implementations of reinforcement learning algorithms
• Data processing and training utilities
• Lightweight tools for standard dataset formats (CSV, JSON, etc.)

Why D for AI?

• Native compilation and excellent runtime speed
• Fine-grained memory management (ideal for embedded or resource-limited environments)
• More expressive than C++, faster than Python

My question to the community:

Do you believe D has the potential to become a serious alternative to Python in AI development? Would you support an open-source initiative to create artificial intelligence libraries entirely in D?

Looking forward to your thoughts and feedback. 🚀

Best regards,
Daniel N.

On Friday, 2 May 2025 at 16:37:13 UTC, daniel wrote:

Hi Daniel,

First of all let's start with your questions:

No. Not even specific languages like Julia and Mojo currently can create any real alternative to Python.
Some small team can try to use it of course for their project, but wider adoption is not the case. Most of the research and real production cases support only Python.

There are already several native D libraries in this area.
Vectorflow, grain, golem, tensorformat.

For any significant improvement and development you need quite big community.
Even if the new DL framework will be released in D - I don't think it will get any wide adoption - because there is no community who will improve and push forward the solution.

For example, Rust has several ML frameworks. And because they are more resource efficient and they have Rust community - they are trying to push it.. but even they can't really compete with Python for now. But they have a chance. D in the current situation - only if some research group will do everything by themselves..
But I don't see any improvements over Python + C++/Rust approach.

Second,
Python AI ecosystem is based on C++ and Rust libs. TF, PyTorch, tokenisers and many other libs used by the industry are based on high perf, portable langs.

D has issues with this as well. Efficient Float8, Float16, BFloat16 and especially GPU support - are essential for current situation.

Where D can be applied from my point of view - "unix way" approach.
Some small task which doesn't require GPU support with some usage of C++/C API Rust libs - maybe possible way to use the language.

Or for educational purposes implement some algorithms in D also could be fine.

On Friday, 2 May 2025 at 19:38:35 UTC, Sergey wrote:

There would be no reason for anyone to adopt it unless it does something unique.

I can't say I've looked at everything Rust has pushed in this area, but their standard thing about rewriting stuff in Rust works for a small Unix utility, and then they can post it to Hacker News and get upvotes because "Woah! Neat!" That strategy will be completely useless in this area because the target audience wants to get stuff done and they don't have a use for a weird, complex language.

It would be worthwhile to do everything before and after calling the ML library in D. It's a matter of someone writing a convenient library. In this community, unfortunately, you'd likely get a bunch of templated crap that will drive people away.

On Friday, 2 May 2025 at 20:31:09 UTC, Lance Bachmeier wrote:

Exactly.
And even if it will have something unique - it will be just ported to Python-ecosystem in 1 week/month. And it will be done.

Actually the API they are creating for such frameworks are quite close to Pythonic.
For example candle and burn 2 quite popular frameworks.
Other libs can be found here: https://www.arewelearningyet.com

The data pipeline is possible but again they are using quite rich ecosystems, which could be either web-service oriented (which is usually very Go specific) or BigData oriented (which is very Java specific with Hadoop and Scala's Spark).