Blog posts

On water and AI: towards sustainable practices for building AI systems

posted on 2024-10-09

According to a recent article the authors behind the assertion that LLMs (GPT-3 in this case) consume 500 ml of water per 10-50 requests have now reportedly revised their estimate upwords by a factor of 4. Having a relatively novel computing paradigm at our disposal comes with opportunities, but for also with responsibilities. Not every problem is a nail to the LLMs hammer and, as AI carpenters, we should make sure that LLMs aren't the only tools in our toolbox.

Running NVIDIA GPU enabled containers in GNU Guix

posted on 2024-10-02

This blog post details the process of enabling NVIDIA GPU support for Docker containers on GNU Guix. The solution includes a workaround involving the bind mounting of /gnu/store inside containers to access required libraries, which comes with some security concerns. As such, the setup is not yet suited for production use, especially in exposed environments.