agunapal

My first thought after reading the blog was, let me share the blog with Claude and ask it how bots can circumvent this.

imo AI bots have significantly affected OSS and we need better qualitative measures to define success

Yes, one can easily setup agents to bump up the stars, increase pip downloads etc

I think it comes down to "Is the juice worth the squeeze"

As someone who worked for a large organization maintaining an OSS project, one issue I faced was how do you show impact? We used to have many organizations really love and use our project , but they would hardly give anything back to the project, including writing blogs where they could have shared some success stories. IMO github stars/pip downloads etc are not good metrics and these are even worser metrics in today's agentic AI world. Its so easy to fake these nowdays.

Here is a paper from few years ago where they talk about 7x speed increase, which equates to savings.

https://arxiv.org/abs/2101.03961

Do you mean model sharding?

Very competitive price for the speed and intelligence being offered!

Nvidia had the first movers advantage. Nvidia spent so many years perfecting CUDA to work well with PyTorch. Before ROCM, there was only CUDA. There were so many developers building their use cases on top of PyTorch+CUDA, and bringing all that feedback to PyTorch, this made CUDA battle ready and stable. AMD can get there, especially now with demand for compute, but as someone already said here, the biggest focus needs to be on PyTorch

If you really think about why MoE came into existence, its to save significant cost during training, I don't think there was any concrete evidence of performance gains for comparable MoE vs dense models. Over the years, I believe all the new techniques being employed in post training have made the models better.