[cs702]

It's not just a commodity, but a very capital-and-labor-intensive one.

Developing, training, retraining, and delivering a competitive AI service today requires an army of expensive people and a huge investment in infrastructure. It's not cheap!

Moreover, any improvement, whether it's better generation, longer context, a new modality, cheaper prices, or something else, is quickly replicated by all major groups.

They're in a race to see who can spend the most the fastest while charging the least to prevent customers from switching.

Meanwhile, free open-source alternatives keep getting better and more efficient.

[dartos]

Which open source models are u referring to? Most are based on llama, which isn’t open source, and some on mistral, which is more open, but we don’t have access to the data, so we can’t make it ourselves (in the spirit of open source)

[fzzzy]

When people say open source model they are really trying to say available weights model.

[Calvin02]

I hear you but I think even partially “open source” models are better than all the closed systems that can only be accessed via APIs.

I would like to acknowledge companies that offer even partially open source models in hopes of convincing others to do the same.

[mofosyne]

At the very least for the engine there is now llama.cpp which is an open source port. Admittedly the real magic is in the weights however.