[brundolf]

I think the market is going to explode (if it hasn't already) for on-prem, or at least private, LLMs on par with ChatGPT. This could be served by companies building their own, or by open-source projects, or by OpenAI or OpenAI's competitors

As a side-effect, this feels like a bright spot in the potentially authoritarian trajectory that AI could take as labor becomes less and less valuable. It encourages development of LLMs that compete with the current default option and can be run on more and more limited hardware. Enterprises might even want separate departments, or separate individuals, to be able to run their own models to prevent leakage

[eachro]

Open source is a race to the bottom. Seems like the only obvious winner then is people selling the shovels aka NVIDIA.

[esafak]

More like a race to the top: you're forgetting all the applications that will be built on top of these open source models.

[eachro]

I'm not so sure. Does anyone pay for pytorch, numpy, tensorflow? In the matter of weeks we've seen llama.cpp, alpaca.cpp released to the public. Barrier to entry in this market is quickly going to zero.

[jerojero]

Has Pytorch and numpy's quality gone down though?

These projects are funded by several organisations that rely on them for their operations. This is the same for Linux, just because the software is accesible for free doesn't mean it's not funded.

I really want to see governments putting more funding on open source as well, public money, public code. As they say.

[esafak]

Those are libraries, not applications. Think of copilots, midjourney, etc.

[pixl97]

No one pays for Windows, SQL server, and Office. That's why Microsoft is poor.

[rvz]

Finally someone is thinking. Stable Diffusion is already at the finish line to the bottom, since their AI model is already open source. Many other open source LLMs and DALL-E 2 alternatives are available competing against O̶p̶e̶n̶AI.com.

They are all gradually catching up and O̶p̶e̶n̶AI.com cannot run their services for free forever or even close to free. Eventually the price hikes will come in.

But as long as the AI industry continues to use inefficient methods of training, fine-tuning and inference via using tons of GPU hardware, NVIDIA will continue to smile at relying on this for a long time until a true breakthrough in efficient training and inference methods in neural networks on everyday desktop or typical servers.