Hacker News new | ask | show | jobs
by dsign 1041 days ago
> I believe it’s hard to build a business on open-source AI at this stage.

I would say it makes sense to build a business on running open-source AI models.

I’m a hacker with side-projects. The outputs of ML models are useful to my side projects. But there is a world of difference between whipping up a tiny bit of Python code, or even using a web interface to query an open source model online, than running that same model locally.

For starters, I would need to invest a few hundred bucks in hard disk alone. My 3090 is no good to run a bunch of models, so I would have to invest in a more beefy setup. And, when the hardware is home and properly installed, I would have to spend tons and tons of hours fiddling with the code and its setup.

I rather put all those resources to the direct function of my side-projects, and not to something that, say, Stability AI can offer me for cents and in seconds, at least for the volume of services that I need.

That’s not to say there wouldn’t be scenarios where the financial equation goes the other way, but for me at least that hasn’t been the case so far.

2 comments

Running AI models is extremely expensive. If you want to offer it to the public, you have to plan capacity very accurately, otherwise, either the quality will suffer or you will bleed too much money, too early.

Take OpenAI for example, even with their enormous popularity, I believe they are still making huge losses. The only way they can survive (right now) is with the huge financial injection from Microsoft. Besides, you can run your AI models on GCP or AWS and other providers easily, no need to invest locally. For a pay-on-hours tariff, it’ll be affordable for experiments. Huggingface is yes taking care of all the “logistics” for you already.

We make private versions of the models and serve them at scale cheap as we have loads of fast chip access when the market has run out.

Had a record month

It’s known as an open core business model

https://twitter.com/emostaque/status/1649152422634221593?s=4...

FWIW, you can definitely run stable diffusion on your 3090. I played around with it on my 3060m (laptop) and it works surprisingly well.

The models are pretty tiny as well, a few GB. I think you could get pretty far with your setup if you wanted too