You Can't Have a Free Software AI Stack | Hacker News Reader

You Can't Have a Free Software AI Stack

4 pointsVox_Leone10 comments

A creeping crisis in free software?

You've probably noticed by now that it's not possible to build a complete AI stack with free software. You can start building your dataset with free tools, in house, but soon you will run into difficulties with dataset annotation and model training. If your work is in CV area (my case) you can choose manual labeling with (merely) Open Source tools, or use a backend to automate the task [Tensorflow, PyTorch, JAX, etc., all proprietary]. Image annotation tools like [often require] working by pulling images from the cloud rather than your workstation. Then you soon find yourself making an account with Microsoft, Google or Amazon.

Are you a genius with plenty of time to do everything on your premises? You won't escape Nvidia's [GeForce] or Intel's [Arc] tentacles.

Then you go back to Tensorflow/PyTorch/OpenCV/Nvidia/Intel to train and put your model into production… You get the idea.

Will there be a place for free software in AI?

Huh?

I wrote my own system for classifying a stream of texts in Python, I might Open Source it one of these days but I have to get it to the point where it is modular enough that I can customize it to do the particular things I want without subjecting people to my whims... I use it every day and I'm not afraid to demo it because it is rock solid.

My understanding is that my system would not be hard to adapt to work on images for certain kinds of tasks.

Pytorch is open source, Huggingface is open source. CUDA isn't. This is

https://labelstud.io/

and for annotating text spans there are so many open source tools

https://github.com/doccano/doccano

I worked for a company a few years back that built annotation tools for projects we sold to customers but never quite got to a polished general purpose annotator. Today there are an overwhelming number of companies in this space and products I never heard of, many of which are cloud based or paid. Looks like a gold rush to me.

I mean Free as in speech, not merely Open Source.

I can't have a fully Free software stack, the kind that Richard Stallman would approve [not that I'm a follower of any FS cult, mind you - I'm just doing a pretty cold assessment].

Regards

Like it has to be GPL not MIT or Apache?

Well, of course you are right in your assertion(s), in a sense. I'm not an expert in Licensing, just a concerned user of tech, but yes, for this exercise I mean GPL, as laid out on the FSF site:

https://www.fsf.org/blogs/licensing/sharing-knowledge-about-...

Consider it a challenge: can you build/setup an AI stack [or environment, or pipeline] that Richard Stallman would use?

> can you build/setup an AI stack [or environment, or pipeline] that Richard Stallman would use?

GPT4All and RedPajama would both be sufficiently open for him to use, though neither is covered under an FSF-approved license. llama-cpp itself should be perfectly acceptable to him, and there are lots of open API wrappers for these models.

"Are you a genius with plenty of time to do everything on your premises? You won't escape Nvidia's [GeForce] or Intel's [Arc] tentacles."

What do you mean exactly?

That you can't have free software without building your whole computers from scratch using raw minerals mined with a pickaxe you made yourself using wood you chopped yourself using an axe you made yourself from a tree you grew yourself?

>What do you mean exactly?

Does that hardware allow for running FSF approved software, as a ThinkPad does [I ask in earnest]?

I would say yes, since we're building it: https://engageusers.ai/ecosystem.pdf

Si si avec des données synthetiques de Stable ml

Why do you think that the libraries that you have mentioned are not open source and what keeps you from using FOSS tools to label your data properly?

Reply on news.ycombinator.com