[127361]

They've joined the Linux Foundation, does that mean the models are going to be eventually censored to satisfy the foundation's AI safety policies? That includes ensuring the models don't generate content that's non-inclusive or against diversity policies?

[pico_creator]

Currently the main policy is only around copyright - and nothing about AI safety: https://www.linuxfoundation.org/legal/generative-ai

Also in the full power of opensource, if LF really force something the group disagree with, we will just fork

All the other alignment policies, are optional for groups to opt-in

So I would not worry so much about that - the group already has a plan in event we need to leave the Linux Foundation - for example: If USA regulates AI training (since LF is registered under USA)

[viraptor]

Downvoted, because it's a very trolly way to ask this. Especially given the foundation doesn't have an AI safety policy from what I've seen. Let's be better than this...

[lhl]

It is trivial to fine tune any model (whether a base model or an aligned model) to your preferred output preferences as long as you have access to the model weights.

[Al-Khwarizmi]

Not trivial for the general public at all, and furthermore, you need much more memory for finetuning than for inference, often making it infeasible for many machine/model combinations.

[lhl]

If you are running a local LLM already (which no one in the "general public is") then the bar is really not that much higher for fine-tuning (either for an individual or community member to do).

And you don't need any additional equipment at all. When I say trivial, I really do mean it - you can go to https://www.together.ai/pricing and see for yourself - a 10M token 3 epoch fine tune on a 7B model will cost you about $10-15 right now. Upload your dataset, download your fine tune weights (or serve via their infrastructure). This is only going to get easier (compare how difficult it was to inference local models last year to what you can do with plug and play solutions like Ollama, LM Studio, or Jan today).

Note also that tuning is a one-time outlay, and merges are even less resource intensive/easier to do.

To put things in perspective, tell me how much cost and effort it would be to tune a model where you don't have the weights at all in comparison.

[Al-Khwarizmi]

Running a local LLM - downloading LM studio, installing on Windows, using the search function to search for a popular LLM, click "download", click the button to load the model, chat.

Fine-tuning - obtaining a dataset for your task (this in itself is not trivial), figuring out how the service you linked works (after figuring out that it exists at all), uploading the dataset, paying, downloading the weights - OK, now how do you load them into LM studio?

It's all subjective, of course, but for me there's a considerable difficulty jump there.